20 Watts Was Enough — Full Concept Book

A biologically inspired R&D blueprint for sparse, grounded, continual, energy-efficient AI.

The adult human brain operates within a whole-organ power budget of roughly 17–20 watts (C-001). That observation does not provide a fair brain-versus-silicon benchmark by itself; it establishes a serious engineering question: which computational constraints behind adaptive biological intelligence can be translated into measurable requirements for artificial systems? The energy evaluation chapter defines the boundary required for a comparison.

This repository turns that question into a traceable research programme: primary evidence becomes scoped claims, cross-domain mechanisms are deduplicated, and retained ideas face explicit null models and failure rules.

Current status: the evidence framework, architecture and bounded experiment machinery exist. An integrated AI system and claim-eligible workstation result do not. Development smoke checks remain NO_RESULT. The generated coverage report and readiness summary record the exact current boundary.

Read the project

The portal and web book list all 24 official EU languages, but publish only translations tied to an exact canonical source digest and human review. English remains canonical. Where a reviewed translation does not yet exist, the language control opens a focused contribution issue instead of unlabelled machine output. See the translation contract.

What the project tests

Useful adaptive intelligence may require much less active computation and data movement than current systems routinely spend. The project investigates whether that gap can be attacked through a system that:

• separates total capacity from active capacity;
• grounds representations in aligned perception, action, and outcome;
• allocates computation according to uncertainty and task demand;
• separates rapid episodic learning from slow structural learning;
• consolidates before pruning or hardening;
• keeps mutable facts in inspectable memory rather than forcing everything into weights; and
• measures quality, energy, data movement, lifecycle cost, and uncertainty together.

Nature is a source of mechanisms, not a shortcut around engineering controls. Observations from neuroscience, biology, ecology, physics, chemistry, mathematics, social systems, and other fields are normalized by their causal operation, deduplicated into shared principle bundles, and tested against strong ordinary engineering alternatives.

Plan and progress

The research roadmap is the canonical dependency order. Its stages run from evidence and experiment contracts through isolated mechanisms, adaptation, grounding, structural maturation and substrate co-design. They are gates, not promised dates.

GitHub milestones project those stages into issues and pull requests. They answer what work is open and who is doing it; they do not decide whether a scientific claim is established. The authorities remain separate:

Closing an issue updates operational progress. Only a reviewed evidence change updates a claim, and only a qualifying run updates result status.

The default normative context is the European Union and Germany. Legal, standards, and conformity statements remain qualified by jurisdiction, role, version, applicability, and date under the normative baseline.

Scientific conduct follows the research integrity baseline, which deduplicates the recurring ALLEA, EU, DFG, and large European university rules for disclosure, stewardship, ethics, correction, and review.

How the repository works

primary sources -> scoped claims -> deduplicated principles
                -> engineering translations -> equations and architecture
                -> equal-budget tests -> bounded evidence -> revised claims

• research/claims.md gives major assertions stable C- identities and evidence status.
• research/principle-registry.md groups equivalent cross-domain mechanisms under stable P- identities.
• concept/ develops the maintained system synthesis.
• math/ defines notation, units, derivations, and testable efficiency models.
• experiments/ separates written protocols, development plumbing, and result authority.
• decisions/ records durable choices without rewriting their history.
• sources/ preserves provenance and publication boundaries; imported discussions and summaries are leads, not evidence.

Canonical source and editing

Git main is the canonical source. The Pages portal, HTML book, and PDF are generated views of the same committed material; Google Docs and imported chats are not synchronized document stores. Work incrementally on one chapter, claim, equation, diagram, test, or decision rather than regenerating the whole project.

Read AGENTS.md and the engineering and research contract before editing. The research-writing skill keeps project prose direct and evidence-aware, with a deliberately narrow automated check for high-confidence generated filler. The repository-rule crosswalk explains which conventions from the maintainer's other projects were adopted, adapted, staged, or rejected here.

For a local live preview:

npm ci
npm run dev:github-pages

The preview runs at http://localhost:5173/ and reloads canonical Markdown, equations, tables, diagrams, and plots as their source files change.

Qualifying releases publish repository tooling and individual experiments as small container images. Use the complete image@sha256:... identity from the release's checksum-bound oci-images.json, not a mutable tag. The release contract explains admission, provenance and supported platforms; the experiment instructions provide the exact isolated commands. For example, the static validation image runs without network access:

image='ghcr.io/lusoris/20-watts-was-enough-20w@sha256:...'
docker pull "$image"
docker run --rm --network none \
  --pull never \
  --mount "type=bind,src=$PWD,dst=/repo,readonly" \
  "$image" \
  validate docs --root /repo

Releases may also attach a verified native Go binary as a convenience. The containers are the portable public default and remain scoped per experiment.

Before a local commit, follow the validation sequence: check the changed contracts and their downstream consumers, combining scopes for mixed changes. Unknown, unsafe or shared-authority scope requires the full gate.

Run the complete npm and Go gates before marking a pull request ready, integrating into main, merging or releasing.

Changes to book sources also require:

npm run generate:book-pdf
npm run validate:book-pdf

See CONTRIBUTING.md for the complete workflow and SUPPORT.md for issue routing. The repository does not maintain a PowerShell, C# or host-specific execution lane.

Citation, support, and licence

Use CITATION.cff and identify the exact release or commit when citing the project. Scientific use should also cite the original sources that support the specific claim; citing this synthesis does not replace them.

If the project is useful, support is available through Ko-fi or GitHub Sponsors.

Project-authored software, scripts, tests, configurations, schemas, and workflows are licensed under the EUPL v1.2 or later. Original project prose, mathematics, diagrams, plots, and presentation are licensed under CC BY-SA 4.0. Third-party and source material retains its own terms. Read LICENSING.md for the controlling boundary.

Scope

This chapter states the project's argument, its engineering requirements, and the limits that keep a biological analogy from becoming mythology.

Thesis

The brain demonstrates that adaptive behavior, continual learning, perception, memory, and action can coexist under a tight energy and communication budget (C-001). Contemporary AI demonstrates a different strength: highly parallel optimization over vast data and parameter spaces.

The project hypothesis is that a capable artificial system should combine the search capacity of modern learning with constraints that biology cannot evade:

1. only a small, relevant fraction of capacity should be active for an event;
2. communication and memory movement must be priced, not hidden behind FLOPs;
3. learning must begin from aligned perception, action, and consequence rather than language alone;
4. rapid acquisition must not directly rewrite stable long-term structure;
5. consolidation must precede destructive compression; and
6. stable skills, mutable facts, and active reasoning should use different storage and execution paths.

The intended result is a system whose cost scales primarily with the information and uncertainty relevant to a task, rather than with its full stored capacity.

Biological observation

Neural signaling is metabolically costly, cortical activity is constrained, and biological learning spans multiple timescales. Sensory and motor experience precedes mature language. These observations motivate design constraints.

They do not establish a single brain algorithm, an equivalent number of digital operations, or a guarantee that every biological mechanism is efficient on silicon.

Proposed AI translation

The blueprint combines four separations:

• capacity versus activity: sparse routers select modules, depth, memory, and precision;
• prediction versus correction: predictable state follows a cheap path; residual uncertainty receives more computation;
• experience versus structure: fast memory captures episodes while slow learning integrates regularities under regression tests; and
• reasoning versus knowledge: stable skills may be compiled or quantized, while mutable facts remain retrievable and attributable.

Development is a controlled lifecycle:

flowchart LR
    seed["Capacity-rich modular seed"] --> ground["Sensorimotor grounding"]
    ground --> plastic["Sparse active learning"]
    plastic --> consolidate["Replay and consolidation"]
    consolidate --> validate["Generalization and stability gates"]
    validate --> prune["Structured pruning"]
    prune --> harden["Quantize, compile, or externalize"]
    harden --> operate["Continual operation"]
    operate --> plastic

Editable source: ../assets/diagrams/developmental-pipeline.mmd.

Efficiency mechanism

The optimization target is constrained quality per unit of total system energy, with latency and uncertainty limits. A model is not efficient merely because it has sparse weights; the runtime must avoid loading and operating on inactive state, and routing overhead must remain smaller than the work skipped.

The top-level objective is developed in the energy model:

Here $Q$ is task quality, $E_{\text{facility}}$ is measured facility energy, $L_{p95}$ is tail latency, and $R$ is a defined risk or calibration metric.

Evidence status

• Metabolic constraints on neural activity are established within the scope of C-001.
• Conditional computation is established in narrower engineered systems under C-003 and C-004.
• The complete developmental pipeline is a speculative integration. No cited paper validates it as one system.

Speculative extensions

• Learn a global energy price that modules must “bid” against when requesting compute.
• Let stable routing motifs become physically colocated or compiled for lower data-movement cost.
• Allow controlled structural growth when no existing module can absorb a new regime without interference.

Failure modes

• Anthropomorphic labels conceal incompatible mechanisms.
• Sparsity saves theoretical FLOPs but increases real communication or latency.
• Energy optimization suppresses rare but important computation.
• Early specialization prevents transfer and produces brittle modality silos.
• “Hardening” turns uncertain or mutable behavior into an uncorrectable path.

Measurable predictions

The integrated hypothesis survives only if, at matched quality:

• active parameters and memory traffic grow more slowly than total capacity;
• average energy falls without unacceptable tail-risk or calibration loss;
• sequential learning retains prior capability better than a matched single-timescale baseline; and
• grounded intervention tasks improve beyond gains explained by additional data or parameters alone.

A capable system should execute, adapt, and consolidate on different timescales while one controller prices energy, latency, and risk.

Scope

This is the shortest complete description of the proposed system. It joins the project's individual mechanisms into one operating architecture and shows where information moves, where learning happens, and which decisions remain reversible.

Architecture at a glance

flowchart TB
    subgraph event["Event loop"]
        direction LR
        world["World · user · tools"] --> fast["FAST RUNTIME<br/>predict · select · execute"]
        fast --> result["Outcome + telemetry"]
        result --> world
    end
    subgraph lifecycle["Learning and lifecycle"]
        direction LR
        adapt["ADAPTATION"] --> provisional["Episodes + provisional modules"]
        provisional --> slow["MAINTENANCE"]
        slow --> stable["Slow model + hardened paths"]
    end
    result --> adapt
    provisional --> fast
    stable --> fast
    budget["Resource + risk controller"] -. "energy · latency · risk" .-> fast
    budget -.-> adapt
    budget -.-> slow
    result --> budget

Editable source: ../assets/diagrams/three-coupled-loops.mmd.

The three loops solve different problems:

The resource controller crosses all three. It can restrict an action because of energy, latency, communication, or risk, but it cannot silently replace the task objective. Measured outcomes return to the controller so estimated costs can be corrected.

Biological observation

Living systems repeatedly separate processes by timescale and locality:

• many components are available, while only a small subset is strongly active together (C-001, C-003);
• local circuits resolve routine events while larger systems coordinate exceptions (C-017, C-024);
• fast traces change behavior before slow structure is rewritten (C-008, C-019);
• replay, forgetting, and structural maintenance are selective processes (C-036–C-042);
• stable structure can be protected and later reopened under bounded intervention (C-043–C-045); and
• local demand can recruit adjacent resource supply (C-049–C-051).

The shared principle is separation of concerns: execution, adaptation, memory, structure, and physical supply change at different speeds and respond to different signals. Digital systems add exact copying, external stores, checkpoints, shadow evaluation, and rollback. Those capabilities make the separations easier to test and safer to revise.

Proposed AI translation

1. Fast runtime

For each event, the runtime forms a predictive state containing expected observations, uncertainty, current goals, and relevant recent context. A hierarchical selector then chooses the cheapest admissible path:

• a hardened low-cost path for a familiar, low-risk event;
• one or more conditionally active experts;
• another layer or reasoning step;
• episodic or factual retrieval;
• a sensor, tool, or environment intervention; or
• escalation to a human or a more capable controller.

The output is more than an answer. It includes measured latency, energy, communication, calibration, and any observed consequence. That telemetry is the input to later learning and resource decisions.

2. Adaptation

An outcome is captured with provenance before it can alter slow parameters. It may update temporary state, create a provisional adapter, change a local connection, or simply become a replay candidate. A surprising event is not automatically a lesson; it may be noise, attack, contradiction, or evidence of a new regime.

There is no ex-nihilo generator in this design. Every proposal is built from inherited structure and experience: complete patterns, fragments, relations, or abstractions are copied and transformed. The generative cycle is therefore explicit:

flowchart TB
    sources["Inherited structure + experience"] --> copy
    subgraph synthesis["Construct a candidate"]
        direction LR
        copy["Copy fragments + relations"] --> compress["Compress + abstract"]
        compress --> vary["Vary + recombine under a goal"]
    end
    vary --> test{"Test in simulation or the world"}
    subgraph selection["Evaluate the result"]
        direction LR
        test -->|"supported"| retain["Retain with provenance"]
        test -->|"fails"| revise["Reject · revise · keep uncertain"]
    end
    retain --> sources
    revise --> vary

Editable source: ../assets/diagrams/generative-recombination.mmd.

Stochasticity can alter which variants are tried. Usefulness comes from the entire cycle: decomposition, recombination, intervention, comparison, and selection. The supporting evidence and unresolved distinctions are tracked in the endogenous generation audit and is falsified first through Candidate 004.

Reconstructive adaptation through external state

Visual art and design cognition make the cycle more operational. Proposal generation is reconstructive: learned categories, exemplars, relations, styles, constraints, tools, and prior transformations shape what can be proposed (C-842–C-860). That does not reduce the process to nearest-neighbor retrieval. A system can externalize a partial state, inspect consequences that were not explicit in its preceding plan, change representation, probe the world, and use the result to create a new branch. The novelty is relative to a declared history; the validity comes from constraints and observed outcomes.

The adaptation loop therefore distinguishes seven event types:

1. Exposure: which sources, examples, styles, and constraints were available before the task;
2. Retrieval: which versioned sources were deliberately requested during the task and at what cost;
3. Reconstruction: which fragments, relations, or abstractions formed a proposal branch;
4. Externalization: which editable representation—text, sketch, scene graph, CAD state, program, simulation, or prototype—made the proposal inspectable;
5. Epistemic action: which render, measurement, simulation, query, or material intervention was selected to reduce a decision-relevant uncertainty;
6. Evaluation and selection: which hard constraints, qualified evaluators, risk limits, and selection rule admitted or rejected a version; and
7. Retention: which source, parent, operation, test, rejected branch, dependency, expiry, and rollback records were preserved.

flowchart LR
    brief["Hidden problem · declared constraints · evaluator contract"] --> expose["Logged source exposure"]
    corpus["Versioned source corpus"] --> expose
    corpus --> retrieve["Metered retrieval · precedents · cases"]
    expose --> reconstruct["Reconstructive generation"]
    retrieve --> reconstruct
    reconstruct --> branches["Independent proposal branches"]
    branches --> externalize["Externalize · sketch · scene graph · CAD"]
    externalize --> inspect["Inspect · query · resegment"]
    inspect --> transform["Transform representation or proposal"]
    transform --> externalize
    inspect --> action["Epistemic action · render · simulate · probe"]
    action --> material["Material or environment feedback"]
    material --> externalize
    externalize --> diversity["Diversity and fixation audit"]
    diversity --> evaluate["Constraint validity · usefulness · risk · cost"]
    evaluate --> select["Blinded evaluation and selection"]
    select --> retain["Retain · expire · recombine · rollback"]
    retain --> lineage["Version graph · sources · operations · tests · rejected branches"]
    lineage --> retrieve
    nulls["Retrieval · CAD/editors · diffusion · QD/BO · scratchpads · active learning"] --> evaluate

Editable source: versioned-reconstructive-design.mmd.

Several separations are non-negotiable. More proposals are not necessarily better proposals; proposal diversity, constraint validity, selection regret, and realized outcome are different measurements. Similarity to an example can transfer a useful structural relation or cause fixation and negative transfer, so similarity and task effect are both reported. A provenance graph can record derivation without proving intention, authenticity, correctness, or originality. Aesthetic response is observer-, context-, expertise-, and history-qualified rather than a universal scalar.

Parallel branches are useful only if they remain genuinely independent before critique. A shared canvas can coordinate work and simultaneously propagate an early fixation. The comparison therefore includes private branches, delayed reveal, source substitution, example removal, evaluator swap, material swap, and representation conversion. Every alternative pays for retrieval, proposals, tools, simulations, human critique, physical material, waste, retention, and recovery.

Fixture F-002 tests this complete loop against a mature composition of retrieval, CAD and editors, generative sampling, explicit search and quality-diversity, scratchpads and shared workspaces, active learning, simulation, version control, provenance, and independent reranking. The mathematical contract keeps exposure, retrieval, relative novelty, reinterpretation, information gain, fixation, negative transfer, constraint margins, evaluator qualification, selection regret, lineage, reconstructability, and lifecycle cost separate. If that ordinary stack ties, the additional architecture is removed while the fixture remains as a negative result.

3. Maintenance

Maintenance receives a budgeted set of possible actions. For any memory, module, or route it may:

• defer while collecting more evidence;
• replay against related and conflicting cases;
• merge redundant records while retaining provenance;
• externalize mutable propositions to an inspectable store;
• reduce influence or plasticity;
• delete state after retention and safety checks;
• reopen mature structure in a reversible branch;
• change logical topology or physical placement; or
• promote stable computation into a cheaper representation.

