Sparse prediction and adaptive compute

Scope

This chapter specifies the online control path that decides, for each event, whether to reuse an existing prediction, exit, invoke another module, read memory, acquire another observation, or escalate. Its objective is not minimum FLOPs. It is minimum measured resource use subject to explicit quality, calibration, deadline, and safety constraints.

The runtime has three coupled but distinct jobs:

1. estimate what changed and what remains uncertain;
2. choose the next computation or observation by expected decision value; and
3. allocate physical capacity without making the fast task path unstable.

Training curricula, long-term consolidation, structural pruning, and skill compilation are handled in other chapters. Their products may be invoked here, but this runtime does not silently rewrite slow model state.

Biological observation

The relevant evidence separates demand detection, distributed context, local response, and physical supply.

Activity is selective and regulated

Neural signaling operates under a strict metabolic budget, which constrains the fraction of cortex that can be strongly active at once (C-001). In the Drosophila mushroom body, feedback inhibition keeps odor codes sparse and decorrelated; disrupting that loop impairs discrimination between similar learned odors (C-025). These results support selective activity and active inhibition. They do not identify top-$k$ routing, a pruning ratio, or a digital energy saving.

Prediction, uncertainty, and action are different signals

Hierarchical prediction with feed-forward residuals accounts for selected visual-cortical response properties (C-005). Active whisker sensing also shows why a passive residual is incomplete: self-generated sensor mechanics affect the observed neural code (C-022). A residual says that a prediction was wrong. It does not by itself say whether the error is irreducible noise, model ignorance, task-relevant novelty, or a reason to buy another observation.

Broad context is interpreted locally and temporally

Context-dependent disinhibition can change gain in a selected cortical population (C-020). More direct causal work on neuromodulation shows three constraints on a possible artificial abstraction:

• antagonistic signals can sustain different distributed behavioral modes (C-046);
• the frequency and duration of one broad signal can produce different and non-monotonic system responses (C-047); and
• receptor identity and location can turn the same source into transient or sustained local responses (C-048).

The engineering candidate is therefore a rate-limited context broadcast whose receivers own their gains, thresholds, and temporal filters. It is not a central command containing a route for every module. Its evidence, limits, and conventional analogues are developed in the neurodevelopment and global-control audit.

Demand and physical supply can be locally coupled

Presynaptic ATP production can be recruited by local activity (C-049); mitochondrial docking changes resource placement and local synaptic dynamics (C-050); and neural activity can recruit an adjacent vascular response in the studied rodent preparations (C-051). The transferable constraint is modest: measure demand near work, account for placement, and let an infrastructure layer adjust supply. The neurovascular paper's model-derived percentage is not an architecture constant, and none of these cellular results provides a conversion from ATP to accelerator joules.

Aggregate headroom is likewise not evidence of delivery or useful uptake at the support where demand occurs; local deficit, arrival, consumption, and export remain distinct records (C-1488).

Taken together, the observations motivate a layered controller: fast local prediction and gating, a narrow context path, and a slower resource plane. The architecture still has to beat standard estimation, decision, routing, and control methods.

Proposed AI translation

Separate the quantities before making a gate

Let $y_t$ be an observation in its declared sensor units and $\hat y_{t|t-1}$ the prediction available before observing it. Define the raw residual

If the predictor supplies a calibrated residual covariance $S_t$, the normalized innovation is

$r_t$ has the units of $y_t$; $\nu_t$ is dimensionless. A large $r_t$ with a large expected $S_t$ may be unsurprising. Conversely, a small residual in a safety-critical variable may justify action. The runtime therefore carries separate fields for:

No implementation may collapse all of these into “confidence” and retain the same claim. Residual magnitude, likelihood, expected value, and risk answer different questions. Predictive regulation adds the same separation between forecast value, feedback correction, integrated action exposure, reserve debit, and later outcome (C-1494).

Relative sensing is a symmetry, not a normalization label

The same separation applies when a gate should respond to change relative to a background. For a positive input $u(t)$ and reference $r(t)$, a log ratio

is invariant to the common transformation $(u,r)\mapsto(pu,pr)$. That algebra does not establish how $r$ is acquired, how long it remains valid, or whether the downstream target is itself scale-invariant. Exact fold-change detection is stronger: the complete output trajectory must remain the same after the input history and its adapted initial state are scaled together (C-1540). Exact adaptation or equal peaks cannot substitute for that test.

Biological evidence shows why the interface must be recorded. Bacterial chemotaxis connects finite-regime intracellular rescaling to active migration (C-1542); nuclear RelA fold change predicts three downstream transcripts in one cell system (C-1544); yet dynamic extracellular cytokine protocols support a rectified dose-change statistic at the upstream cytokine-to-NF-κB interface (C-1545). The pathway name therefore cannot select the statistic.

