Research roadmap

A publication checkpoint records progress. Completion is earned by closing a gate, retiring a hypothesis, or producing a reproducible result.

Scope

This roadmap converts an open-world research ambition into bounded iterations. Breadth remains unlimited: any scientific field may contain a transferable operation. Work remains finite because each iteration must change a canonical claim, principle, chapter, equation, diagram, decision, or experiment—or record that the new evidence changes none of them.

The stages are dependency gates rather than dates. Evidence discovery, architecture writing, null-model comparison, and measurement design continue in parallel, but an integrated system does not inherit a mechanism merely because its audit is interesting.

Live operational view

This document owns the stage order, gate meaning and exit conditions. The canonical milestone manifest projects Stages 0–5 into GitHub milestones, where issues and pull requests can be assigned, discussed and closed. The Go metadata command recreates that projection from the manifest; GitHub is not a second roadmap source.

The views answer different questions:

A milestone percentage therefore measures operational progress only. It cannot promote a claim, turn a smoke run into evidence or waive a later gate. Due dates are omitted until a real external commitment justifies one.

Biological observation

Development, adaptation, and maintenance operate on different timescales. Variation precedes selection; temporary states can precede durable change; stable organization remains repairable; and recurring behavior can migrate into structure. Relevant observations are separated across P-003, P-004, P-009, and P-010.

The engineering consequence is progressive commitment. Cheap reversible work comes first; expensive structural or hardware commitments require stronger evidence and longer reuse horizons.

Proposed AI translation

One research iteration

flowchart LR
    discover["Discover across fields"] --> scope["Scope evidence + boundary"]
    scope --> dedupe["Deduplicate into P-principle"]
    dedupe --> null["Name strongest null model"]
    null --> contract["Write decisive contract"]
    contract --> result{"Result"}
    result -->|"supports"| promote["Promote scoped mechanism"]
    result -->|"ties"| merge["Merge into conventional primitive"]
    result -->|"fails"| retire["Retire translation; keep evidence"]
    promote --> update["Update claim · chapter · math · diagram"]
    merge --> update
    retire --> update
    update --> discover

Editable source: ../assets/diagrams/research-iteration-loop.mmd.

An iteration has three independently reviewable outputs:

Parallel workstreams

Four workstreams move continuously, with different exit conditions:

1. Discovery and evidence. Sample under-covered fields by constrained problem, audit primary sources, extract causal operations, and preserve boundary conditions.
2. Principle synthesis. Deduplicate observations using the mechanism tuple in chapter 07, then update the shared primitive rather than accumulating themed components.
3. Concept closure. Turn retained principles into readable data contracts, control paths, state ownership, equations, null models, failure signatures, graphics, and measurable predictions.
4. Decisive testing. Compare one operation at a time against the strongest conventional method at equal interface, budget, hardware, and quality–risk envelope.

Stage 0 — Evidence, synthesis, and contracts

Stage 0 is complete for a mechanism only when:

• its observation has at least one scoped claim and primary source;
• its nearest existing AI and engineering analogues are recorded;
• its normalized operation maps to a P- principle or a held candidate;
• the affected chapter exposes state, information flow, timescale, physical boundary, and failure behavior;
• every equation defines symbols and units;
• a conventional null, ablation, measurement envelope, and rejection condition exist; and
• link-only discussion provenance records remain source leads rather than evidence.

The repository structure satisfies this workflow. Individual mechanisms remain at different positions inside it; the domain inventory and adoption matrix expose those differences.

Historical evidence has a separate promotion gate

Longitudinal evidence enters the roadmap as one of four typed claims:

1. Accounting identity: the declared stocks and flows reconcile.
2. Descriptive regularity: a relation exists in a sampled period, place, coding system, and surviving archive.
3. Causal estimate: an intervention or design-specific contrast identifies an effect under explicit assumptions.
4. Prospective prediction: a frozen model succeeds on a period, place, lineage, or regime unavailable during development.

No row upgrades itself into the next. A cohort matrix can be arithmetically correct and forecast poorly under changing rates. An adoption curve can fit while influence, common cause, homophily, and command remain unidentified. A historical instrument can estimate a local effect without predicting a future system. A surviving record can support chronology while omitting destroyed or never-produced evidence.