The formal action model is defined in ../math/memory-lifecycle.md. Structural changes remain shadowed and reversible until regression, calibration, and energy tests pass.

4. One event end to end

1. A sensor, user, or tool changes the current state.
2. The runtime predicts what matters and estimates uncertainty.
3. The selector chooses depth, modules, memory, precision, and possible intervention under the current budget.
4. Execution produces an outcome and physical telemetry.
5. The adaptation loop stores an attributable episode and may construct a bounded hypothesis or provisional module.
6. The maintenance loop later replays the episode against related and conflicting evidence.
7. A validated regularity may be consolidated, compiled, quantized, externalized, relocated, or used to change future routing.
8. Regression or fragility tests can reject the change and restore the previous state.

5. State ownership

Efficiency mechanism

Efficiency is an event-level constrained decision. At time $t$, the selector chooses an admissible action $a$ from $\mathcal{A}_t$:

subject to

Here $s_t$ is the predictive state; $\Delta V(a)$ is expected task-value improvement; $E(a)$ is energy in joules; $L(a)$ is latency in seconds; $R(a)$ is the declared risk measure; and $B_t^E$, $B_t^L$, and $\epsilon_t^R$ are event-specific limits. Measured cost after execution recalibrates the estimator.

Lifecycle comparison includes the work usually hidden outside inference:

An architecture advances only if it improves the joint frontier of quality, risk, latency, and lifecycle energy against its strongest conventional baseline. Skipped FLOPs alone are not an efficiency result.

Evidence status

The engineering analogue audit maps the architecture to feedback control, value of information, caches, schedulers, change detection, adaptive routing, replicated state, and system identification. Each experiment must compare against the strongest relevant version of those methods.

Speculative extensions

• Receiver-specific fast and slow context channels under a hard bit budget.
• Modules bidding for compute by expected value improvement while a global controller enforces physical and safety constraints.
• Active fragility probes on shadows or replicas before topology changes.
• Capability-gap repair that admits complementary modules instead of restoring arbitrary capacity.
• Joint learning of logical routing, memory placement, and hardware locality.

Each remains outside the core architecture until its isolated contract beats the appropriate engineering baseline.

Failure modes

• Dense control disguised as sparsity: routing, broadcast, or monitoring moves nearly as much state as dense execution.
• Controller conflict: runtime, adaptation, maintenance, and resource loops oscillate or optimize incompatible signals.
• False maturity: a shortcut appears stable and is hardened before compositional and intervention tests.
• Memory laundering: unsupported episodes become slow parameters without provenance or contradiction checks.
• Maintenance dominates cost: replay, probes, migrations, and regression testing consume the savings of conditional execution.
• Fixation propagates through shared state: early examples or collaborators collapse independent branches before alternatives are tested.
• Generation impersonates evaluation: the proposal model's own score is reported as independent evidence of validity or usefulness.
• Lineage impersonates truth: a complete derivation record is treated as proof of correctness, intention, authenticity, or originality.
• Renaming without mechanism: a familiar cache, controller, or scheduler receives a biological label without a different causal operation.

Measurable predictions

The architecture earns implementation only if isolated experiments support the following predictions:

1. Conditional execution reduces measured data movement and wall energy at matched quality, calibration, and tail risk.
2. Fast attributable memory improves adaptation latency while slow-model regression remains bounded.
3. Reversible maturity gates reduce catastrophic drift without blocking necessary relearning or newcomer admission.
4. Structured recombination plus targeted intervention produces more valid, useful novelty than matched-budget stochastic sampling alone.
5. Editable externalization, inspection, representation change, and epistemic action improve hidden-constraint validity, selection regret, and transfer beyond the complete retrieval/editor/generation/search/tool stack without increasing fixation or lifecycle cost.
6. Maintenance improves retention and adaptability after replay, monitoring, migration, and recovery costs are counted.
7. At least one proposed mechanism beats the strongest ordinary controller, scheduler, cache, router, or system-identification baseline.

Until those tests pass, this is a working architecture: precise enough to reject, modular enough to simplify, and explicitly incomplete.

Scope

This chapter defines how the project may borrow from living systems without turning a biological implementation into a specification. The brain is an existence proof for efficient adaptive intelligence, but neither evolution nor neural tissue optimized the same objective, hardware, scale, or reliability contract as an artificial system.

The working rule is:

Reproduce the useful constraint or computation, not the substrate accident.

Biological observation

Biological intelligence is shaped by slow, noisy, failure-prone components; metabolically expensive communication; physical growth; local chemical signals; and an inability to checkpoint, clone, or roll back a whole organism. Brains compensate with local processing, sparse activity, redundancy, specialized cell types, multiple timescales, and continual maintenance.

Those compensations are informative, but they are not automatically optimal on silicon. Digital systems offer different capabilities: fast switching and interconnect, exact copying, explicit addressing, external storage, reversible experiments, global synchronization when it is worth its cost, and precision that can be selected per operation. At the same time, digital systems still pay heavily for data movement, memory access, communication, cooling, and idle capacity. The relevant question is which constraints and useful computations survive a change of substrate.

The speed comparison must stay qualitative until a shared task and latency boundary exist. Axonal conduction, synaptic integration, transistor switching, GPU kernels, and cluster collectives are different operations. Selecting the fastest number from each domain would recreate the invalid comparison rejected by C-016.

Proposed AI translation

For each biological candidate, record five transformations:

1. Observed function: what the organism demonstrably achieves.
2. Biological constraint: which physical or evolutionary limit shaped it.
3. Candidate invariant: the computation that may survive a substrate change.
4. Silicon implementation: the least literal engineered mechanism that tests the invariant.
5. Escape hatch: the biological limitation that engineering should not inherit.

This transformation is a project decision, not evidence that any row will produce a gain. The candidate inventory is maintained in the neuroscience opportunity map and comparative-biology map. Its relationship to existing AI is tracked in the adoption matrix.

Efficiency mechanism

The approach preserves useful functions in the cheapest available form and uses digital capabilities—including exact copies, reversible branches, and direct addressing—where they improve the measured system.

For a biological candidate $m$, the engineering experiment compares at least three systems: a conventional baseline $B$, a literal or close biological translation $L_m$ when meaningful, and a silicon-native abstraction $S_m$. Using the measurement contract from the energy model, the candidate is interesting only if $S_m$ improves the quality–risk–energy frontier:

where $Q$ is task quality, $R$ is the declared risk metric, $E$ is energy within the declared boundary, and $L$ is latency. The symbol $\succ$ means Pareto dominance under pre-registered tolerances; it does not collapse unlike units into a decorative score.

The decision geometry below is illustrative. Its points and simultaneous regions are not measurements; they show why a favorable mean on one coordinate cannot override uncertainty or a failed protected-outcome gate.

Evidence status

• Neural energy constraints are established only within the scope of C-001.
• Nonlinear dendritic subunits, homeostatic scaling, neuromodulated plasticity, and specialized inhibitory control are biological observations under C-017 through C-020; their proposed artificial abstractions are not thereby validated.
• Performance of the complete cross-substrate system remains speculative until matched engineering tests establish it.

Speculative extensions

• Learn when global communication is worth buying instead of banning it.
• Compile frequently reused local circuits into fast deterministic paths while retaining a slower plastic path for exceptions.
• Search jointly over algorithm, memory hierarchy, interconnect, precision, and physical embodiment rather than treating hardware as a final deployment detail.
• Import organizational principles from organisms without centralized brains, including plants, immune systems, cephalopod limbs, and adaptive transport networks.

Failure modes

• A metaphor is mistaken for a mechanism.
• “Biological” becomes an unearned synonym for efficient or intelligent.
• “Silicon is faster” is used to ignore memory movement, synchronization, or thermal limits.
• A negative result for one implementation is treated as falsifying the underlying biological abstraction—or vice versa.
• The project claims an idea is absent from AI after finding only that it is absent from mainstream foundation models.
• Evolutionary fitness is confused with task accuracy, truthfulness, or human values.

Measurable predictions

This framing earns its place if, across isolated mechanism experiments:

• silicon-native abstractions equal or outperform literal translations at matched quality and system boundary;
• retained biological constraints predict where energy or interference is saved;
• removing a proposed invariant removes the benefit even when the biological surface form remains; and
• at least one hybrid uses a capability unavailable to biology—such as exact rollback or external versioned memory—to improve continual learning without increasing the declared risk metric.

Different sciences often describe the same constrained operation in different nouns. The architecture should pay for the operation once.

Scope

This chapter turns the project's open-world research policy into an architectural method. Findings may originate in neural tissue, plants, immune systems, animal collectives, ecosystems, control theory, databases, or materials. They enter the design only after their causal operation has been normalized and compared with the principles already present.

The canonical unit is therefore neither a paper nor an organism. It is a versioned problem–solution invariant with a stable P- identifier, scoped evidence, an engineering null model, and an experiment that can reject its AI translation. The detailed registry remains in the principle ledger; this chapter explains how its bundles compose into one system.

Biological observation

The current evidence corpus repeatedly encounters five pressures.

1. Scarcity: more states, candidates, or possible actions exist than can be active at once. Examples include constrained cortical signaling (C-001), sparse insect odor codes (C-025), immune selection (C-028), and congestion-triggered reserve routes (C-035).
2. Locality: a complete central description is too slow or expensive. Dendritic branches integrate locally (C-017); cephalopod limbs retain peripheral control (C-024); and collective systems can coordinate through sparse influence or shared environmental state (C-031, C-054).
3. Reversible change: recent evidence must affect behavior before it deserves permanent structure. Eligibility traces (C-019), plant priming (C-026), selective replay (C-036), and retrieval-sensitive memory (C-039) all separate a temporary state from a later commitment decision.
4. Stability: selection and learning create positive feedback that must be bounded. Synaptic scaling (C-018), feedback inhibition (C-025), reconstruction around a target organization (C-033), and ecological recovery dynamics (C-058, C-059) expose different parts of that control problem.
5. Repeated work: a stable solution should migrate from expensive general search into topology, placement, representation, or material. Relevant observations include body–controller co-development (C-029), structured pruning (C-012), lower-precision representation (C-013), and activity-dependent resource placement (C-050).

These recurrences do not make the underlying mechanisms identical. They reveal where multiple fields expose the same engineering pressure and where one shared experiment can replace several renamed proposals.

Proposed AI translation

From domain language to a mechanism record

Every retained observation is rewritten as the following tuple:

where:

The tuple is a structured research record, not a numerical embedding. Its fields retain units and provenance. Two findings are candidates for one principle only when their problem, causal operation, information topology, and timescale agree at the abstraction needed by an experiment.

Source organization and scale are not optional metadata. A material structure, an organism-level control loop, and an ecosystem-level interaction can deliver a similar function through different causal paths, authority, and timescales; they do not merge until those differences are shown irrelevant to the discriminating experiment.

Biomimetic transfer is a search method, not an evidence grade

Biomimetics gives this project useful process language, but it does not create a shortcut through the evidence ledger. The search runs in both directions:

1. Phenomenon push begins with a measured effect in a living system and asks which artificial problem shares its function, constraints, causal operation, and failure boundary.
2. Problem pull begins with a measured artificial-system failure and asks which scientific fields contain a mechanism that solves the normalized problem under a comparable budget.

Phenomenon push broadens the established biology-push/solution-driven route; problem pull broadens technology-pull/problem-driven design. The broader names matter because this project searches physical, formal, social, and engineering sciences as well as biology.

Both routes converge on $M$, then pass through deduplication, silicon-native redesign, a strong conventional null, and a falsifiable equal-budget test. A visual likeness without transferred function is rejected; a transferred function without a causal and cost model remains only a lead; and a validated source mechanism does not establish that its artificial translation works.

flowchart TB
    bio["PHENOMENON PUSH<br/>observe a reproducible effect"] --> bmeasure["Measure function,<br/>mechanism, cost, boundary"]
    problem["PROBLEM PULL<br/>declare failure + budget"] --> pspec["Specify function,<br/>constraints, strongest null"]
    bmeasure --> abstract["FUNCTIONAL ABSTRACTION<br/>normalized mechanism record"]
    pspec --> abstract
    resemblance["Shape-only or story-only resemblance"] --> reject["REJECT AS LEAD<br/>no functional transfer"]
    abstract --> dedup["DEDUPLICATE<br/>merge · discriminate · hold"]
    dedup --> substrate["RE-ENGINEER<br/>use silicon affordances"]
    substrate --> test["FALSIFY<br/>equal quality · risk · lifecycle budget"]
    test --> keep["KEEP / MERGE / RETIRE"]

    classDef bio fill:#087f5b,stroke:#34d399,color:#ffffff,stroke-width:2px;
    classDef pull fill:#c2410c,stroke:#fb923c,color:#ffffff,stroke-width:2px;
    classDef shared fill:#6d28d9,stroke:#a78bfa,color:#ffffff,stroke-width:2px;
    classDef gate fill:#1d4ed8,stroke:#60a5fa,color:#ffffff,stroke-width:2px;
    classDef pass fill:#166534,stroke:#4ade80,color:#ffffff,stroke-width:2px;
    classDef fail fill:#991b1b,stroke:#fb7185,color:#ffffff,stroke-width:2px;
    class bio,bmeasure bio;
    class problem,pspec pull;
    class abstract,dedup,substrate shared;
    class test gate;
    class keep pass;
    class resemblance,reject fail;

The public descriptions of biomimicry are useful for discovering biological strategies and unifying themes, while ISO 18458:2015 and VDI 6220 Part 2 provide more precise terminology and process frames. None is evidence that a particular mechanism transfers or improves efficiency. The dated method audit records the source hierarchy and the parts adopted here.

Five solution families

The thirteen current P- principles remain distinct, but they compose into five navigational families:

flowchart TB
    evidence["Primary observations from many fields"] --> record["Normalize problem · state · operation · topology · timescale"]
    subgraph online["Online pressure"]
        direction LR
        scarcity["Scarcity<br/>P-001 · P-007"] ~~~ locality["Locality + coordination<br/>P-002 · P-008 · P-011 · P-013"]
    end
    subgraph lifecycle["Lifecycle pressure"]
        direction LR
        change["Reversible change<br/>P-003 · P-004 · P-005 · P-012"] ~~~ stability["Stability + maintenance<br/>P-006 · P-009"] ~~~ structure["Structure as computation<br/>P-010"]
    end
    record --> scarcity
    record --> locality
    record --> change
    record --> stability
    record --> structure
    scarcity --> primitive["Shared engineering primitives"]
    locality --> primitive
    change --> primitive
    stability --> primitive
    structure --> primitive
    primitive --> test["Null model · ablation · lifecycle measurement · rejection"]

Editable source: ../assets/diagrams/recurring-solution-families.mmd.

Deduplicate before architecture expansion

A new finding passes through six decisions:

1. Scope the evidence. Record the observed system, intervention, result, uncertainty, and boundary as a C- claim.
2. Normalize the mechanism. Fill every field of $M$ without organism-themed names standing in for operations.
3. Search the registry. Compare against all existing P- records, including held candidates and negative evidence.
4. Merge or discriminate. Merge when the same control loop would be tested; otherwise name the smallest experiment that distinguishes the mechanisms.
5. Name the strongest null. A scheduler, cache, controller, estimator, database, routing rule, or standard learning method gets the same interface and budget.
6. Change one canonical object. Update a principle, primitive, equation, diagram, or experiment instead of adding another themed architecture.

This creates a many-to-one evidence graph:

many observations → scoped claims → fewer principles → shared primitives → decisive experiments

Recurring outcomes still need a construct firewall

Different fields often expose the same broad problem while assigning different causal content to it. Ecological succession describes ordered community change; an ecological priority effect requires an arrival-order intervention; continual-learning task order concerns sequential parameter updates; curriculum learning chooses an order; and a scheduler may merely decide when unchanged jobs execute. Calling all five “succession” would erase the experiment.

The history-conditioned modular-succession audit therefore reuses C-008, C-056, C-057, C-574, Fixture F-014, and Candidates 004 and 019 without creating another principle. Its residual test holds the task multiset, eligible module identity set, exogenous presentations, update ceilings, capacity, optimizer, evaluator and budget fixed, randomizes order, and separately cuts age, routed acceptance, capacity pre-emption, shared-state modification, facilitation and lock-in. If canonical replay, a non-learning scheduler, a mature continual-learning method, or a search-budget-charged curriculum explains the outcome, the biological transfer adds nothing.

New mechanisms can sharpen a principle without multiplying it

The developmental, regenerative, plant-plasticity, and multiscale-reduction depth passes add new failure boundaries without adding another universal principle. That is a useful result, not a failed search.

• Positional memory and repair separate what state says should exist from which workers can currently rebuild it (C-1506, C-1507). They sharpen P-005, P-009, and P-012 by making write authority, reconstruction capacity, and current context separately testable.
• Compensating sources, finite scaling, receiver geometry, delayed reinforcement, conditional redundancy, and local/global fields (C-1508–C-1515) refine locality, maintenance, and structural offloading. They do not imply that every duplicated route, boundary, or organizer is useful.
• Plant memory divides population-level digital switching, lifecycle reset, writing, trace maintenance, retrieval, systemic route identity, action-conditioned growth, and joint environmental context (C-1516–C-1525). These are discriminating state and protocol dimensions inside P-002, P-005, P-006, P-009, P-012, and P-013—not a plant-themed architecture.
• Projection memory, normal-hyperbolicity limits, macro-to-micro queries, and lift dependence (C-1526–C-1529) supply mathematical rejection tests for coarse state. They constrain when an existing runtime or memory primitive is valid; they are not biological claims at all.