The runtime carries a typed candidate-statistic record

where $\tau_t^{(\mathrm{ref})}$ is reference age in seconds and $\chi_t^{(\mathrm{sup})}$ is a typed support/validity state. A selector may route a ratio-qualified task through the relative channel, but the record's fields retain their individual units and types; it is not an ordinary numeric vector. An absolute load, dose, damage, or safety target retains a calibrated absolute path. Near zero, after sign changes, under clipping, or with stale reference state, the log-ratio path must abstain or fall back.

This design must beat explicit log-ratio, difference, future-free streaming normalization, state-space and compact recurrent nulls while charging reference updates, selector execution, fallback and state writes. The exact model, interface firewall, scale-orbit plot, and trajectory discrepancy are derived in Interface-qualified scale symmetry; the reciprocal ten-track test is Fixture F-026.

Command, interface work, transport, and observation are different stages

An adaptive gate often mistakes its command for the work that actually crossed an interface. Electrochemistry supplies a sharp counterexample. In one common charge-transfer form,

where $j$ and $j_0$ are current densities in amperes per square metre, $\eta$ is overpotential in volts, $\alpha_a$ and $\alpha_c$ are dimensionless transfer coefficients, $F$ is the Faraday constant in coulombs per mole, $R_g$ is the molar gas constant in joules per mole-kelvin, and $T$ is absolute temperature in kelvin. Even when this local closure is correct, bulk transport, double-layer state, and series loss still separate commanded voltage from realized flux and terminal observation (C-1530).

The corresponding artificial contract records a five-stage chain:

1. the controller issues a command under declared authority;
2. receiver-local state converts the command into a driving force;
3. an interface transfers bounded useful work;
4. internal transport accepts, delays, or saturates that transfer; and
5. an observation operator reports a noisy, delayed projection of the result.

A candidate may compress this chain only after an intervention shows that the omitted stage is conditionally redundant. Otherwise, a low command error can coexist with wrong realized service. A validity check may reject an observation before interpretation, but passing such a check does not identify a unique mechanism (C-1532). Likewise, a low prediction residual cannot give physical semantics to unidentifiable latent parameters (C-1537).

Observation is affected by action

The acquisition model must also represent how sensing and response change the future observation channel. Operational telemetry can be delayed, incomplete, pooled, or policy-coupled: robust residual methods remain strong nulls (C-124); current counts can omit events that occurred but have not yet arrived (C-126); pooled signals can trade attribution for coverage (C-127); and a fast behavioral proxy can drift with platform, attention, or policy (C-128).

More importantly, an action can alter both hidden state and telemetry. Isolation may reduce propagation while also removing the isolated component's messages; throttling can suppress an error count by suppressing all traffic; a warning can change user behavior; and a new sensor policy changes which events can be observed. The controller therefore versions coverage, delay, ascertainment, observation-model state, and every sensing or response action. A drop in the post-action residual is not independent recovery evidence (C-131).

Candidate 007 tests joint process/observation estimation against residual CUSUM/GLR, nowcasting, maximum coverage, value-of-information sampling, and delay-aware POMDP or model-predictive control. It must use true data vintages, preserve subgroup coverage, estimate counterfactual observation paths, and charge investigation and action capacity as well as compute.

Price a menu of acquisitions

At event $t$, the system considers a finite action menu $\mathcal A_t$:

exit | reuse cached state | run local refinement | invoke expert
     | read episodic/factual memory | acquire sensor action | call tool
     | escalate to a larger model or human

For acquisition $a$, estimate the expected improvement in decision utility $\Delta U_t(a)$ and its full energy, latency, and traffic costs. A Lagrangian form of the selection rule is

subject to hard task constraints such as

$\mathcal I_t$ is the information legally available when event $t$ is decided, $\delta\in[0,1]$ is the preregistered maximum conditional probability of an unsafe outcome, and $L_{\max}$ is the latency ceiling in seconds. $\lambda_E$, $\lambda_L$, and $\lambda_B$ have units utility/joule, utility/second, and utility/byte respectively, converting the three costs into the declared utility scale. When those conversions cannot be defended, use constrained multi-objective selection and report a Pareto surface instead of inventing a scalar score. An action with high information gain but no expected decision benefit receives no credit.

The next figure uses a hypothetical action ledger to expose that rule. Point positions, costs, probabilities, and the highlighted action are analytical illustrations only; they are not measured values or a recommended controller.

Three runtime timescales

The event loop cannot raise its own power or bandwidth allocation. It requests capacity through the resource plane. The resource plane cannot change task semantics or fabricate confidence; it changes supply and prices. The context loop cannot transmit a per-module command table under the name “broadcast.” Equal mean state does not collapse delayed local regulation, anticipatory action, and slow structural change into one loop (C-1492, C-1494, C-1496).

A timescale label is not a closure certificate

Calling a loop fast, slow, or coarse describes its schedule. It does not show that the state exposed to that loop is sufficient to predict its own future. Eliminating an unobserved variable can leave history, an unresolved initial-condition term, or lift-dependent evolution behind. The runtime may use a reduced path only after one of the following four contracts is made literal; otherwise it retains or restores the fuller state.

