Structural growth, specialization, and conditional routing

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.