The deduplicated output is therefore richer in contracts while remaining at thirteen P-series principles. Fixtures F-022, F-023, and F-024 preserve the distinctions so later implementations can test whether a proposed composition adds value beyond its strongest ordinary null within a declared workload and transfer boundary.

Composition without double counting

Principles can cooperate while remaining separately testable. For one event:

1. prediction-error allocation decides whether more evidence is valuable;
2. selective allocation chooses a module within the active budget;
3. local autonomy runs that module near its state;
4. a temporary trace records unresolved credit;
5. the maintenance plane later decides whether to replay, protect, merge, or forget it; and
6. structural offloading compiles only the recurring path that survives those tests.

Calling the entire chain “sparsity,” “memory,” or “homeostasis” would hide which operation produced a gain. Each experiment therefore removes one principle at a time while holding the other interfaces constant.

Interconnection is an intervention

A small or typed interface does not by itself make two modules dynamically modular. In transcriptional regulation, covalent-modification cycles and yeast synthetic circuits, attaching a downstream target can alter the upstream state or transient response (C-1550–C-1555). Substrate competition in Drosophila MAPK signalling supplies a distinct, potentially functional case (C-1559). The shared mechanism record is connection-induced back-action; the biochemical literature calls one important family retroactivity.

The domains do not collapse into one mechanism. Direct binding can sequester a signal, multiple substrates can compete for an enzyme, and nominally separate modules can instead couple through shared transcriptional, translational, CPU, memory, or scheduler resources (C-1558). The causal intervention therefore crosses connection state and shared-resource load independently. It also includes a genuinely nondestructive immutable-read control, because silicon can often copy or address state without the molecular consumption that motivated the analogy.

Insulation is bounded rather than automatically desirable. It must preserve both the producer and useful downstream service while exposing latency, staleness, copies, buffers, replicas, operations, maintenance and physical energy. Weak coupling may spend less while losing tracking or robustness (C-1556, C-1557); substrate competition can also be a useful integrator rather than a fault (C-1559). Fixture F-027 therefore compares suppressing, preserving and deliberately using back-action. If immutable messages, ordinary queues, resource isolation, explicit filters, stop-gradient paths or conventional controllers reach the same frontier, the translation is retired without adding a principle.

Passive self-organization is a mandatory null

Efficient-looking structure does not establish sensing, represented goals, or counterfactual action. Fracture sets, drainage networks, and river avulsions can arise from local stress, gravity, flow, conservation, thresholds, and stored geometry (C-232–C-242). The project calls a mechanism adaptive control only when it specifies:

1. a service variable external to the adaptation law;
2. observations with spatial/temporal support and latency;
3. a decision that could choose differently under different evidence;
4. authority and an actuator;
5. movement, reserve, monitoring, and recovery budgets;
6. persistent state and reset cost; and
7. a guarantee or tested failure envelope.

This criterion makes passive physics a stronger baseline. If local flow–structure feedback produces the same topology and service without a controller, the controller must justify its sensing, decision, switching, and maintenance cost. Connectivity is not throughput (C-234); material removed by “pruning” must appear as transport, storage, or output elsewhere (C-239); and a topology change can reassign or destroy service rather than improve it (C-242). A closed-form branching law is therefore a regime-qualified physical null, not a portable architecture rule (C-1489).

Transported fields are not messages

Microbial signals, metabolites, fungal nutrients, plant hormones, process streams, queues, and network packets can all create a local field that later receivers use. The common systems object is not “communication in nature.” It is a transported, transformed, delayed, and locally decoded quantity whose physical path may also change future capacity (C-563, C-578, C-580).

For transported type $k$, use the local balance

where concentration $c_k$ is mol/m³, diffusion coefficient $D_k$ is m²/s, velocity $\mathbf u$ is m/s, production and uptake rates $p_k,q_k$ are mol/(m³ s), and first-order loss $r_k$ is 1/s. Digital fields use an analogous typed byte or item ledger; they do not inherit molecular units or diffusion laws. Geometry, boundary conditions, sampling support, and receiver state are part of the contract. Countercurrent orientation is one finite exchanger arrangement inside that contract; it does not establish effectiveness or a generic bidirectional-computation benefit by itself (C-1490).

Several fields can be superposed at the same endpoints. If the observed outcome is

where $G$ is the number of typed channels, $g$ indexes one channel, $c_g$ is its transported concentration or inventory, $h_g$ is its receiver-dependent effect, $\theta_g$ is channel state, $b$ is background, and $\varepsilon$ is measurement error. Observing $y$ or cutting every channel together does not identify any one $h_g$. Each claimed channel needs an independently targetable intervention or a justified identification model. Rillig et al.'s concurrent-fungal-network perspective makes this problem explicit for fungal guilds sharing plant endpoints; the project adopts it as an engineering inference, not as evidence that such guild combinations have a demonstrated beneficial function (DOI, C-217, C-584).

flowchart LR
    P["Local production, release, or deficit"] --> T["Transport: flow · diffusion · delay · loss"]
    G["Geometry, topology, and boundaries"] --> T
    T --> C["Local concentration or typed inventory"]
    C --> R["Receiver state + uptake + response"]
    R --> O["Task outcome + local cost"]
    O --> U["Update source, sink, conductance, or topology"]
    U --> G
    C --> A["Accumulate · transform · consume · export · purge"]
    A --> C
    F["Damage, contamination, incompatibility"] --> X["Contain · isolate · reroute · regrow"]
    X --> G
    O --> V{"Independent validation"}
    V -->|"useful"| U
    V -->|"not useful"| D["Decay, disperse, or retire"]

Editable source: transported-field-contract.mmd.

The diagram deliberately spans several fields while retaining their differences:

• microbial quorum output depends on production, confinement, flow, loss, and receptor state; it is not a vote count (C-563, C-575);
• cross-feeding can help, harm, or create a weakest-link dependency, so local transfer is not system utility (C-571);
• fungal flow can both deliver material and remodel conductance while rapid containment remains distinct from later bypass (C-577–C-581); and
• distinct symbiotic partners can exchange unequally under conflicting objectives (C-584); and
• concurrent typed channels sharing endpoints remain causally unresolved when the experiment observes or disables only their aggregate.

The held fixture belongs jointly to P-011/P-013 and existing Candidates 001 and 013 (C-585). It must beat typed pub/sub or queues, fixed-graph adaptive routing, backpressure/primal–dual allocation, and make-before-break state transfer. Charge production, transport, storage, cleanup, contamination, topology edits, failed delivery, reserve, and measurement. If geometry and cleanup do not affect the task, use the simpler digital primitive.

Inactive is not one state

The immune evidence in C-727–C-747 does not add another architecture principle. It contributes a stricter state contract. Representation, recognition, authentication, authorization, activation, suppression, deletion, impairment, contraction, memory, and recovery answer different questions and have different reversal costs.

flowchart LR
    G["Generate candidate rule / module"] --> R["Represent sampled evidence"]
    R --> V{"Developmental validation"}
    V -->|"insufficient coverage"| I["Ignorant · unresolved"]
    V -->|"unsafe with high support"| D["Delete + retain decision record"]
    V -->|"provisional pass"| N["Eligible but inactive"]
    N --> C["Context · provenance · current system state"]
    C --> A{"Recognition + authority + risk gate"}
    A -->|"permit"| X["Activate · expand · execute"]
    A -->|"monitor"| M["Observe without authority increase"]
    A -->|"quarantine"| Q["Anergic analogue · reversible disablement"]
    A -->|"external control"| S["Suppressed · rate-limited · scoped"]
    X --> O["Outcome · collateral loss · resource use"]
    O --> K["Contract · clean up · retain memory"]
    X -->|"persistent load or dysfunction"| H["Exhausted / impaired trajectory"]
    K --> P["Maintained memory · placement · expiry"]
    P -->|"qualified recurrence"| C
    I --> R
    M --> C
    Q -->|"revalidate"| C
    S -->|"authority changes"| C
    H -->|"diagnose + bounded intervention"| C
    O -->|"harm or false permit"| D

Editable source: typed-tolerance-lifecycle.mmd.

An activation bit cannot distinguish an absent module from an unobserved one, a policy-suppressed module from a damaged one, or post-response contraction from retained memory. That ambiguity makes reactivation unsafe and makes resource accounting wrong. Each transition therefore records evidence support, authority, expected useful effect, collateral-loss terms, resource budget, expiry, recovery target, and observed outcome.

Recognition supplies evidence; it does not grant authority. Deletion requires a higher support threshold than reversible monitoring or quarantine because a rare useful capability can be lost. Local copies pay maintenance, refresh, invalidation, and inconsistency costs. Recovery ends only when useful service and reserve are both restored.

The typed lifecycle mathematics compares this contract against a complete ordinary stack: typed state machines, calibrated risk and abstention, least privilege, constrained control, anomaly detection, evolutionary or ensemble search, replay, placement, and resource-aware scheduling. If the ordinary stack reproduces the decisions and frontier, the biological vocabulary is removed and the state contract remains.

Coordination without a privileged clock

Music cognition provides a combined stress regime for mechanisms that already exist elsewhere in the project: multiscale prediction, local timing control, transformation-aware memory, constrained generation, role protocols, repair, and cumulative transmission. The evidence in C-748–C-783 does not justify a music-specific principle.

Its useful residue is a benchmark: agents must preserve literal phrase constraints and bounded relative timing while tempo, expressive microtiming, delay, roles, motifs, leaders, and partners change, without reading a privileged shared clock. Synchrony still requires common-input controls, edge interventions, and a protected functional endpoint before it can count as useful coordination (C-1495).

flowchart LR
    task["Score · cue · literal task constraints"] --> phrase["Phrase · motif · boundary state"]
    partner["Partner action"] --> channel["Delay · jitter · loss · acoustic/physical path"]
    channel --> observe["Timestamped local observation + uncertainty"]
    observe --> phase["Local phase · tempo · drift estimate"]
    memory["Partner · role · repertoire memory"] --> phrase
    phase --> controller["Role-qualified local controller"]
    phrase --> controller
    controller --> action["Timed expressive action"]
    action --> partner
    action --> evaluate["Literal task · synchrony · response · quality · cost"]
    evaluate --> repair["Correct · clarify · rehearse · update convention"]
    repair --> phrase
    repair --> memory
    nulls["PLL · Kalman · MPC · retrieval · typed protocol · rehearsal"] --> evaluate
    perturb["Tempo drift · role switch · leader loss · new motif · turnover"] --> channel
    perturb --> controller

Editable source: shared-clock-free-coadaptation.mmd.

The fixture refuses several easy substitutions:

• similar mean tempo is not pairwise synchronization;
• stimulus-frequency neural power is not evidence of endogenous prediction;
• low onset error is not the correct phrase or role;
• familiar-partner performance is not co-adaptation when retrieval explains it;
• novelty is not validity, responsiveness, or usefulness; and
• within-group convention compression is not cross-group task success.

Fixture F-001 compares the full composition with fixed correction, PLL/adaptive oscillator, Kalman identification, distributed MPC, a paid centralized conductor, transformation-aware retrieval, and typed role/version protocols. The mathematical contract keeps local clock error, phase, phrase prediction, role safety, message bytes, latency, rehearsal, recovery, and joules separate. If online estimation plus retrieval and explicit protocol state ties, the residual is retired.

Preserved bits are not preserved meaning

Library and archival science adds a dependency that storage abstractions often hide: information remains useful only for a declared interpreter community with the formats, schemas, vocabularies, identity mappings, authority evidence, software, units, permissions, and practiced capability needed to use it. The audit evidence will occupy C-784–C-803.

flowchart LR
    object["Payload · metadata · evidence links"] --> package["Versioned preservation package"]
    community["Designated interpreter community + capabilities"] --> package
    dependencies["Format · schema · vocabulary · software · authority"] --> package
    queries["Registered queries · units · tolerances"] --> gate
    package --> transform["Retain · compact · migrate · emulate · delete"]
    transform --> candidate["Candidate version + manifest"]
    candidate --> gate{"Query · evidence · authenticity · access gates"}
    gate -->|"pass"| publish["Publish version + supersession lineage"]
    gate -->|"fail"| fallback["Retain source · narrow contract · repair"]
    publish --> drift["Community · dependency · vocabulary · authority drift"]
    drift --> invalidate["Invalidate · re-test · migrate · retire"]
    invalidate --> gate
    requests["New unregistered information need"] --> abstain["Answer if supported · otherwise abstain / recover source"]
    publish --> abstain
    outcomes["Task outcome · error · effort · cost"] --> queries
    publish --> outcomes

Editable source: query-registered-preservation.mmd.

This separates several failure modes that are otherwise mistaken for one success:

• fixity detects bit change; it does not establish truth or authenticity;
• provenance records lineage; it does not validate the assertion;
• a valid ontology constrains representation; it does not prove instance data;
• a citation edge is not endorsement or comparable impact;
• findability and interoperability do not establish correctness or safe reuse; and
• retaining records does not retain the tacit capability needed to interpret or act on them.

The query-registered preservation mathematics defines finite native-unit tolerances, evidence reachability, interpreter and dependency validity, migration regression, unregistered-query abstention, and complete lifecycle cost. The Candidate 017 track must beat ordinary OAIS/PREMIS-style packaging, versioned schema/vocabulary, retained-source recovery, and query-regression tests. If it does not, the ordinary preservation stack remains the implementation.

Candidate production is not acceptance

Mathematical practice supplies the sharpest version of a distinction already needed throughout this project: producing a candidate and earning acceptance are different operations. Conjectures, analogies, abstractions, proof sketches, retrieved lemmas, numerical patterns, solver traces, and learned tactic choices can all direct search. None of those origins establishes the resulting claim. The evidence and formal-system boundaries are tracked in C-861–C-880.

The transferable lifecycle is typed:

1. Propose. Bind a candidate to the exact problem, definitions, logic, axioms, library, corpus, proposal ancestry, and accessible evidence.
2. Challenge. Search examples, counterexamples, countermodels, boundary cases, definition changes, and adversarial encodings before commitment.
3. Decompose. Turn the target into a dependency DAG whose child artifacts reconstruct the parent goal through named interfaces and side conditions.
4. Prove or refute. Produce a proof term, checked finite reduction, solver certificate, admissible model, or exact counterexample—not only a verdict.
5. Check. Bind the artifact to the exact instance and library version, then run a small independent checker while publishing shared trust roots.
6. Publish. Retain proved, refuted, tested, unknown, disputed, and retracted as different states with evidence, provenance, and cost.
7. Invalidate. A definition, axiom, checker, dependency, or library change quarantines its reverse-dependency closure until rebuild or recheck.

flowchart LR
    identity["Immutable identity<br/>problem · definitions · logic · library · corpus · checker"] --> propose["PROPOSE<br/>conjecture · analogy · abstraction · candidate proof"]
    split["Dependency-safe split<br/>ancestor · sibling · restatement · source isolation"] --> propose
    propose --> challenge["CHALLENGE<br/>examples · countermodels · adversarial cases · critique"]
    challenge -->|"valid counterexample"| refuted["REFUTED<br/>typed counterexample + domain witness"]
    challenge -->|"survives finite tests"| tested["TESTED<br/>support range + precision + generator"]
    challenge --> revise["Revise statement · definition · abstraction"]
    revise --> propose
    challenge --> decompose["DECOMPOSE<br/>goal DAG · lemmas · interfaces · side conditions"]
    abstraction["Abstraction obligations<br/>soundness · adequacy · reconstruction · excluded cases"] --> decompose
    decompose --> prove["PROVE<br/>proof term · solver trace · model · exact finite reduction"]
    prove --> check["CHECK<br/>independent kernel or certificate checker"]
    separation["Generator/checker separation<br/>frozen input hash · explicit shared roots"] --> check
    check -->|"accepted certificate"| proved["PROVED<br/>scoped proposition + dependency closure"]
    check -->|"failed or unsupported"| unknown["UNKNOWN<br/>timeout · incomplete theory · malformed certificate"]
    refuted --> publish["PUBLISH<br/>versioned artifact · state · evidence · provenance · cost"]
    tested --> publish
    proved --> publish
    unknown --> publish
    disputed["DISPUTED<br/>formalization · assumptions · significance · review"] --> publish
    publish --> registry["Release registry<br/>proved · refuted · tested · unknown · disputed · retracted"]
    registry --> change["Dependency, definition, checker, or library change"]
    change --> invalidate["INVALIDATE<br/>reverse dependency closure · quarantine · recheck"]
    invalidate --> retracted["RETRACTED<br/>superseded state + reason + affected versions"]
    invalidate --> challenge
    nulls["ITP · ATP · CDCL · SMT · CP/MIP · CEGAR/CEGIS<br/>formal libraries · learned guidance · exact numerics · peer review"] --> compare{"Equal lifecycle budget<br/>hidden families · human-hours · joules"}
    registry --> compare
    compare --> retain["Retain measured residual<br/>otherwise retire composition"]

Editable source: versioned-proof-discovery-lifecycle.mmd.

This lifecycle prevents five common substitutions:

• finite agreement is evidence for a tested range, not a universal result;
• a counterexample rejects the encoded universal claim but does not choose its unique repair;
• individually plausible lemmas do not close a parent goal unless their proof artifacts reconstruct it;
• a checker establishes the encoded proposition relative to its logic, definitions, axioms, libraries, preprocessing, and trusted implementation; it does not validate the informal intent or importance; and
• participant or model count does not establish independent checking when sources, parsers, libraries, prompts, or failure modes are shared.