For interval $\Delta t$ seconds, the basic population identity is

where $N$ is qualified units at a timestamp and $B,D,I,O$ are entry, retirement, immigration, and outmigration counts during the interval. The identity finds bookkeeping errors; it does not explain or forecast a flow. The full cohort, selection, diffusion, collapse-vector, and prospective-test contract is derived in population-observation math and folded into Candidate 014 rather than promoted as a new principle.

Stage 1 — Isolated mechanism experiments

The first tests isolate control operations before composing them.

Stage-1 exit gate: at least one isolated operation changes the measured quality–risk–latency–energy frontier relative to its strongest null. A tie merges the biological translation into the conventional primitive; a loss retires the translation while preserving the source evidence.

Stage 2 — Runtime plus adaptation

Combine the event controller from chapter 30 with attributable episodic capture from chapter 40. Keep slow structural state protected while a branch adapts.

Required composed ablations:

• runtime only;
• adaptation only;
• runtime plus adaptation without maintenance pricing;
• full two-loop system with storage, replay, rollback, and controller cost; and
• dense or monolithic continual-learning baselines at the same capacity and observation stream.

Stage-2 exit gate: sequential adaptation improves a declared task stream without a retention, calibration, rare-event, rollback, or lifecycle-energy regression hidden by the average score.

Stage 3 — Grounded developmental curriculum

Introduce temporally aligned perception, action, outcome, and language. Compare passive observation, action-conditioned prediction, targeted acquisition, and text-centric learning at matched capacity and interaction budget.

The environment must expose interventions, counterfactual structure, missing modalities, sensor noise, delayed outcomes, and distribution shifts. Language may describe, query, or compress the learned state; it cannot be the only path by which every evaluation variable enters the model.

Stage-3 exit gate: held-out intervention, temporal prediction, transfer, and composition gains survive leakage, memorization, simulator-state, and language-shortcut controls.

Stage 4 — Structural maturation

Apply the reversible maturity cycle from chapter 50 and the separated hardening targets from chapter 60. Candidate operations include structured pruning, module fusion, compilation, placement, quantization, retrieval externalization, write protection, and local reopening.

Every operation receives its own invalidation and recovery test. Parameter count is diagnostic; resident bytes, transferred bytes, wall energy, tail latency, maintenance, and rollback determine the system result.

Stage-4 exit gate: benefits remain positive after discovery, validation, migration, shadow traffic, failed promotions, recovery, and the observed reuse horizon are included.

Stage 5 — Substrate co-design

Compare accepted mechanisms on optimized conventional hardware and on event-driven, near-memory, neuromorphic, or other appropriate substrates. C-015 establishes feasibility for a scoped class of event-driven systems; it does not select a winner for this workload.

Algorithm, precision, memory layout, interconnect, batching, cooling boundary, and hardware control policy become joint experimental variables. The same qualified event and quality–risk envelope apply to every substrate.

Stage-5 exit gate: an independently reproducible end-to-end lifecycle advantage transfers beyond the workload or hardware used to tune it.

Breadth-wave decisions and next queue

The latest completed audits already changed the architecture rather than merely adding citations:

The music, library/archival, comparative-cognition, visual-design, mathematical-practice, fluid-dynamics, sports, optics, semiconductor, acoustics, information-thermodynamics, and olfaction/chemical-sensing queues are closed by explicit no-promotion decisions. The music Fixture F-001 retains E-MUSIC-07 as a cross-candidate benchmark; the archival refinement stays inside Candidate 017's query-registered preservation track; and comparative cognition contributes Fixture F-003 as an opportunity/history audit across Candidates 004/006/007/014/018/019; and visual design contributes Fixture F-002 as a versioned reconstruct–externalize–inspect–transform–evaluate audit across Candidates 004/014/017/018/019. It treats fixation and negative transfer as measured failure modes, qualifies evaluators and material evidence, and keeps lineage separate from truth, intent, and recoverability; and mathematical practice contributes Fixture F-004 across Candidates 004/009/010/011/014/017/019. It keeps proposal provenance separate from correctness, requires counterexamples and proof-DAG reconstruction, preserves typed lifecycle state, prevents dependency leakage, propagates reverse-dependency invalidation, and includes proof, checker, human, rebuild, and joule cost; and fluid dynamics contributes Fixture F-005 across Candidates 002/003/006/007/012/014. It requires signed scale transfer, tail fidelity, detector-qualified structures, explicit closure/model support, objective-qualified reduction, total adaptive work, actual observation and sensor operators, observability and posterior calibration, stable net-benefit control, vector mixing outcomes, transition history, extreme calibration, propagated uncertainty, and a full lifecycle energy boundary. Its null is the complete relevant CFD, reduced-model, estimation, sensing, control, rare-event, metrology, and accounting stack. Sports contributes Fixture F-006 across Candidates 002/004/006/007/009/012/014/019. It preserves actual observation/action coupling; separates anticipation from interception, practice from retention and transfer, relevant exploration from variance, and fatigue/readiness/damage from staged return; challenges coordination with turnover and perturbation, deceptive cues with calibration, selection with counterfactual opportunity, and apparent efficiency with full athlete, coach, facility, sensor, recovery, medical, failed-pathway, wall-time, and joule budgets. Its null is the complete calibrated prediction, retrieval, planning, control, curriculum, assurance, coordination, causal-inference, and lifecycle- accounting stack. Optics contributes Fixture F-007 across Candidates 001/006/007/010/014/017/018. It binds inferred state to the physical operator, acquisition action, photons/noise, calibration, ambiguity, prior support, saturation, typed uncertainty, route eligibility, retained-query obligation, and complete resource boundary. Its adversarial tracks test null- space honesty, prior mismatch, active photon allocation, multiplex crossover, drift attribution and fallback, optical/digital service curves, multi-device fabrication and thermal variation, and physical compaction under future queries. Its null is the complete inverse-method, experiment-design, control, fusion, digital/analog acceleration, calibration, and lifecycle-accounting stack. The first T1 slice is now executable as the F-007 null-space-honesty smoke harness: it must expose false specificity, accept justified abstention, and exactly tie the complete mature active null rather than manufacture a novelty advantage. Semiconductor reliability contributes Fixture F-008 across Candidates 001/005/006/009/010/012/014/017/018. It binds accepted service to manufactured-unit hierarchy, mission profile, mechanism-specific acceleration, censoring, monitor age, electrothermal concentration, fault geometry, correction and recovery state, protected exact boundaries, physical- weight age, endurance, and fabrication-through-retirement resources. Its ten tracks test hierarchical transfer, accelerated-life support, sparse wear, correlated faults, voltage authority, approximation containment, analog/digital crossover, nonideality transfer, wear placement, and keep/repair/repurpose/ replace policy against the complete mature reliability stack. Acoustics contributes Fixture F-009 across Candidates 002/006/007/009/012/014. It binds each result to source and emission action, propagation/room state, receiver geometry and motion, calibrated operator support, context/history, literal target, uncertainty, independent unit, feasible action, exposure, and complete resource boundary. Its nine tracks test level-dependent front ends, sparse event timing, active emission/reception, localization under operator shift, reverberation transfer, grouping/separation, efferent-like gain, multi-emitter interference, and full lifecycle cost against the complete calibrated acoustics, inference, control, spectrum, uncertainty, and accounting stack. Information thermodynamics contributes Fixture F-010 across Candidates 001/005/006/009/010/012/014/017/018. It binds every physical- efficiency result to one accepted service while keeping logical, device, circuit, workload, facility, and lifecycle quantities distinct. Its tracks challenge lower-bound substitution, finite-time and finite-error operation, reversible-history closure, thermal reliability, measurement/controller reset, uncertainty-relation applicability, distribution mismatch, topology, data movement, PUE allocation, and fabrication-to-retirement ranking. Its null is the complete compression, reversible-computation, stochastic-thermodynamics, circuit/architecture, workload-metering, facility, and lifecycle-assessment stack. Olfaction and chemical sensing contribute Fixture F-011 across Candidates 002/006/007/009/010/012/014/017/018. It binds source and release, turbulent or reactive transport, realized sniff/pump/motion, receiver path, dynamic response and recovery, cross-sensitivity, calibration and exposure history, operator support, literal outcome, feasible action, safety, independent unit, and complete lifecycle cost. Its fourteen tracks test concentration–identity factorization, sensor coverage, normalization, deconvolution, active acquisition, bilateral/serial/wind cues, plume statistics, search policy, embodiment, sparse total cost, drifting representations, mixture masking, sensor poisoning, analytical confirmation, and safe tiered action against the complete calibrated sensing, analytical, estimation, control, standards, human-work, consumable, and lifecycle stack. The next breadth field is selected by expected ability to split, merge, reject, or re-baseline a current mechanism. Breadth alone does not close a row.