Fixture F-004 compares the joined lifecycle with complete interactive and automated theorem provers, formal libraries, CDCL SAT, DPLL(T) SMT, constraint and model finding, CEGAR/CEGIS, learned premise/tactic guidance, exact numerical and exhaustive methods, and ordinary expert review. The mathematical contract defines immutable identities, leakage closures, abstraction obligations, proof-DAG reconstruction, certificate cost, typed states, invalidation, human effort, and joules. If the mature stack ties, the composition is retired; the state distinctions and negative result remain.

Efficiency mechanism

Deduplication saves two different resources.

First, it reduces research duplication. If plant priming, eligibility traces, and cache admission all motivate a temporary state before commitment, the project maintains one promotion interface and tests the domain-specific differences as variants. Papers and claims remain separate; architecture and instrumentation are reused.

Second, the resulting families define where the runtime should avoid repeated work:

• scarcity families reduce unnecessary activation and acquisition;
• locality families reduce movement and synchronization;
• reversible-change families prevent every event from rewriting durable state;
• maintenance families move repair and integration off the critical path; and
• structural offloading reduces the recurring cost of mature behavior.

For principle $j$, lifecycle acceptance uses the energy contract from chapter 80:

where $N$ is a dimensionless count of qualified events, event terms are joules per qualified event, and introduction, maintenance, and expected recovery are joules over the same observation horizon. A positive $\Delta E_j(N)$ is only an energy result; quality, risk, calibration, latency, and resilience must also remain inside their declared envelopes.

Evidence status

Speculative extensions

• Represent claims, principles, primitives, experiments, null models, and failures as a queryable versioned graph while keeping Markdown canonical.
• Track negative results as first-class edges so a rejected translation is not repeatedly rediscovered under another domain name.
• Estimate reviewer agreement on mechanism tuples before allowing a new principle ID.
• Search specifically for counterexamples: fields where the same pressure produces a different stable solution or where the recurring solution fails.
• Use contradictions between domains to generate new experiment regimes rather than averaging the difference away.
• Maintain silicon-native escape routes beside every principle: exact copying, direct addressing, typed storage, rollback, high-speed communication, and variable precision can change which operation is cheapest.

Failure modes

• Naming duplication: the same feedback loop appears repeatedly as a new organism-inspired component.
• False convergence: similar diagrams hide different sensed variables, causal operations, or timescales.
• Overcompression: a family becomes so broad that no ablation can isolate its mechanism.
• Evidence laundering: recurrence is treated as proof of independence, optimality, or transfer to AI.
• Null-model neglect: a familiar cache, scheduler, controller, or estimator is omitted because the biological story sounds novel.
• Interface drift: two variants share a principle ID while receiving different inputs, budgets, or evaluation envelopes.
• Positive-result bias: failed translations disappear, so the same proposal returns with a new metaphor.
• Double-counted savings: two principles claim the same avoided operation or compare against different baselines.
• Verifier laundering: checker acceptance is reported as truth about an informal statement without exposing formalization, axioms, dependencies, or the trusted implementation.
• Finite-to-universal leakage: a sampled numerical or test range is silently promoted from tested to proved.
• Correlated checking: producer and verifier share the parser, library, preprocessing, or defect that matters while being counted as independent.
• Taxonomy lock-in: stable IDs are mistaken for immutable scientific truth.

Measurable predictions

1. Independent reviewers given only normalized mechanism records agree on merge-versus-separate decisions more often than reviewers given titles and domain descriptions alone.
2. As domain coverage grows, the number of supporting claims per accepted principle rises faster than the number of principles; a near one-to-one ratio signals failed deduplication.
3. At least one proposed organism-specific mechanism is experimentally indistinguishable from an existing primitive at matched interface and cost and is removed rather than renamed.
4. Experiments built around shared principles reuse telemetry, null models, and failure regimes across more than one source domain.
5. Per-principle ablations attribute lifecycle savings to distinct avoided operations; overlapping savings disappear when all variants use one common baseline and boundary.
6. A principle promoted from recurrent evidence survives at least one regime derived from a domain outside the one that originally motivated its AI translation.
7. A typed propose–challenge–prove–check lifecycle reduces bad acceptance, localizes repair, and improves dependency invalidation beyond a generator plus checker alone after search, certificates, libraries, human effort, and joules are charged; otherwise Fixture F-004 retires the composition.

Scope

Define how a modular system acquires new capacity without running, training, or retaining every possible module for every event. Growth is admitted only for a measured capability gap, new capacity earns traffic in probation, and mature capacity remains eligible for merge, reopening, or retirement.

The central object is a capacity lifecycle, not a continuously expanding pool. Birth, routing, specialization, placement, consolidation, and removal are separate decisions with separate costs.

Biological observation

Developing nervous systems generate and reorganize more cells and connections than remain in mature circuits. In the studied mouse retinogeniculate system, relative activity and complement signaling changed microglial engulfment and retention of developing inputs (C-043). This supports activity-sensitive structural refinement by a slower maintenance process, without specifying one general pruning rule.

Protection is not necessarily permanent. Targeted extracellular and receptor interventions reopened specific forms of adult visual-cortex plasticity (C-044, C-045). The resulting engineering states are candidate, consolidating, protected, reopened, and retiring. A module can be stable without becoming impossible to revise.

Other biological systems expose control operations that recur at different scales:

• germinal-center affinity maturation combines variation, selection, expansion, and later protection of useful lineages (C-028);
• Physarum and fungal networks couple use-dependent reinforcement, exploration, fusion, and contraction to changing flow (C-027, C-034);
• ants can open reserve routes under crowding before throughput falls (C-035); and
• activity can recruit local energy production, alter resource placement, and change local vascular supply in scoped neural preparations (C-049–C-051).

These mechanisms are not interchangeable. Together they impose a useful system constraint: variation and reserve capacity consume resources; selection needs an independent test; frequently used structure may stabilize; unused or duplicated structure must be able to leave the hot path.

A defined microbial-community experiment adds a practical design tactic: adding candidates selected for missing functions repaired the tested community better than merely restoring organism count (C-056). The inverse problem also matters. Functional redundancy was associated with poor newcomer engraftment in two small reanalyzed human microbiome datasets (C-057). Mature capacity can resist both harmful and beneficial entrants, so an artificial newcomer needs protected evaluation traffic rather than permission from incumbent routing logits alone.

Proposed AI translation

The capacity lifecycle

flowchart LR
    subgraph birth["1 · Detect and birth"]
        direction TB
        gap["Persistent capability gap"] --> proposal["Clone · seed · compose"]
        proposal --> probation["Probation<br/>bounded state + budget"]
    end
    subgraph compete["2 · Route and specialize"]
        direction TB
        trial["Exploit · explore · reserve traffic"] --> evidence{"Useful, distinct,<br/>and affordable?"}
        evidence -->|"yes"| specialize["Local specialization"]
        evidence -->|"no value"| discard["Discard + retain result"]
        evidence -->|"duplicate"| merge["Merge or distill"]
    end
    subgraph lifecycle["3 · Consolidate or release"]
        direction TB
        gate["Maturity gate"] --> protected["Protected module"]
        protected --> monitor["Quality · traffic · cost · fragility"]
        monitor -->|"redundant"| retire["Drain + retire"]
    end
    probation --> trial
    specialize --> gate
    merge --> gate
    monitor -->|"new gap"| gap
    retire --> released["Released hot capacity"]

Editable source: ../assets/diagrams/structural-growth-routing.mmd.

The initial system contains modality encoders, predictive shared state, conditionally addressable experts, hierarchical routers, episodic and factual memory interfaces, and a declared reserve. Total addressable capacity may be large, but each event receives only a bounded route. Reserve capacity is stored, placed, and periodically tested; it is not free merely because it is inactive.

1. Detect a capability gap

Growth begins with a gap record, not a global loss spike. A valid record groups attributable episodes that existing routes fail in a consistent way and asks whether the failure is better explained by:

• missing capability;
• insufficient active compute or depth;
• missing context or memory;
• router error or capacity congestion;
• interference inside an existing module;
• distribution drift; or
• corrupted data, tools, or feedback.

The maintenance plane attempts the cheaper explanations first. New capacity is eligible only when the gap persists across resampling or recurrence, existing modules cannot absorb it inside their interference and resource bounds, and a candidate has a declared validation contract. This ordering prevents every hard example from becoming an expert.

2. Choose a birth operation

The proposal names both the new structure and what it is expected to repair:

Every birth creates a versioned provisional module. It cannot write the slow model, replace an incumbent, or claim permanent memory during probation. Initialization source, training episodes, expected role, resource ceiling, and discard path are recorded before it receives traffic.

Composing or admitting that module is itself a perturbation. The newcomer can pin shared activations, add router or queue pressure, alter normalization, create a gradient path, mutate shared cache state, or intentionally feed back. Before probation credit is assigned, compare the incumbent producer under the same input and seed with the connection absent and present. Cross connection state independently from shared CPU, memory and scheduler load so ordinary contention is not mislabeled as connection-induced back-action (C-1550, C-1558). Immutable snapshots, bounded queues, resource isolation, stop-gradient adapters and explicit feedback remain the minimum nulls. The full qualification and retirement rules are in Fixture F-027.

Birth order is another intervention, not a background detail. A newcomer that arrives first can be older, receive more examples, pre-empt finite capacity, modify shared state, facilitate a later module, or create a costly lock-in. Those mechanisms predict different remedies. The same task multiset and eligible module identity set must therefore be replayed across randomized orders with equal exogenous presentations, task-local update ceilings, capacity, optimizer, evaluator and lifecycle budgets. Realized active, consolidated, merged, or retired state remains an outcome. Routed acceptance remains a measured mediator and is equalized only in the exposure-cut cells. Canonical replay, ordinary scheduling, replay/EWC/OGD, fixed and capacity-matched MoE, random curricula, optimized curricula with their search cost, population-based training and quality-diversity search are the required null stack. The exact factorial and kill rules are in the history-conditioned succession contract.

3. Give the candidate probation traffic

The router divides admitted work into three explicit budgets:

• exploit traffic goes to the strongest validated route;
• exploration traffic compares plausible candidates on informative events; and
• reserve traffic preserves failover and tests paths that would otherwise decay unnoticed.

A newcomer receives a capped share of gap-relevant episodes plus matched control episodes. The control traffic reveals whether it learned a capability or merely a narrow identifier for the failure cluster. Incumbents cannot reduce the evaluation share through their own confidence, but the maintenance plane can stop the trial for quality, risk, latency, memory, or energy violations.

For event $x$, use a dimensionless routing objective

where $\widehat E(x)$ is estimated joules/event, $\widehat B(x)$ is estimated bytes/event across named memory and network boundaries, and $E_0$ and $B_0$ are declared reference scales with the same units. All $\lambda$ coefficients and losses are dimensionless. The physical joule and byte measurements remain separate reported outcomes; normalization does not turn them into task quality.

Balance keeps one expert from taking all traffic. Churn is applied only after a route has accumulated evidence of stable specialization; penalizing early movement would protect arbitrary initialization.

When module reports become strategic

Ordinary routing remains the default. Shadow-price feedback already represents scarcity under declared controller conditions (C-133), and auction or matching language adds nothing when the router can observe costs and every module shares the system objective. A market-like mechanism is in scope only when a persistent module holds decision-relevant private information, can improve its future traffic by misreporting, and faces a real opportunity consequence it cannot reset or evade.

That regime creates specific failures. Selection pressure on a visible metric can damage poorly measured substitute tasks (C-139); proper scoring needs an independently verified outcome and does not establish competence or causal contribution (C-140); peer agreement can reward shared error or collusion (C-141); and fixed-agent truthfulness does not solve false identities or a biased allocator (C-143).

Candidate 008 therefore begins with a cooperative applicability control that it should not beat. Only then does it introduce hidden costs, adaptive metric gaming, protected outcomes, entrants, identity resets, collusion, and allocator deviation. Withheld audits, lineage-bound consequences, protected entrant traffic, and replayable commitments remain only if they improve external task, risk, energy, and latency outcomes after their evaluation and storage costs.

4. Measure specialization rather than naming it

A candidate becomes useful when it improves a defined region of behavior while remaining distinguishable from existing modules. Evidence includes:

• causal improvement when the candidate is admitted and regression when it is ablated;
• reduced interference on incumbents or protected history;
• consistent advantage on held-out gap and recurrence episodes;
• a stable but non-exclusive routing region;
• calibration and rare-case behavior inside its declared envelope; and
• physical cost that remains inside its allocation.

Low routing entropy alone is not specialization: a router can collapse onto a module for the wrong reason. High activation diversity alone is not useful: a pool can fragment one capability across many expensive duplicates. The test is complementary causal contribution at a measured lifecycle cost.

5. Hand off to merge, protection, or retirement

At the end of probation, the candidate has three normal outcomes:

1. Discard. It adds no reliable capability. Preserve the negative result so the same proposal is not regenerated indefinitely.
2. Merge or distill. It reproduces an incumbent or several modules have converged on one operation. Build a compact branch, rerun intervention and recurrence tests, then drain duplicates.
3. Consolidate. It contributes a distinct reusable capability. Pass it to the maturity lifecycle for protection, reopening rules, structured pruning, and rollback.

A protected module remains monitored for traffic, unique contribution, physical placement, recovery, and newcomer exclusion. Persistent redundancy returns it to a merge-or-retire gate. Logical topology changes can be evaluated with Candidate 001, which charges reconfiguration, migration, reserve, controller, and recovery costs.

Conventional null models

The growth controller must beat ordinary ways of allocating or restructuring capacity, not only a frozen weak model:

Equalize initial capacity, maximum stored capacity, active work, optimizer updates, training examples, router information, tuning trials, migration bytes, validation work, and wall-clock opportunity. Charge candidate failures and discarded births. Otherwise growth buys more search while the null models are asked to solve the task in place. A stylized optimum such as a fixed branching exponent belongs in this null stack only inside its verified objective, transport law, and constraint regime (C-1489).

A topology change carries state

Changing an edge label is not the same as transferring a working system. The process-engineering audit makes the missing state visible: an installed path may hold inventory, energy, contamination, pending work, actuator authority, wear, calibration, maintenance obligations, and shared protection dependencies (C-501, C-510). Two endpoint configurations can both be feasible while the path between them is unsafe (C-512). Static branching relations do not remove that transition state (C-1489); fast allocation and slow structural adaptation must retain separate action, deficit, build, stranded-capacity, and recovery histories (C-1496).

For digital modules, use a typed transition inventory rather than pretending that bytes obey material conservation. For state class $k$,

where $I_k$ is bytes present at a named boundary, $A_k$ is admitted bytes, $D_k$ is deliberately deleted bytes, $X_k$ is exported bytes, and $R_k$ is internally replicated bytes over $[t_0,t_1]$. Each term is bytes and carries a provenance, validity, and ownership version. This is an accounting contract, not a claim that information is physically conserved (C-517).

flowchart LR
    P["Proposed graph version"] --> G{"State, cost, and safety gate"}
    I["Queued · cached · provisional state"] --> G
    A["Authority · dependencies · fallback"] --> G
    G -->|"admit"| T["Shadow and bounded transfer"]
    G -->|"reject"| K["Keep fixed graph"]
    T --> V{"Observed postcondition"}
    V -->|"pass"| N["New active version"]
    V -->|"fail"| R["Drain and roll back"]

Editable source: conservation-qualified-reconfiguration.mmd.

Candidate 001 therefore includes a physical stress track. It must beat fixed-graph adaptive control, multi-mode supervisory control, and offline redesign after installed reserve, transition state, flushing, downtime, maintenance, and rollback are charged. If those nulls tie it, “adaptive topology” describes an implementation choice rather than an efficiency mechanism (C-516).

Positional instruction, repair capacity, and structural fields

The developmental and regeneration residue sharpens the capacity lifecycle by separating operations that a generic “grow and repair” controller would hide. The claims are source-domain results; their engineering translations remain prospective and are tested only by the preimplementation F-022 fixture.

1. Keep repair instruction distinct from repair workforce. In the studied planarian experiments, differentiated muscle carried position-control expression while neoblasts supplied replacement capacity (C-1507). A modular system should therefore test a two-plane design: surviving task modules retain bounded reconstruction constraints, while a replaceable worker pool executes them. A worker reserve is not positional knowledge, and retained instructions are not proof that enough capacity survives. DEV-T02 compares this separation with complete checkpoint-and-log restoration and requires abstention when surviving instruction support is insufficient.
2. Treat positional memory as writable, gated state. Transient intervention rewrote persistent positional state during the scoped axolotl regeneration window, but not in comparable uninjured cells (C-1506). The translation is neither immutable metadata nor unrestricted self-editing: retain versioned local role state, open writes only inside an authenticated repair episode, validate against independent service evidence, and preserve rollback. DEV-T01 makes valid, locally corrupted, and common-mode-corrupted memory separate cases so saved reconstruction work cannot excuse confident wrong-role repair.
3. Separate source compensation, scale support, and receiver observability. Opposed BMP-related sources compensated for scoped perturbations in Xenopus (C-1508); DEV-T03 asks whether reciprocal sources beat distributed robust control after delay, stability, and double-source detection are charged. Dpp-gradient scaling in the studied wing disc had an experimentally exposed finite envelope (C-1509); DEV-T04 therefore requires an explicit beyond-support disposition rather than a scale-invariance claim. In the cited micropattern assay, edge access and receptor orientation changed which signal a cell could observe (C-1510); DEV-T05 compares local boundary-relative inference with exact distributed graph distance. A compensating field can still be unreadable at the receiver, and a readable field can still be outside its calibrated size range.
4. Consolidate paths and boundaries only after persistent evidence. Tension recruited and stabilized Myosin II in the scoped tissue experiments (C-1511), motivating capped, decaying load-path reinforcement rather than permanent reinforcement after a burst; DEV-T06 includes load reversal and event-triggered backpressure as controls. An actively maintained Myosin-II barrier limited compartment mixing (C-1512), motivating delayed, reversible interface hardening tested against a mature soft boundary in DEV-T07. Apparent enhancer redundancy improved robustness only across particular perturbations in the studied loci (C-1513); DEV-T08 accordingly tests partially input-diverse small gates against full duplicate controllers and preserves common-mode corruption as a hostile null.
5. Compose local patterning with coarse global context without conflating their roles. The limb-pattern study supports a plausible, non-unique reaction–diffusion interpretation modulated by global gradients (C-1514). DEV-T09 therefore requires local-only and global-only ablations and compares the composition with central constrained assignment. Chick perturbations also support local-feedback symmetry breaking, while the proposed molecular model remains non-unique (C-1515); DEV-T10 tests temporary organizer election against a complete randomized leader-election protocol, including duplicate detection and re-election.

These distinctions add three gates to structural growth: verify that surviving instruction is sufficient before allocating repair workers; verify that the field is both inside its support envelope and observable by the receiving module; and verify that reinforcement or boundary writes amortize their write, maintenance, reopening, and fallback costs. Every F-022 track remains NO_RESULT; a written protocol does not establish an efficiency advantage.

Competence-gated structural transition

A signal is not a complete command. Developmental experiments show that concentration, duration, position, stage, and receiver state can change the response to the same extracellular input (C-539–C-549). Commitment can later be redirected, but reopening is an intervention with selection, resource, integrity, and safety costs rather than free reversal (C-550, C-551).

For module $m$ and transition-contract version $v$, define the dimensionless admission predicate

$q_m(t)$ is typed local competence state; $\mathcal C_v$ is the permitted state region; $t$ is seconds; $W_v$ is a declared time interval in seconds; $e_m$ and evidence threshold $\theta_v$ are dimensionless; resource vectors $\mathbf r_m$ and $\mathbf r_v^{\min}$ are compared componentwise in their native units; and $R_{m,v}$ is a dimensionless predicate that a tested rollback path exists. The signal is eligible to create a provisional transition only when $G_{m,v}=1$.

stateDiagram-v2
    [*] --> Incompetent
    Incompetent --> Eligible: local state enters window
    Eligible --> Provisional: signal + evidence + resources
    Provisional --> Committed: validation passes
    Provisional --> Incompetent: reject + reset
    Committed --> Protected: structural postcondition
    Protected --> Reopening: explicit trigger + budget + rollback
    Reopening --> Provisional: bounded retrial
    Reopening --> Protected: abort + restore
    Protected --> Retiring: replacement or lost utility
    Retiring --> [*]: drain + retain evidence

Editable source: competence-gated-transition.mmd.

The fixture splits responsibilities instead of adding another candidate:

• Candidate 002 tests low-bandwidth signal decoding by versioned receivers;
• Candidate 009 owns admissibility, authority, dependencies, and invalidation;
• Candidate 010 owns provisional evidence before commitment and reset;
• Candidate 006 owns structural write, health, shadow, and physical reopening; and
• Candidate 014 owns support, stage, vintage, and the observations defining competence.

Reject the composition if an ordinary gate, schedule, staged verifier, and versioned migration procedure matches quality, interference, rollback, availability, bytes, latency, and joules. The fixture is useful only if the interaction among signal, receiver history, finite window, commitment, and reopening predicts failures those separate controls miss (C-562).

When the unit of adaptation changes

A cooperating set of modules is not automatically a higher-level unit. The claim becomes testable only after defining its boundary, child-configuration event, inherited state, within-lifetime state, descendant relation, collective performance, member-level incentives, and conflict-control cost. This separates aggregation from individuality and a demographic bottleneck from a reproducible founder boundary (C-282–C-295).

flowchart LR
    F["Versioned founder / propagule"] --> C["Reproducible collective configuration"]
    C --> V["Bounded within-lifetime variation + specialization"]
    V --> T["Collective task outcome"]
    T --> P["Partition collective gain vs member shortcuts"]
    P --> G{"Joint gain survives conflict cost?"}
    G -->|"no"| X["Use ordinary modular control"]
    G -->|"yes"| H["Select reproducible higher-level lineage"]
    H --> F
    V --> A["Audit · budget · sanction · reproductive separation"]
    A --> P

Editable source: conflict-bounded-unit-transition.mmd.

For collective lineage $k$, let $Z_k$ be task-native collective performance and $W_k$ its dimensionless admitted-descendant weight. The selection accounting is

The first term is change among declared collectives and the second is transmission change. A nested partition separately reports member-level shortcuts. The identity is accounting, not causal proof; group and descendant definitions are preregistered. Candidate 016 must beat typed modular systems, permissions and tests, clean versioning, external evaluation, routed experts, and ensemble/population selection after enforcement, false-sanction, interface, founder, reserve, and recovery costs.

Efficiency mechanism

For experts $i=1\ldots n$, let $g_i(x)\in\{0,1\}$ be the dimensionless event gate, $C_i(x)$ be executed operations/event under a declared precision, and $C_{\mathrm{router}}(x)$ use the same convention:

This separates addressable parameter capacity from executed work, as sparse mixture-of-experts systems demonstrate in specific implementations (C-003). It does not price parameter reads, dispatch, all-to-all communication, imbalance, cold starts, or maintenance.

The lifecycle energy per served event is therefore

where every $E$ term is in joules, $E_i$ is expert execution energy/event, and $N_{\mathrm{served}}$ is the number of events over the comparison horizon. Report stored parameter and optimizer bytes, bytes moved/event, latency, utilization, quality, calibration, and risk alongside energy.

Growth is efficient only if conditional execution saves more than candidate search, idle reserve, placement, validation, and later contraction consume. Lottery-ticket results show that competitive sparse subnetworks can exist in tested settings (C-012); they do not establish that this lifecycle discovers them or realizes energy savings on a target system.

Evidence status

Speculative extensions

• Let modules request a birth trial with a compact capability-gap certificate; maintenance allocates the trial, not the requesting module.
• Maintain seed modules at several parameter and precision scales so a new role need not start from the largest available structure.
• Learn placement jointly with specialization only after migration cost and rollback are measurable.
• Use recurrence-aware cold storage: retire a module from hot execution while retaining enough checkpoint and routing evidence to restore it if its regime returns.
• Allow two candidates to share an encoder or memory interface while keeping their update authority and resource accounts separate.
• Test local birth/retirement against periodic global architecture optimization under recurrent rather than one-way task sequences.

Failure modes

Measurable predictions

1. Capability-gap-driven births close held-out failure clusters with fewer admitted candidates and lower lifecycle energy than random birth, periodic fixed growth, and capacity-matched adapter baselines.
2. A protected probation share lets genuinely better newcomers establish causal value faster than incumbent-logit admission without increasing harmful promotions at the same validation budget.
3. Successful specialization produces a stable, non-exclusive routing region and positive unique ablation value; routing entropy alone predicts promotion less reliably.
4. Conditional growth reduces executed operations and bytes moved/event relative to a capacity-matched monolith while preserving quality, calibration, and rare-case performance.
5. Causal merge/retire gates preserve recurring and intervention capability better than usage- or magnitude-only pruning at matched hot capacity.
6. Reserve capacity improves recovery after faults or returning regimes enough to justify its stored bytes, idle energy, and periodic test traffic.
7. Joint routing and placement lowers communication energy only after migration, cold-start, and rollback work are included.
8. A collective lifecycle advances only when cost-adjusted between-collective selection and inherited capability persist under member shortcuts and turnover beyond ordinary lifecycle governance.
9. The complete lifecycle advances only if it improves the quality–risk–latency– energy–adaptability frontier over N0–N5; a parameter-count or FLOP reduction alone does not satisfy the prediction.

Scope

Grounding means that a model’s state is constrained by temporally ordered observation, action, and consequence—not that every concept must be reducible to pixels or motor commands. This chapter defines the trajectory record from which such state can be learned, how language attaches to it, and the tests that separate action-conditioned structure from passive multimodal correlation.

The central unit is a versioned episode containing asynchronous sensor events, commanded and realized interventions, exogenous events, and language with provenance. The model must predict and act under missing modalities, timing error, partial observability, and finite sensing energy. A representation is treated as grounded only to the extent that it supports held-out prediction, intervention, control, and reference at a declared uncertainty.

Biological observation

Biological perception is embedded in a closed loop. During active whisker exploration, measured contact mechanics predicted primary sensory-neuron firing better than whisker angle alone in the scoped mouse experiment (C-022). The result establishes that self-generated sensing mechanics can matter to the sensory code; it does not make embodiment necessary for every capability.

Other evidence constrains parts of the translation:

• sparse coding can learn localized structure from natural images in the model studied under C-002;
• latent target prediction can learn semantic image representations without pixel reconstruction under C-006;
• children in a controlled toy study targeted actions toward unresolved causal structure under C-062;
• pedagogical demonstration narrowed subsequent exploration in the scoped toy experiments under C-063; and
• causal transparency changed copying of irrelevant demonstrated actions in a comparative puzzle-box experiment under C-064.

Together these findings motivate aligned experience, active evidence collection, and separation of imitation from outcome learning. They do not yet establish the project-level hypothesis that embodied multimodal training yields more general physical concepts than comparable text-centric training. That claim remains speculative under C-007.

Proposed AI translation

Canonical trajectory record

One episode is

where $\tau$ is an episode record and all five components are immutable after publication; corrections create a new version.

Metadata $\nu$ contains episode and environment identifiers; train, development, or test split; actor or policy version; simulator or device version; clock domains; calibration versions; random seed when applicable; and parent episode for a counterfactual branch. Frames from one episode never cross data splits independently.

An observation event from modality $r$ is

where:

• $r$ is a modality identifier such as vision, audio, touch, proprioception, force, temperature, or depth;
• $k$ is an event index;
• $t^{\mathrm{cap}}_{r,k}$ is physical capture time in seconds;
• $t^{\mathrm{recv}}_{r,k}$ is the time in seconds at which the learner could first access the event;
• $v_{r,k}$ is the sensor value with the sensor’s declared unit, shape, encoding, and quantization;
• $m_{r,k}$ is a dimensionless missingness/status code;
• $q_{r,k}$ is a quality record with named quantities and units, such as exposure time in seconds or signal-to-noise ratio in decibels; and
• $p_{r,k}$ is a provenance pointer to source, calibration, transformation, and checksum records.

Capture and receipt time are both required. Offline alignment uses capture time; a deployable causal policy may use an event at time $t$ only when $t^{\mathrm{recv}}_{r,k}\le t$. This prevents a delayed sensor packet from becoming accidental future information.

An action event is

where $t_j^{\mathrm{cmd}}$, $t_j^{\mathrm{on}}$, and $t_j^{\mathrm{off}}$ are command, realized onset, and realized offset times in seconds; $a_j$ is the commanded action; $\widetilde{a}_j$ is the measured realized action; $\pi_j$ is the behavior-policy probability for a discrete action (dimensionless) or probability density for a continuous action (in reciprocal action-volume units) when it is known; and $p_j$ is its provenance pointer. Each action component declares its physical unit—for example newtons, newton-metres, metres/second, radians, or a discrete tool identifier. Command and realization are never silently substituted for one another.

An exogenous event $x_l\in\mathcal{X}$ records event and receipt timestamps $t_l^{\mathrm{event}}$ and $t_l^{\mathrm{recv}}$ in seconds, event type, measured magnitude and unit, and provenance. Examples include another agent’s action, an uncommanded collision, lighting change, or simulator reset. Known simulator state $s_t^*$ may be stored for evaluation in its declared physical units, but it is oracle information and is withheld from learned systems unless a baseline explicitly receives it.

A language event is

where the two timestamps are seconds, $w_n$ is text or a token sequence, $c_n$ identifies speaker and communicative role, $r_n$ is a set of referent pointers to objects, events, trajectory intervals, or external records, and $p_n$ is source provenance. A retrospective caption, an instruction available before action, a question, a report from another agent, and model-generated text have different roles even when their words match.

Observation is a versioned contract

A sensor record is not the latent state. Propagation, foreground, exposure, instrument response, background, calibration, reconstruction, threshold, and selection intervene before a downstream model sees it (C-218–C-221). For latent target $\theta$, nuisance state $\eta$, measurement $y$, and response version $H_v$,

where $y$ and $H_v$ have matching declared units and $\epsilon$ follows a declared noise and background model. If the analyzed record was selected by event $S=1$, inference is conditional on that event:

The trajectory record therefore carries response and calibration version, exposure, selection state, detection power for non-detections, reconstruction choice, and data vintage. Cross-modal fusion also retains association uncertainty and shared dependencies; a different sensor type is not evidence of conditional independence (C-224).

Candidate 014 tests whether propagating this contract with each claim adds value beyond a complete typed, calibrated, lineage-aware, selection-aware, and simulation-checked evidence stack. The candidate must abstain on response null spaces and exact degeneracies rather than buy more observations that cannot identify the missing direction (C-229).

Derived observations inherit an interface and a support

A ratio, difference, derivative, normalized innovation, or latent coordinate is an observation operator, not a neutral rewrite of the raw signal. The operator record must name its input interface, reference state, causal window, positive-domain and saturation support, units, timestamp, calibration and any absolute information it removes. The statistic can legitimately change across the external-input, internal-state, downstream-readout and action interfaces: the Wnt/β-catenin result is an internal fold readout (C-1543), while the two NF-κB studies address different upstream and downstream arrows (C-1544, C-1545).

An invariant observation can make a transformed scale unidentifiable from that interface without proving that the information is universally destroyed. Adding a calibrated absolute observation or changing the output map can change recoverability (C-1549). The trajectory record therefore retains both the derived value and the smallest raw or calibrated side channel required by registered downstream targets.

Missingness is observed state, not a zero tensor

For modality $r$ and decision time $t$, define $M_t^{(r)}\in\{0,1,2,3,4\}$ as a dimensionless status:

The raw value is not imputed before status is retained. If an imputed value is needed, the model receives both the imputation and $M_t^{(r)}$. Training dropout is labeled as synthetic missingness and is sampled in contiguous outages as well as isolated frames. Confirmatory tests include natural missingness, failure, corruption, and combinations not seen during training.

Missingness may itself be informative—for example a tactile sensor becomes available only after contact. Evaluation therefore distinguishes performance gained from legitimate availability structure from shortcut prediction of a label or environment identifier.

Temporal alignment contract

Each modality has a calibrated clock offset $\delta_r$ in seconds and residual jitter scale $\sigma_r$ in seconds. The corrected capture time is

Alignment does not mean forcing all modalities onto one frame rate. Encoders consume timestamped events or aggregate them inside declared causal windows. For a decision at time $t$, the available history is

where $\mathcal{H}_t$ is a set of records, not a unit-bearing scalar, and $\alpha_j^{\le t}$ contains only action fields measured by time $t$; a future realized offset or actuator trace is not revealed with the earlier command. The equation admits a language event after its end time; a streaming system may instead add explicitly timestamped token prefixes. Outcome targets may use later events, but the state used to choose an action may not.

Alignment robustness is tested by adding known per-modality offsets and jitter, dropping clock-synchronization messages, and withholding one calibration version. A mechanism that works only at exact simulator step boundaries has not learned a robust temporal relation.

Observation, intervention, and consequence

An action token is useful only when its consequence can be distinguished from background change. Training retains four separable cases:

1. passive observation with no controlled action;
2. policy-chosen action with known or estimated behavior probability $\pi_j$;
3. randomized or scripted intervention with known assignment; and
4. paired simulator branch from the same saved initial state with one changed action.

Only the third and fourth directly support an interventional comparison. Policy-chosen logs are confounded by the policy’s state and require a declared system-identification, propensity, or model-based assumption. Exogenous events remain separate from realized action so the model cannot credit itself for an outside cause.

For a horizon $\Delta>0$ seconds, an action-conditioned predictor estimates

where $z_t\in\mathbb{R}^d$ is a dimensionless latent state of width $d$, $\theta$ is the parameter vector, $\widetilde{a}_{[t,t+\Delta)}$ is the realized action sequence over the interval with per-component physical units, and $M_{\le t}$ is missingness history. The predictor is not called causal merely because action is an input. Causal interpretation depends on the data case and evaluation intervention.

This realized-action form is used for training and retrospective scoring. A prospective rollout conditions on a candidate command $a$ and integrates over the calibrated distribution of actuator delay and realized action; it does not assume that a command is executed exactly.

Predictive state and action loop

The observation encoder forms

where $E_\phi$ is an encoder with dimensionless parameters $\phi$ and $z_t$ is dimensionless. Separate heads predict future latent state, selected sensor targets, task outcomes, and calibrated uncertainty. Pixel or waveform reconstruction is used only when the task requires that detail; normalized latent prediction follows the narrower evidence in C-006.