Efficiency mechanism

Stage gates reduce the cost of compounding weak mechanisms. Shared principle bundles reuse telemetry, baselines, and test regimes; isolated rejection avoids paying for full-system integration; exact provenance prevents repeated searches; and negative results retire duplicated work.

For stage $k$, let $E_{k,\mathrm{research}}$ and $E_{k,\mathrm{integration}}$ be joules spent within declared research and compute boundaries. Let $p_{k,\mathrm{survive}}$ be the empirically estimated fraction of candidates that pass the isolated gate. The expected integration energy avoided by testing first is

This quantity is reported only when both energy boundaries are actually measured; its purpose is to expose the cost of evaluation, not to assume that research automatically saves energy.

Evidence status

• The ordering and gates are project decisions.
• Mechanism-specific evidence remains in the claims ledger; roadmap placement does not promote a claim.
• The candidate and fixture catalogues contain written contracts and bounded development harnesses, but no claim-eligible result.
• The six-stage composition and its lifecycle benefit remain speculative until the gates produce measured results.
• Sensorimotor transfer remains speculative under C-007.

Speculative extensions

• Maintain a machine-readable experiment ledger linked to commits, hardware records, raw telemetry, derived figures, and claim-status changes.
• Use active literature search to select the next field by expected impact on an unresolved principle, not by publication volume.
• Allocate compute to experiments by expected decision value and uncertainty reduction once the proxy is calibrated against actual outcomes.
• Run adversarial review agents whose only task is to find conventional explanations, hidden costs, or domain evidence that breaks a principle.
• Publish negative results and retired translations as reusable research outputs rather than silently deleting them.

Failure modes

• Perpetual breadth: new fields accumulate without changing principles or experiments.
• Premature composition: several plausible mechanisms are integrated before any one survives its null model.
• Gate drift: thresholds or comparison boundaries change after results are visible.
• Mechanism creep: a component enters the system without a claim, principle, ablation, or owner.
• Proxy capture: FLOPs, parameter count, or average accuracy substitutes for measured physical cost and qualified behavior.
• Negative-result loss: a failed translation disappears and later returns under another domain name.
• Benchmark enclosure: the roadmap optimizes a convenient task family that no longer tests the intended operation.
• Documentation lag: equations, diagrams, or claim status stop matching the executed artifact.
• Hardware lock-in: a mechanism is tuned to one substrate before its functional interface is stable.
• Narrative immunity: biological appeal preserves a mechanism after its measurable predictions fail.

Measurable predictions

1. Each breadth wave changes at least one principle boundary, null model, experiment regime, or explicit no-change decision; otherwise the sampling policy is revised.
2. The ratio of scoped claims to deduplicated principles increases as coverage grows, while duplicate organism-themed components decline.
3. At least one Stage-1 candidate ties a conventional baseline and is merged, and at least one loses and is retired; a roadmap in which every mechanism wins is not discriminating.
4. Shared telemetry and null-model infrastructure reduce the incremental cost of testing later principles relative to the first isolated contracts.
5. Mechanisms that pass isolated gates still sometimes fail composition, producing explicit interaction terms rather than post-hoc stories.
6. Lifecycle accounting reverses at least one apparent saving based on FLOPs, parameter count, or inference-only energy.
7. Every completed stage leaves reproducible baselines, scoped claim changes, negative results, and a smaller set of unresolved integration decisions.