For target-event set $\mathcal{G}$, a probabilistic prediction loss is

where $|\mathcal{G}|$ is a target count, $v_g$ is a declared target under a declared quantization or likelihood on dimensionless normalized coordinates, $t_g$ is target context time in seconds, $\Delta_g$ is its prediction horizon in seconds, and $\mathcal{L}_{\mathrm{pred}}$ is nats/target. A normalized latent-distance loss may be reported in dimensionless units, but it is not added to negative log likelihood without a declared conversion weight.

At a decision, the system may act on the world, request a modality, ask a question, consult sourced memory, or wait. Let $b\in\mathcal{B}_t$ denote one such acquisition or intervention choice. A conventional one-step value-of-information policy chooses

where $V$ is expected task-utility improvement in declared utility units, $E(b)$ is full sensor, communication, and compute energy in joules, $T(b)$ is added latency in seconds, $\lambda_E$ has units utility/joule, and $\lambda_T$ has units utility/second. wait is an explicit zero-acquisition option. This is the strongest conventional null for claims that prediction error should route sensing or compute, as described under P-007 and in the engineering analogue audit.

Controlled observability and plant binding

The observation operator is partly controlled. Sensor pose, locomotor microstructure, emission timing, sampling density, and contact can change which state directions are observable (C-589–C-592). At the same time, body, tool, attachment, payload, compliance, and contact change how commands become motion and sensory consequences (C-599, C-603–C-605).

Separate the immediate task controller from the sensing controller:

where $b_t$ is a dimensionless belief state, task action $u_t$ retains its physical units, sensing action $a_t^{\mathrm{sense}}$ retains pose, emission, sampling, or motion units, $\mathcal O_v$ is observation-contract version $v$, and $\Pi_v$ is the controller–plant binding

with body/tool identity $B$, task $T$, attachment/payload $A$, contact model $C$, sensor/actuator calibration $S$, task Jacobian $J$, impedance/passivity envelope $Z$, delay model $D$, and safety/authority envelope $E$. Each field is typed, versioned, and linked to the controller, estimator, data, and tests that depend on it.

flowchart LR
    G["Task goal + risk envelope"] --> TC["Task controller"]
    B["Belief + uncertainty"] --> TC
    B --> SC["Sensing controller"]
    TC --> U["Motor command + reflex settings"]
    SC --> A["Pose · emission · sampling action"]
    U --> P["Versioned body · tool · actuator · contact plant"]
    A --> P
    P --> W["World + medium"]
    W --> O["Propagation + sensor physics"]
    P --> O
    O --> Y["Versioned observations"]
    U --> E["Action record / efference copy"]
    Y --> X["State + consequence estimator"]
    E --> X
    X --> B
    V["Plant · attachment · calibration · delay · impedance · safety version"] -.-> P
    V -.-> X
    V -.-> TC
    H["Health · passivity · wear · energy · contact"] -.-> TC
    H -.-> SC
    Z["Counterfactual swap + selective invalidation"] -.-> V

Editable source: controlled-observability-plant-binding.mmd.

Two held tests share this record:

1. Controlled observability. Keep the task policy fixed while a sensing policy maintains information through pose, emission, or sampling. It must beat fixed excitation, random acquisition, one-step expected value of information, active SLAM, observability MPC, and dual control at equal action, time, risk, compute, and energy (C-596).
2. Counterfactual plant binding. Swap one body, tool, attachment, payload, sensor, delay, contact, or impedance field at a time. Dependencies should predict which estimator/controller state is safe to reuse and which must be invalidated. It must beat reset/retraining, unversioned adaptation, model banks, gain scheduling, operational-space impedance control, and adaptive MPC (C-606).

This composition avoids two attribution errors. Passive mechanics and compliance are credited before a controller (C-597); an action-correlated internal signal is not assumed to be a complete accurate forward model (C-601).

Opportunity- and history-qualified adaptive action

Comparative cognition adds a sharper condition to the observation contract: the same apparatus does not create the same task when bodies, sensors, prior training, rewards, demonstrators, handlers, or feasible actions differ. The evidence in C-804–C-841 spans tool manufacture, causal transfer, future preparation, event memory, uncertainty control, imitation, teaching, exploration, negative transfer, and central-versus-peripheral control. Its durable contribution is not a species ranking. It is a requirement to record the opportunity that made an outcome possible.

Extend each episode with

where $X$ is task and apparatus state, $S$ is the sensory information actually available, $A$ is the feasible realized-action set, $H$ is prior training and social exposure, $R$ is reward and stopping state, $C$ is the intervention and control set, $\tau$ is elapsed time in seconds, and $U$ is the sampling unit. The detailed definitions and units are maintained in the opportunity/history mathematics.

This record prevents several category errors:

1. Terminal success is not transfer. First-trial response after a material, geometry, cue, relation, body, or history change is retained separately from later test-time learning.
2. One copied outcome is not one copied process. Demonstrations are decomposed into action form, trajectory, end state, location, demonstrator, and functional relation. Ghost and result-only controls identify which component changed the learner.
3. A future-directed action is not a unique memory mechanism. Value tables, successor representations, planning, and retrieval remain nulls for delayed selection and resource reservation.
4. Adaptive checking is not a unique uncertainty mechanism. Public difficulty cues, response strength, calibrated confidence, selective prediction, and ordinary value of information must be beaten.
5. Local control is not independent control. Body mechanics, local sensing, peripheral feedback, central goal selection, override latency, and message traffic are measured separately.
6. Repeated trials are not independent agents. Episode, learner, dyad, group, site, rearing or training cohort, and population remain distinct sampling levels.

flowchart LR
    task["Task · apparatus · hidden regime"] --> channel["Actual sensory channel + receipt time"]
    body["Body · tool · feasible action set"] --> channel
    history["Training · reward · social exposure · handler cues"] --> state["Opportunity/history-qualified state"]
    channel --> state
    controls["Intervention · ghost · cue reversal · history swap"] --> task
    state --> policy["Act · inspect · abstain · retrieve · communicate"]
    policy --> plant["Realized action through compliant plant"]
    plant --> outcome["Outcome + first-trial transfer + safety + cost"]
    outcome --> diagnosis{"Which relation transferred?"}
    diagnosis --> affordance["Affordance · causal model · search"]
    diagnosis --> memory["Event · semantic · prospective memory"]
    diagnosis --> social["Action form · result · location · convention"]
    diagnosis --> control["Central · local · mechanical contribution"]
    nulls["RL · POMDP/MPC · retrieval · VOI · imitation · hierarchical control"] --> diagnosis
    diagnosis --> revision["Retain · narrow · reject · revise"]
    revision --> history

Editable source: opportunity-history-qualified-action.mmd.

Fixture F-003 turns the contract into eight separable tracks: manufacture transfer, causal inversion, future resource reservation, event memory, costed uncertainty, copy-target decomposition, social acquisition, and local/central/mechanical control. It compares the full composition with learned affordances and search, model-free and model-based RL, POMDP/MPC, semantic and episodic retrieval, selective prediction and value of information, imitation learning, central replay, and hierarchical control under one lifecycle boundary.

The fixture earns architectural credit only when opportunity qualification changes a conclusion, exposes a reproducible confound missed by the existing stack, or produces robust transfer beyond the complete nulls. Otherwise the episode contract remains useful instrumentation and the additional mechanism is retired.

Grounded learning loop

flowchart LR
    world["World and body"] --> events["Aligned multimodal<br/>events"]
    events --> model["Grounded predictive<br/>state"]
    language["Language and<br/>sourced knowledge"] <--> model
    model --> assurance["Uncertainty and<br/>provenance"]
    assurance --> decision{"Act, query,<br/>or wait?"}
    decision --> action["Bounded action or<br/>sensor request"]
    action --> world

Editable source: ../assets/diagrams/grounded-learning-loop.mmd.

Language attaches to grounded state

Language serves at least four different functions:

1. reference: name an object, relation, action, event, or trajectory span;
2. instruction and query: change the agent’s task or request information;
3. compression and communication: summarize learned regularity for another time, module, or agent; and
4. testimony: introduce facts and abstractions not available through direct sensorimotor experience.

Reference is trained from the $r_n$ links in language events, including negative and ambiguous referents. Instructions are available only from their receipt time and are treated as interventions that may change action. Testimony retains source, version, trust policy, and retrieval trace; it is not forced into a fictional physical observation. A concept can therefore combine direct experience, action–outcome evidence, and sourced symbolic knowledge without erasing which support came from where.

The staged design learns predictive physical state before or alongside a small language adapter, then compares late attachment, joint training, and language-first training under equal data and capacity budgets. Physical probes use paraphrases and unseen names; language probes include relations with no direct sensory referent. Demonstrated action and demonstrated outcome are separate targets, following the distinction motivated by C-064.

Uncertainty and provenance contract

Each prediction exposes a distribution or calibrated interval at the level at which action is chosen. Report negative log likelihood in nats/target, Brier score for declared categorical events, interval coverage in percent, and calibration error with binning fixed before evaluation. Calibration is broken out by modality status, prediction horizon, intervention type, environment, and linguistic versus direct evidence.

A single confidence scalar does not replace modality-specific uncertainty. The system records whether uncertainty arose from noisy observation, missing modality, disagreement among plausible dynamics, unseen composition, ambiguous language, or untrusted testimony when the estimator supports that distinction. If it does not, the output remains an undifferentiated predictive uncertainty.

Every training and evaluation event resolves through $p_{r,k}$, $p_j$, or $p_n$ to:

• raw or generated source identifier and checksum;
• simulator, device, actor, and policy version;
• capture and receipt clock domains;
• calibration and transformation chain;
• missingness and corruption operations;
• split assignment and counterfactual parent; and
• license or access policy where applicable.

Answers and plans can then cite direct observation, intervened outcome, retrieved testimony, or model rollout separately. Provenance does not make a prediction correct; it makes its evidential path inspectable and reversible.

Strongest conventional baselines and nulls

B7 is an upper bound, not a superiority baseline. Comparisons match trajectory seeds, raw sensor exposure where logically possible, model capacity, optimizer updates, action count, sensor acquisitions, context bytes, latency ceiling, and lifecycle energy. The active-data question and the action-conditioning question are evaluated separately: hiding action tokens in the same log is not the same experiment as allowing an agent to collect a different log.

Staged experiment

Stage 0 — Data and clock qualification

Build a deterministic replay harness before training a world model. It must reproduce event order, verify checksums and split isolation, expose capture and receipt time, reconstruct commanded versus realized actions, and audit every missingness code. Inject known offsets, jitter, packet delay, dropped intervals, and actuator lag. Reject a dataset version when these perturbations cannot be recovered or bounded from its metadata.

Stage 1 — Passive action-conditioned prediction

Use a controlled compositional environment with vision, audio, proprioception, touch/force, and optional depth. Objects vary independently in shape, color, mass, surface friction, and containment. Actions include look, approach, push, lift, rotate, drop, occlude, and tool contact. The held-out split contains:

• unseen combinations of known object properties;
• new visual textures and camera placement;
• changed mass or friction under familiar appearance;
• action delays and partial actuator failure; and
• one-modality and multi-modality outages.

Training, development, and test split by episode seed, object combination, and environment configuration—not by frames. Simulator state $s_t^*$ is stored for evaluation and B7 only.

Train B1–B5 and the candidate predictive state on the same logged trajectories. Compare intact realized actions with actions masked, time-shifted, and permuted within episode. Primary outcomes are future-state and sensor NLL in nats/target, physical-property probe error in declared units, held-out intervention error, control success percent with a frozen small controller, calibration, bytes, and joules/qualified episode.

Stage 2 — Paired interventions and active acquisition

From a saved initial state, run paired branches differing in one action or sensor request while holding the environment seed fixed. Separately allow each active policy the same maximum world actions, sensor acquisitions, wall time, and energy. Compare random exploration, entropy seeking, one-step EVSI, and the candidate policy.

Measure prediction change in the correct physical direction, intervention effect error, successful disambiguations/action, downstream task utility, unsafe interventions, sensor byte, action energy, total joules, and latency. This stage tests the open issue in C-022 and the targeted-exploration translation from C-062.

Stage 3 — Language attachment

Attach language using trajectory-span and referent links. Compare late attachment to a frozen physical core, joint training from initialization, and language-first pretraining followed by grounded data. Equalize language tokens, grounded episodes, trainable parameter count, optimizer updates, and tuning budget.

Evaluate new names for known referents, paraphrased instructions, ambiguous reference, imitation versus outcome emulation, physical questions about unseen compositions, testimony-only facts with changed source versions, and removal of language at control time. Report whether language improves sample efficiency without becoming the only route to physical success.

Stage 4 — Asynchronous physical transfer

Move the frozen comparison to one named real sensorimotor platform. Preserve native clocks, packet delay, actuator mismatch, calibration changes, occlusion, and sensor failure. Recalibrate energy at device and node boundaries using the energy evaluation contract. Simulation-to-real transfer, recovery, and negative results are reported by intervention and missingness stratum rather than one average.

The real platform adds two contracts that simulation accuracy cannot supply. First, estimation is qualified by integrity. An illustrative position protection level is

where $PL$, position standard uncertainty $\sigma_{\mathrm{pos}}$, and alert limit $AL$ are in metres, while $K_{\mathrm{int}}$ is dimensionless and tied to a declared fault and risk allocation. A learned confidence score does not become integrity without bias bounds, dependence, coverage, and time-to-alert (C-446).

Second, a controller handoff must preserve exclusive effective authority while the world keeps moving. The transfer binds the controlled effect, epoch, outgoing and incoming controller, actual and pending mode, observation basis, headroom, time-to-boundary, faults, outstanding commands, fallback reachable set, expiry, acknowledgement, atomic activation, and observed postcondition.

stateDiagram-v2
    [*] --> StableOwner
    StableOwner --> TransferProposed: effect + epoch + observation
    TransferProposed --> TransferRejected: stale / unsupported / not ready
    TransferProposed --> Armed: receiver ready + fallback reachable
    Armed --> NewOwner: atomic activation + old authority revoked
    Armed --> Fallback: timeout / fault / margin exhausted
    NewOwner --> StableOwner: postcondition verified
    NewOwner --> Fallback: postcondition fails
    TransferRejected --> StableOwner
    Fallback --> StableOwner: recovery + explicit reassignment

Editable source: asynchronous-authority-transfer.mmd.

This is a refinement of Candidate 012, not a new autonomy layer. It must beat mode annunciation, interlocks, epoch leases, runtime assurance, explicit arbitration, Candidate 014, and the recoverable-initiative contract under split-brain, stale-command, no-owner, delayed-link, common-mode, and failed- postcondition injections (C-447–C-461).

Efficiency mechanism

Grounding adds sensors and interaction, so its efficiency claim is conditional: the learned state must avoid more future sensing, compute, data movement, or failed action than it costs to acquire and maintain.

The candidate levers are:

• event-driven encoding: update modality state on timestamped change rather than resampling every channel at the highest rate;
• latent prediction: predict task-relevant state without reconstructing every high-bandwidth detail;
• local preprocessing: reduce raw sensor traffic near the source while retaining calibration and uncertainty metadata;
• conditional acquisition: power, transmit, or process another modality only when its expected utility exceeds energy and latency cost;
• temporal reuse: carry forward stable state with an uncertainty increase instead of re-encoding unchanged observations; and
• language compression: communicate grounded referents and relations when sending the underlying sensor history is unnecessary.

Every experiment reports sensor-on time in seconds, sensor energy in joules, raw and processed bytes by boundary, acquisitions/episode, world actions/episode, encoder and world-model operations, latency in seconds, controller energy, maintenance/calibration energy, and total lifecycle joules/qualified episode. A lower model FLOP count is not a grounding-efficiency result when sensor collection or interaction dominates.

The relevant efficiency comparison is a frontier over task utility, risk, latency, energy, and physical interaction. One-step EVSI is the minimum null for conditional acquisition. A passive model trained on a larger recorded dataset is an additional null when interaction energy, not online autonomy, is the proposed benefit.

Evidence status

• Sparse natural-image representation under C-002: established for the cited model and images.
• Joint-embedding image prediction under C-006: established for the cited image experiments.
• Active sensor mechanics under C-022: established for the measured whisker task.
• Targeted exploration under C-062, pedagogical narrowing under C-063, and imitation/emulation difference under C-064: established for the scoped behavioral tasks.
• A generally useful predictive-coding abstraction under C-005: plausible, not a complete grounding theory.
• Robust general physical concepts from the integrated multimodal curriculum under C-007: speculative.
• Navigation integrity, protection-envelope scope, fault-stage separation, degraded/fallback distinctions, and assurance boundaries under C-445–C-460: established or plausible within the audited vehicle domains.
• Validated asynchronous authority transfer under C-461: speculative.
• Opportunity, history, causal-transfer, event-memory, social-learning, and distributed-control observations under C-804–C-840: established, plausible, or disputed only within their declared tasks; the integrated action contract under C-841 is speculative.

Speculative extensions

Controllability curriculum

Order early tasks by what an agent can reliably change and observe: persistence and contact, motion and occlusion, material response, tools and other agents, then linguistic abstraction. Curriculum order becomes an ablation rather than an assumed developmental law.

Learned morphology and sensor placement

Jointly adapt sensor placement, body parameters, and control only after the fixed-body grounding loop is understood. The simulation evidence under C-029 motivates the question, while lifecycle energy and reality-gap robustness remain required outcomes.

Socially supplied interventions

Treat demonstration, correction, question answering, and pedagogy as actions that change the learner’s evidence policy. The system can preserve uncertainty after a demonstration and deliberately test affordances not shown, rather than assuming instruction is exhaustive.

Counterfactual branch memory

Store compact paired branches from matched initial states as high-value causal records. Promotion into long-term memory depends on reuse and provenance, not on intervention novelty alone.

Failure modes

Measurable predictions

H-G1 — Action-conditioned consequence

With identical logged sensor events, intact realized actions will improve held-out intervention NLL and physical-effect error over masked, shifted, and permuted actions. If only reconstruction or in-distribution probe accuracy improves, the result does not support action grounding.

H-G2 — Temporal alignment

Models trained with explicit capture/receipt time and calibrated clock uncertainty will degrade more gradually under held-out offset, jitter, and packet delay than fixed-step concatenation. The comparison reports error as a function of injected milliseconds and identifies the offset at which the quality envelope is crossed.

H-G3 — Missing modalities

Typed missingness plus modality-specific predictive heads will preserve a better quality–calibration frontier under contiguous sensor outages than zero imputation or independent encoders fused only at the output. Performance is reported separately for absence, failure, occlusion/corruption, and unknown status.

H-G4 — Active evidence collection

At the same action, acquisition, latency, and joule budgets, a learned active policy will require fewer interventions to identify held-out mass, friction, containment, or causal relations than random and entropy-seeking exploration. It supports a new mechanism only if it also improves beyond calibrated one-step EVSI; otherwise EVSI becomes the accepted implementation of P-007.

H-G5 — Compositional physical transfer

Action-conditioned predictive state will yield lower sample count to a fixed control-success threshold on unseen property combinations and dynamics than text-only, passive multimodal, and behavioral-cloning baselines of matched capacity. The threshold, maximum episodes, and unsuccessful runs are fixed before evaluation.

H-G6 — Language attachment

Late or joint language attachment with explicit referent and provenance links will improve new-name, paraphrase, and testimony-version tests without reducing pre-language physical probes beyond a preregistered equivalence margin. If language-first training alone matches intervention transfer at equal grounded episodes, the additional grounding curriculum has not earned its cost.

H-G7 — Calibrated abstention

Prediction uncertainty will increase under unseen interventions, multi-sensor failure, and ambiguous reference, and calibrated abstention or acquisition will reduce high-cost physical errors at a declared false-abstention rate. Average confidence without stratum calibration does not satisfy this prediction.

H-G8 — Lifecycle efficiency

Conditional sensing, event-driven updates, and latent prediction will reduce sensor byte, processed byte, and node joules/qualified episode at matched task utility, risk, and latency relative to always-on sensing and fixed-rate encoding. The claim is rejected if the confidence interval for lifecycle energy includes the preregistered equivalence margin after collection, calibration, language, maintenance, and failed-action costs are included.

H-G9 — Integrity-qualified authority transfer

Under delayed links, boundary misclassification, degraded sensors, outstanding commands, common-mode faults, and failed postconditions, a validated transfer record will reduce double-owner, no-owner, stale-command, and unsafe-boundary time beyond mode annunciation, interlocks, epoch leases, runtime assurance, ordinary arbitration, and the HCI/observation contracts at equal sensing, reserve, communication, training, review, latency, and joule budgets. A tie merges its useful fields into those conventional mechanisms.

H-G10 — Opportunity-qualified transfer

At equal episodes, interventions, action feasibility, search, storage, communication, human effort, and lifecycle joules, explicit opportunity and history state will improve first-trial functional transfer and calibration under held material, causal-relation, body/tool, acquisition-history, and social-channel changes relative to an otherwise identical learner. The claim is rejected if the gain disappears after actual sensor cues, feasible actions, prior training, reward schedule, handler/site, or later test-time learning are modeled, or if the complete affordance/search, planning, retrieval, VOI, imitation, replay, and hierarchical-control stack matches the result.

Promotion requires H-G1 plus at least one transfer result from H-G3–H-G6, valid uncertainty under H-G7, and no lifecycle contradiction under H-G8. Stage-4 authority-transfer claims additionally require H-G9. A comparative-cognition-derived transfer claim additionally requires H-G10. A positive result remains scoped to its trajectory schema, intervention family, environment shift, hardware boundary, and evaluated language role.

Scope

This chapter turns the sports expertise, adaptive performance, and team coordination audit into a system-wide performance contract. A result is interpretable only when it travels with the information that was available, the actions that were feasible, the history that produced the policy, the opponent and team it faced, the feedback it received, its current resource and damage state, and the selection process that determined which systems were observed.

The contract constrains:

1. sensing and observation provenance;
2. routing and allocation under deadlines;
3. action and closed-loop control;
4. memory of practice, opponents, damage, and recovery;
5. curriculum and online adaptation;
6. assurance, degradation, and staged return;
7. coordination and communication; and
8. population-level evaluation and lifecycle accounting.

It connects the sensorimotor grounding chapter, sparse predictive computation, memory and consolidation, system synthesis, and the energy model. Its executable specification is Fixture F-006, with notation and derivations in the mathematical contract.

Biological observation

Perception is coupled to a possible response

Experts can extract predictive information from early opponent kinematics, but the useful cue depends on the task, feature, opponent, prior, viewing window, and response mode (C-926–C-930). Predicting a label, moving a joystick, initiating a partial movement, and intercepting a physical event under its real deadline are therefore different experiments. Temporal or spatial occlusion can locate when or where information becomes useful; it does not identify a unique representation.

Training transfers more reliably when relevant information and action coupling are preserved, but “representative” is not synonymous with visually realistic (C-931–C-932). Timing, feasible action, opponent response, consequence, and resource state can matter even when surface appearance changes little.

Practice is not one outcome

Practice variability can improve learning when it explores a relevant task dimension at a useful challenge level; the effect is not universal, and larger perturbations do not automatically help (C-933–C-936). Accumulated practice correlates with expertise, yet amount, type, opportunity, selection, survival, access, and retrospective classification are entangled (C-937–C-939).

This creates three separate questions:

1. Does performance improve while repetition, instruction, and feedback are present?
2. Does the change remain after a declared delay without that scaffold?
3. Does it survive a change in cue, opponent, body, task, rule, deadline, or resource state?

The answers are practice performance, delayed retention, and transfer. They must not be substituted for one another.

Performance follows a frontier, not a single score

Movement speed and accuracy trade off under task-specific conditions, and throughput can appear stable while speed and error move in opposite directions (C-940–C-941). Risk, physical work, latency, and recovery add further axes. Optimizing only accuracy can hide slower action; optimizing only latency can hide errors or unsafe events; optimizing only completed work can hide depletion.

Fatigue changes the state from which an action is controlled. Opponent behavior, remaining work, feedback, recovery, and sleep can alter output, but effects remain task- and outcome-specific (C-942–C-948). Internal load, external load, readiness, damage, and injury risk are different constructs. A workload ratio or a single measurement cannot certify an individual system's safe capability (C-949–C-950).

Recovery and return are staged and reversible

Return to participation, return to the full task, and sustainable recovery of prior performance form a staged continuum. Time and functional evidence can carry nonredundant information, while an apparently high function score can be misleading after early return (C-951–C-953). Each promotion exposes the recovering system to more load and therefore creates new evidence. Deterioration must permit regression to a safer stage.

Teams coordinate through information and adjustment

Joint practice can increase predictive knowledge, and shared displays can change later team state. Yet synchrony is only an observable; similar movement can follow common input without reciprocal coordination (C-954–C-957). A coordination claim needs perturbation, lagged compensation, reduced task error, turnover, and cross-play. A shared-information claim additionally needs records of messages, observations, beliefs, and their discrepancies.

Deceptive behavior exploits an observer's cue policy. Expertise can reduce some susceptibility without removing it, and confidence may rise while the judgment becomes wrong (C-958–C-960). Feedback effects are likewise conditional on content, timing, task, autonomy, and whether the outcome is immediate performance or later learning (C-961–C-962).

Selection changes the population it measures

Relative age, maturity, present capability, coaching access, playing time, opponent quality, and later observation are coupled by selection. A talent system can amplify early differences and then mistake the resulting population for evidence that its original ranking was correct (C-963–C-965). Excluded systems have missing counterfactual careers precisely because the selection policy withheld the experience needed to produce them.

Finally, metabolic estimates and wearable estimates do not equal complete system energy. Human attention, facilities, sensors, equipment, computation, recovery, failed trials, maintenance, medical or safety work, and displaced opportunity alter the efficiency comparison (C-966–C-968). The transferable result is therefore an evaluation contract (C-969).

Proposed AI translation

Carry the state that makes a comparison valid

For agent $i$ in episode $e$, preserve

where:

• $X_e$ is physical state, task, rules, deadline, consequence, and hidden regime;
• $O_{i,e}$ is information actually received, including source, support, latency in seconds, occlusion, noise, loss, and calibration;
• $A_{i,e}$ is the feasible action set under the current body or actuator, equipment, authority, rate, range, and safety limits;
• $H_{i,e}$ is timestamped practice, feedback, opponent, teammate, damage, exclusion, reward, and opportunity history;
• $M_e$ is opponent and teammate composition, role, policy history, turnover, and communication topology;
• $F_{i,e}$ is feedback identity, information in bits, delay in seconds, and provider;
• $R_{i,e}$ is the resource and fatigue vector in its native units;
• $D_{i,e}$ is damage or fault state, diagnostic uncertainty, protected envelope, and return stage;
• $G_{i,e}$ is the selection and opportunity policy;
• $C_e$ is the intervention, paired control, counterfactual seed, and stopping rule;
• $U_e$ is the sampling unit, such as action, episode, agent, dyad, team, site, or cohort; and
• $B_e$ is the complete ceiling in events, bytes, seconds, person-hours, joules, damage, unsafe events, replacements, and opportunity.

For method $m$ and literal outcome $k$, the comparison target is

where $Y_k$ uses the registered unit for outcome $k$. If a baseline and a new method receive different observations, actions, histories, opponents, feedback, resources, or selection opportunities, $Q_{m,k}-Q_{b,k}$ does not isolate method $m$ from baseline $b$.

Treat representativeness as an inspectable vector

For training distribution $P_{\mathrm{tr}}$ and target distribution $P_{\mathrm{te}}$, record

where the components compare actual observation $O$, feasible action $A$, deadline and consequence $T$, teammate/opponent state $M$, feedback $F$, resource state $R$, damage and return state $D$, and selection policy $G$. Each $d_z$ is a declared divergence between the corresponding training and target distributions: dimensionless for a statistical divergence or in the declared ground-cost unit for optimal transport. The components remain visible; one weighted “realism” number would conceal which relation transferred.

Keep outcomes separate

Separate useful exploration from noise

For action variable $A$, reached-outcome variable $Z$, and method $m$, measure

where both entropies $H_m$ and mutual information $I_m$ are in bits, $K_m$ is the dimensionless fraction of the registered feasible region covered, $Q^{\mathrm{tr}}_m$ is later transfer in the task's literal unit, and $\mathbf C_m$ is the complete cost vector. More action entropy is useful only when it improves outcome information, transfer, or later task value at an acceptable cost.

An adaptive curriculum therefore needs the initial skill and history stratum, the perturbed task dimension, perturbation dose, retry cost, feedback bits, criterion exposure, and delayed tests. This directly constrains Candidate 004 and the memory boundary in the consolidation chapter.

Route and act through resource-qualified state

The runtime policy should expose the state on which pacing and authority depend:

where $a_t$ is the commanded action or power target in its native unit, $o_{\le t}$ is causally received observation history, $\widehat r_t$ is the estimated resource/fatigue vector, $\widehat d_t$ the estimated damage state, $s_t$ remaining work in metres, seconds, events, or joules, $\widehat\pi_{\mathrm{opp},t}$ the opponent-policy estimate, $\widehat b_{\mathrm{team},t}$ the teammate-state estimate, and $f_t$ available feedback. Every estimate and channel receives a separate ablation.

This state conditions sensing, sparse routing, action authority, memory access, and recovery. It sharpens Candidate 002, Candidate 006, Candidate 007, and Candidate 012.

Compare a frontier, not a winner

For one task, retain

where $T_m$ is latency in seconds, $\epsilon_m$ task error in its declared physical or task unit, $p^{\mathrm{unsafe}}_m$ dimensionless unsafe-event probability, $E^{\mathrm{life}}_m$ lifecycle energy in joules, and $H^{\mathrm{human}}_m$ role-stratified effort in person-hours. A method is more efficient only through a preregistered utility or a Pareto improvement with non-inferiority on protected outcomes.

Make readiness an action envelope

At decision time $t$, admissible actions are

where $A_t$ is the currently feasible action set, $Z_{t:t+h}$ the multidomain outcome vector over horizon $h$ in hours, $\mathcal Z_{\mathrm{safe}}$ the registered safe envelope, $\mathcal I_t$ information available at time $t$, and $\alpha$ the dimensionless tolerated risk. An empty envelope triggers abstention or escalation.

Let $g_t\in\{0,1,2,3,4\}$ denote protected, modified, controlled, full-load, and adversarial operation. Promotion requires the next stage's envelope; a violation requires that $g_{t+1}<g_t$ remain possible. This turns staged return into a concrete test for Candidate 009 rather than a one-time health classification.

Distinguish coordination from shared input

For agent contributions $u_i(t)$ and $u_j(t)$, apply perturbation $do(\eta_i)$ to agent $i$ and estimate

where lag $\ell$ is in seconds, $X_t$ is task state, $u_i$ and $u_j$ retain their native contribution units, $\eta_i$ is a registered intervention in the unit of $i$'s action or state, and $\Gamma_{ij}$ has the product unit of the two contributions. A useful response must also reduce error in the protected task variable without increasing risk. Teammate turnover, role reassignment, message ablation, common-input controls, and never-co-trained cross-play separate reciprocal adjustment from synchrony. The nephron-coupling boundary adds the same requirement: coherence alone is not direct coupling, information transfer, or functional benefit (C-1495).

Preserve selection and evaluator lineage

Population evaluation carries the policy that granted training, observation, feedback, compute, role, and survival opportunity. Selected and rejected systems remain in the analysis, with censoring and missing outcomes explicit. Additional opportunity near selection thresholds is randomized when admissible or handled with a declared causal design. This extends Candidate 019 and the observation lineage in Candidate 014.

flowchart TB
    contract["Versioned episode contract<br/>task + observation + feasible action<br/>history + opponent/team + feedback<br/>resource + damage + selection"]
    manip["Sealed representative interventions<br/>cue window · action coupling · consequence<br/>fatigue · return stage · turnover · deception"]
    contract --> policy["Predict · act · query · abstain<br/>pace · coordinate · recover"]
    manip --> policy
    policy --> plant["Embodied task and adversarial environment<br/>real deadlines · realized actions · damage"]
    plant --> firewall["Outcome firewall<br/>anticipation ≠ interception ≠ cue use<br/>practice ≠ retention ≠ transfer<br/>exploration ≠ adaptability<br/>readiness ≠ return ≠ performance<br/>synchrony ≠ coordination ≠ shared information"]
    firewall --> compare{"Equal-budget comparison<br/>first trial + learning curve<br/>held-out opponent/team/task/state"}
    nulls["Mature null stack<br/>RL/POMDP · system ID · curriculum<br/>robust/adaptive control · VOI<br/>readiness/survival models · MARL/comms<br/>imitation/opponent + causal selection"] --> compare
    ledger["Complete ledger<br/>events · bytes · seconds · person-hours<br/>operational + embodied joules · harms"] --> compare
    compare --> keep["Retain literal residual<br/>only on protected outcomes"]
    compare --> retire["Retire mechanism claim<br/>preserve measurement contract"]
    keep --> history["Version history and future qualification"]
    retire --> history
    history --> contract

Editable source: representative-resource-qualified-performance.mmd.

Efficiency mechanism

The contract can improve efficiency through five measurable effects:

1. Less proxy optimization. Coupled perception/action tests prevent spending training and inference resources on a symbolic score that does not improve the physical or operational task.
2. Higher-value variation. History-qualified curricula direct perturbations toward task-relevant uncertainty instead of buying undirected entropy.
3. Resource-aware allocation. Routing, sensing, and authority respond to remaining work, fatigue, damage, opponent state, and recovery rather than treating every episode as fresh.
4. Reversible exposure. Staged return seeks a better availability--recurrence frontier than either permanent exclusion or an immediate full-load restart.
5. Complete comparison. Population selection, human effort, failure, recovery, and embodied resources are charged before a claimed saving is accepted.

Lifecycle energy for method $m$ is

where every term is in joules under one declared service interval. The terms are training, inference, sensing, actuation, communication, facility, recovery, maintenance, and amortized embodied energy. Human design, demonstration, coaching, labeling, tuning, monitoring, repair, safety, and medical effort are reported separately in person-hours. A lower device-energy reading does not establish a lifecycle saving.

Evidence status

The proposed composition remains a benchmark target until it beats the complete ordinary stack in F-006. A complete stack includes calibrated prediction and retrieval; model-free and model-based RL; POMDP/MPC; system identification; domain randomization and curriculum; robust/adaptive control; value of information and selective prediction; workload, readiness, survival, canary, and rollback models; multi-agent control and explicit communication; imitation and opponent models; and causal selection models.

Speculative extensions

Resource-conditioned sparse routing

Routing could condition on expected remaining work, resource uncertainty, damage, and safe fallback rather than only token or task features. It must beat a conventional estimator plus constrained control at equal sensing, compute, reserve, and failure allowance.

Retention-aware curriculum control

A curriculum controller could value a perturbation by delayed retention and held-out transfer rather than practice loss. It must beat fixed augmentation, domain randomization, active learning, Bayesian experimental design, novelty, quality-diversity search, and automatic curriculum at equal attempt, perturbation, feedback, evaluator, time, and energy budgets.

Deception-calibrated opponent memory

Opponent memory could retain policy versions, cue conflicts, confidence, change points, and abstention value. It must improve calibration and adaptation on new deceptive opponents beyond Bayesian opponent models, fictitious play, self-play, recurrent policies, retrieval, and conformal abstention.

Turnover-resilient communication

Teams could compress communication after shared practice while retaining typed repair, acknowledgements, and belief discrepancy when membership changes. Lower message volume counts only if cross-play, repair latency, task quality, and safety survive.

Counterfactual development policies

Population management could compare selection with broad-development or threshold-lottery policies and explicitly model opportunity-mediated outcomes. It must predict later capability in new cohorts and recover false negatives without hiding attrition or development cost.

Failure modes

Measurable predictions

F-006 contains the full protocols and hard retirement rules. The chapter-level commitments are:

Mechanism ablations are selective: removing actual-channel state should damage cue interventions; removing resource state should damage pacing and recovery; removing reversible stages should damage recurrence or availability; removing team-belief state should damage turnover and cross-play; removing selection lineage should damage prospective calibration. If every ablation merely reduces capacity and degrades every outcome, the proposed modules have not isolated their claimed roles.

Scope

Sound is not a ready-made sequence of objects. A receiver obtains pressure over time through a medium, room, body, aperture, transducer, clock, calibration state, and preprocessing chain. Sources overlap; direct paths mix with delayed copies; motion changes the operator; an emitted probe changes both the evidence and the physical scene. The system must therefore infer sources and task state from a versioned, action-dependent measurement.

This chapter turns the acoustics, hearing, and auditory-scene-analysis audit into an engineering contract. The detailed equations live in operator-qualified acoustic inference, and Fixture F-009 tests the contract across nine hostile tracks. The editable systems diagram is kept in operator-qualified-active-acoustic-inference.mmd.

The chapter connects four existing parts of the project:

1. sensorimotor grounding, because head, body, receiver, and emitter actions change what can be known;
2. operator-qualified sensing, because acoustic inference is conditional on a physical forward operator;
3. sparse predictive compute, because multirate and event-driven representations can reduce traffic only under a complete task-and-energy test; and
4. the energy model, because sensing, emission, motion, communication, maintenance, and embodied hardware remain inside the service boundary.

The intended output is a calibrated decision, retained acoustic artifact, or safe action—not one universal “hearing” score. Detection, discrimination, intelligibility, grouping, separation, dereverberation, localization, ranging, uncertainty, and lifecycle efficiency remain distinct outcomes.

Biological observation

The ear is a level-dependent, multirate measurement system

Acoustic level has meaning only with a reference pressure, frequency and time weighting, integration interval, sensor position, calibration, and uncertainty (C-1054). Room quantities likewise depend on a declared procedure, frequency band, source-receiver geometry, and spatial and temporal support (C-1055). These constraints are part of the biological stimulus, not bookkeeping added after a model runs.

The cochlea implements frequency-dependent mechanical filtering whose response near characteristic frequency is compressively nonlinear. Outer-hair-cell motility contributes causally to that amplification, while psychophysical filter estimates depend on centre frequency and fitting assumptions (C-1056–C-1058). Temporal coding also has regimes: envelope sensitivity is band-limited, auditory-nerve phase locking degrades with frequency, and downstream circuitry can sharpen synchrony in scoped low-frequency conditions (C-1059–C-1061).

“Cochlear,” “spiking,” and “neuromorphic” are therefore insufficient mechanism descriptions. A transferable front end must state its filter shapes, bandwidths, sample rates, level dependence, compression history, saturation, thresholds, refractory periods, timestamp precision, and calibration state. FFT, wavelet, mel, ERB, gammatone, gammachirp, modulation, and learned filters remain competing implementations; cochlear-like filtering is already conventional signal processing (C-1095).

Spatial hearing combines bounded cues with movement

Coincidence-sensitive circuits can encode interaural time differences, but the useful cue changes with frequency content, head and pinna transfer functions, room response, and source spectrum. Interaural level differences and spectral cues contribute on different bands and geometries (C-1062–C-1063). Head motion can disambiguate otherwise similar observations (C-1064); it is an information-acquisition action, not nuisance augmentation.

Early-arriving spatial information can dominate later copies, while spatial separation can improve recognition under masking (C-1065–C-1066). That does not make reverberation irrelevant. Stable spectrotemporal context and prior room exposure can alter later judgments; monaural context can compensate for some reverberant speech distortion; direct-to-reverberant energy contributes to distance judgments (C-1070–C-1073). The useful state therefore includes direct, early, and late components plus their time-varying relation to source and receiver motion.

Auditory objects are inferred under masking

Energetic masking concerns overlap at the receiver. Informational masking also depends on whether target and masker are perceptually segregated, and selective listening changes strongly with ear and source configuration (C-1067–C-1068). Temporal coherence across frequency channels is a plausible grouping mechanism, not a settled universal decomposition rule (C-1069).

This makes four outcomes non-interchangeable:

• detection: whether task evidence is present;
• grouping: which events belong together and how uncertain that assignment is;
• separation: recovery of one or more waveforms under an explicit reference and permutation contract; and
• causal identity: which physical source produced a recovered component.

A separator can improve a waveform metric while switching identities, damaging localization, or changing perceptually important structure. Published separation metrics can hide severe waveform changes (C-1094), so no one metric is allowed to stand in for the scene.

Sparse activity and efferent control have narrow causal interpretations

Awake auditory cortex can represent sounds with low population activity, and sparse-coding objectives can learn cochlea-like kernels in some sound ensembles. The learned code depends on the training ensemble (C-1074–C-1076). Sparse activity is evidence about representation, not a direct measurement of system joules. Sensors, clocks, memory traffic, event routing, decoding, idle power, and training may dominate (C-1097).

Efferent activation can causally reduce cochlear gain and improve masked-tone auditory-nerve responses in the studied preparations (C-1077). A general human speech-in-noise benefit from medial olivocochlear strength remains disputed (C-1078). An AI analogue earns the name only when an intervention on task-conditioned front-end gain changes a registered external endpoint after distortion, recovery time, calibration, and energy are charged. An attention map or smaller hidden activation is not that evidence.

Echolocation is active observation with physical consequences

Echolocators vary call timing with spatial task difficulty, adapt emission level with distance, steer or reshape the acoustic field of view, and coordinate head or pinna motion with emission (C-1079–C-1082). Human experts can use self-generated echoes, but the demonstrated capability includes extensive experience and neural reorganization (C-1083). Spectral jamming avoidance in bats is context-dependent rather than universal (C-1084).

The transferable loop is emit, receive, infer, move or re-aim, and decide. Its benefit is conditional on waveform ambiguity, target reflectivity, range, clutter, association, other emitters, detectability, acoustic exposure, and the energy required to produce and receive the probe (C-1085, C-1096).

Proposed AI translation

Preserve the episode, operator, and action identity

For episode $e$, retain the contract

where:

Calibration identity can become part of the learned task when a model depends on device, room, or preprocessing state (C-1098). Train-test splits must group clean sources, convolved variants, neighboring windows, impulse-response and renderer lineages, HRTF or array families, devices, speakers or listeners, and derived labels. Capture time and receipt time remain distinct, and only causally received observations may affect a decision.

Keep pressure, level, exposure, and energy separate

For pressure $p(t)$ in pascals over interval $T$ in seconds,

where $p_{\mathrm{rms}}$ is RMS pressure [Pa], reference pressure $p_0=20\,\mu\mathrm{Pa}$ in air, and $L_p$ is sound-pressure level [dB re $20\,\mu\mathrm{Pa}$]. Frequency weighting, time weighting, band, position, orientation, and calibration must be explicit. Sound exposure $\int p^2(t)dt$ has units $\mathrm{Pa^2\,s}$; it is not acoustic emission energy [J], electrical energy [J], or a complete biological risk model.

For source pressure $x_s(t)$, received pressure $y_m(t)$, and noise $n_m(t)$, all in pascals, use the time-varying mixture

where $m$ is receiver-channel index [1], $s$ is source index [1], source count $N_s$ is dimensionless, $t$ and delay $\tau$ are in seconds, and propagation kernel $h_{m,s}(t,\tau)$ is in reciprocal seconds. A fixed convolution is only a special case; moving sources, receivers, boundaries, media, and emitters make the operator time-varying.

Build a qualified multirate front end

The front end may combine:

1. an FFT, wavelet, constant-Q, mel, ERB, gammatone, gammachirp, modulation, or learned analysis bank;
2. level- and history-dependent compression with declared attack, release, saturation, and distortion;
3. masking and source-count hypotheses rather than a single global noise scalar;
4. dense samples or sparse onset, offset, change, phase, or threshold events with timestamp and refractory identity; and
5. a decoder or downstream task that is charged for reconstructing information discarded by the encoder.

Band selection and update rate should follow the causal time scales needed by the task. Fine timing cannot be assumed above the supported band; weak sustained signals and rare hazards cannot be discarded merely because an event count falls.

Fuse spatial evidence without erasing ambiguity

For path-length difference $\Delta r$ [m] and sound speed $c_{\mathrm{snd}}$ [m/s], interchannel delay is

where $\Delta t$ is in seconds. For left and right RMS pressures $p_L,p_R$ [Pa],

is interaural level difference [dB]. Both cues are band-, source-, room-, body-, and calibration-qualified. Spectral cues and head or receiver motion remain separate intervention axes. A correlation surface under reverberation may be multimodal; reducing it to one delay without calibrated uncertainty loses the ambiguity the controller needs.

The learned path must be compared with generalized cross-correlation, matched filtering, steered-response power, delay-and-sum, robust and minimum-variance beamforming, MUSIC or ESPRIT, Bayesian delay estimation, and source tracking. GCC, MVDR, and subspace localization are mature nulls, not discoveries created by a biological label (C-1086–C-1088).

Separate scene inference from metric substitution

Each output head declares one literal construct and uncertainty model. The minimum firewall is:

Independent-component methods, time-frequency sparsity methods, deep clustering, permutation-invariant training, and time-domain separation are mature but assumption-bound baselines (C-1089–C-1093). The strongest compatible composition—not an isolated weak separator—is the decisive null.

Close the action loop

At decision time $t$, policy $\pi_q$ for method $q$ selects

where $\mathcal H_t$ is causally received history, $\widehat X_t$ is estimated scene state, $\widehat O_t$ is operator and calibration state, $\widehat U_t$ is uncertainty, $\mathcal A_t^{\mathrm{safe}}$ is the feasible action set, and $B_t$ is remaining componentwise budget. Emission actions set waveform, spectrum [Hz], level [referenced dB], aim [degree or rad], and time [s]; body actions set pose and motion [m, rad, s]; sensor actions select aperture [m], channels, and sample rate [sample/s]; gain is dimensionless or declared in dB; the task action uses its native unit.

For monostatic echo emission and receipt times $t_{\mathrm{emit}}$ and $t_{\mathrm{recv}}$ [s], nominal range is

where $\widehat r$ is range [m]. The interval must propagate clock error, target motion, refraction, multipath, ringing, waveform ambiguity, association, and detector threshold. Multiple emitters add collision, authentication, fairness, adversarial imitation, and jamming state. The proposal must beat TDMA, FDMA, CDMA, random access, listen-before-talk, power control, coded probes, cancellation, authentication, and game-theoretic allocation at equal spectrum, power, messages, and exposure.

One system, tested against a complete null

flowchart TB
    scene["Sources · interferers · targets<br/>geometry · medium · room · motion"] --> field["Time-varying acoustic field<br/>direct path · reflections · masking · clutter"]
    emit["Active emission<br/>waveform · level · spectrum · aim · timing"] --> field
    body["Receiver/body/array action<br/>pose · head motion · aperture · gain"] --> operator["Versioned observation operator<br/>transducer · HRTF/array · clock · calibration · support"]
    field --> operator
    operator --> received["Causally received waveform/events<br/>pressure · timestamps · missingness · uncertainty"]
    received --> front["Front end<br/>filterbank · compression · sparse events"]
    front --> infer["Infer and calibrate<br/>detect · group · separate · localize · range"]
    infer --> decide["Act · emit · move · query · abstain"]
    decide --> emit
    decide --> body
    infer --> firewall["Outcome firewall<br/>detection · intelligibility · grouping · separation<br/>direction · range · calibration · risk"]
    nulls["Mature null stack<br/>ERB/FFT/AGC · GCC/MVDR/MUSIC<br/>ICA/DUET/WPE/PIT/TasNet<br/>Bayes/POMDP/MPC/VOI · spectrum protocols"] --> compare{"Equal operator/action/lifecycle budget"}
    firewall --> compare
    ledger["Complete ledger<br/>samples · bytes · seconds · person-hours<br/>emission/exposure · operational + embodied joules"] --> compare
    compare --> retain["Retain literal track residual"]
    compare --> retire["Retire composition<br/>preserve acoustic contract"]

Editable source: operator-qualified-active-acoustic-inference.mmd.

The contract is owned by existing components:

• Candidate 002 — Multiscale context broadcast for multirate context;
• Candidate 006 — Reversible physical skill for reusable analog or beamforming paths;
• Candidate 007 — Endogenous observation surveillance for action-dependent observation;
• Candidate 009 — Graded assurance envelopes for calibration, uncertainty, exposure, and fallback;
• Candidate 012 — Latency-qualified authority for travel time, evidence age, and decision authority; and
• Candidate 014 — Versioned observation contract for operator and calibration identity.

Efficiency mechanism

The proposed composition can improve efficiency through four literal pathways:

1. Multirate analysis. Slow envelopes and long context need not update at the rate required for fine timing. Savings must appear in samples, memory traffic, latency, and measured joules without degrading a protected band.
2. Conditional events. Onsets, offsets, changes, or uncertainty triggers can suppress redundant transport and downstream work. The decoder, timestamping, missed sustained signals, false events, and idle event hardware remain in the comparison.
3. Selective physical action. A head turn, receiver move, gain change, or acoustic probe can resolve an ambiguity more cheaply than processing another passive interval. Its motion, emission, exposure, delay, detectability, and interference costs must be lower than the decision value it creates.
4. Reusable calibrated structure. Stable filter, beamforming, correlation, or control paths can move to DSP, FPGA, ASIC, analog, or neuromorphic hardware when reuse amortizes conversion, characterization, drift monitoring, repair, and replacement.

For one registered service interval, lifecycle energy is

where method $q$ has data-acquisition, training, acoustic-emission, sensing, inference, motion, communication, storage, calibration, maintenance, and amortized embodied-energy terms, each measured in joules over the same service interval. The complete ledger also reports:

• pressure samples, sparse events, optimization steps, queries, and bytes;
• wall time [s], task latency [s], peak memory [byte], and network traffic [byte];
• acoustic exposure [$\mathrm{Pa^2\,s}$], unsafe-event count [1], and harm or detectability in a registered task-specific unit;
• emitted acoustic energy [J], electrical input [J], sensing [J], motion [J], compute [J], cooling [J], and facility allocation [J];
• recording, labeling, listening tests, calibration, tuning, monitoring, maintenance, and repair [person-hour]; and
• transducers, arrays, processors, actuators, replacements, wear, embodied hardware, and opportunity cost.

An arm that exceeds any binding budget is infeasible. Resources released by an ablation remain unused. Lower event count, lower FLOPs, quieter emission, or a faster accelerator is not an efficiency result unless accepted task service improves on the complete protected vector.

Evidence status

The stable claim ledger mirrors all 46 audit-local claims without promoting the audit itself into evidence:

No row establishes that the complete proposed system beats the conventional stack. The evidence supports the physical constraints, biological phenomena, and engineering nulls. The composition remains an experiment.

Speculative extensions

Operator-conditioned auditory state

Learn a compact state that jointly represents room response, body or array transfer, clock uncertainty, source hypotheses, and front-end state. It must predict held-out calibration probes and remain identifiable under scene changes. Compare it with explicit system identification, adaptive filtering, room banks, and Bayesian state-space models.

Counterfactual acoustic action

Train the policy to predict not only what a probe or head turn might reveal but also its exposure, detectability, collision, motion, and lifecycle cost. Use paired simulator seeds and physical replay so action value is estimated against the same latent scene, not easier episodes selected after acting.

Cross-timescale front-end control

Let fast protection and compression, intermediate event routing, and slower room or task adaptation share state without sharing unrestricted authority. Test whether the joint controller beats independently tuned AGC, PCEN, adaptive filters, selective prediction, and recurrent context after coordination cost.

Population-qualified acoustic hardware

Route recurring transforms among digital, analog, array, and event-driven front ends by measured device identity, calibration envelope, workload reuse, and remaining service life. Hold out hardware classes, lots, transducers, and array geometries. A nominal-device gain does not establish fleet-level benefit.

Strategic multi-emitter sensing

Model emitters as cooperative, selfish, deceptive, imitating, colluding, or turnover-prone agents. Joint waveform and scheduling policies must retain source authentication and calibrated uncertainty under never-coordinated population cross-play. Ordinary spectrum protocols remain the null.

Failure modes