Projection can move omitted state into memory

For the dimensioned linear example

$x$ and $y$ share unit $U$, while $\alpha,\beta,\gamma,\lambda$ have unit $\mathrm{s}^{-1}$. Eliminating $y$ does not generally give a history-free equation:

The first term is instantaneous, the second retains unresolved initial state, and the third is memory. The kernel $K(\tau)=\beta\gamma e^{-\lambda\tau}$ has unit $\mathrm{s}^{-2}$, so its integral has unit $U\,\mathrm{s}^{-1}$ like $\dot x$. The general projection result also contains an orthogonal-dynamics term; calling it “noise” does not make it independent, Gaussian, or negligible (C-1526). A finite history window $H$ seconds therefore needs a measured tail-error boundary, not a convenient buffer size.

For the single analytical kernel $K(\tau)=K_0e^{-\tau/\tau_m}$, with memory time $\tau_m$ seconds, the normalized mass omitted after retaining $H$ seconds is $R(H)=e^{-H/\tau_m}$. The figure is an exact illustration of that one kernel; it is not evidence that an artificial workload has exponential or finite memory.

A slow manifold has a geometric boundary

A named fast/slow split also does not establish a slow manifold. In the dimensionless fold normal form

prime denotes differentiation with respect to $\theta=t/\tau_0$, where the declared reference time $\tau_0$ is measured in seconds, and $\varepsilon$ is the dimensionless fast/slow timescale ratio. On the attracting critical branch $x^*(y)=\sqrt y$, the dimensionless normal spectral margin is $\gamma_N(y)=2\sqrt y$. It tends to zero as $y\to0^+$, while the sensitivity $|dx^*/dy|=1/(2\sqrt y)$ diverges. Ordinary normal-hyperbolic persistence is therefore qualified only on a declared compact region whose margin stays away from zero; it does not continue through the fold by naming the route “slow” (C-1527).

Coarse computation must expose how detail returns

Two different execution contracts cover cases in which a coarse state remains useful without pretending that fine detail vanished:

1. Local micro-query contract. A heterogeneous multiscale method declares a compression $Q_c:u\mapsto U$, a reconstruction $R(U,\xi)\mapsto u$, and the consistency residual $\|Q_cR(U,\xi)-U\|$. Here $u$ and $U$ retain their native fine- and coarse-state units and $\xi$ indexes admissible unresolved detail. When the macro update needs an unavailable datum such as a flux, it runs a bounded local microproblem to estimate that datum. Microcell support, boundary treatment, relaxation, sampling, failed solves, coefficient queries, iterations, bytes, and reconstruction error all remain in the cost and uncertainty ledger (C-1528). “Local” is not synonymous with “cheaper.”
2. Lift--heal--evolve--restrict contract. With fine propagator $\Phi_T^f$ over burst time $T$ seconds, the lift-specific coarse map is $\Phi_T^c(U;\xi)=Q_c\Phi_T^f(R(U,\xi))$. Several admissible lifts with the same $U$ are evolved through a declared healing time $t_h$ seconds before restriction. Post-healing disagreement is measured in the native norm of $U$. If materially different lifts still give different coarse derivatives or rollouts, the proposed coarse variables are not closed at that state and horizon; the route must abstain, add state, or fall back to fine evolution while charging all attempted work (C-1529).

The complete equations, dimensional checks, and non-overlapping compute, traffic, and microstep ledgers are in the multiscale reduction contract. Fixture F-024 registers matched-information tests against Markov, full-state, analytic homogenization, identified coarse-state, and continuously fine nulls. NO_RESULT: the fixture is pre-implementation; neither plot contains measurements, and this section establishes no accuracy, compute, traffic, energy, or readiness result.

Concrete per-event sequence

1. Timestamp and classify the event. Record source, modality, freshness, deadline, and risk class. Missing timestamps make prediction error and queue latency ambiguous.
2. Run a local change test. Compare the event with the last accepted local state using a sensor-noise model. Below-threshold change creates a no-op or cached-prediction candidate, not an automatic exit.
3. Predict before incorporating the event. Produce $\hat y_{t|t-1}$, residual covariance $S_t$, and the task outputs available at the cheapest depth.
4. Decompose mismatch. Compute $r_t$ and $\nu_t$; update a separate sequential change statistic for persistent drift; estimate epistemic uncertainty and task consequence.
5. Test the cheapest valid exit. Exit only if calibration for the event's risk stratum satisfies its error bound and no mandatory provenance, freshness, or tool check remains.
6. Update the context receivers. If a fast or slow broadcast update is due, each subscribed module applies its own causal filter, bounded gain, and local-state-dependent threshold. The received bitstream and local response are logged.
7. Value the remaining acquisitions. Estimate $\Delta U_t(a)$ and full costs for the available experts, memories, sensors, tools, and escalation paths. Reject actions whose expected value does not exceed their priced cost or whose completion misses the deadline.
8. Route under capacity constraints. Select the smallest useful set of modules. Include queue delay, expert load, communication, and state placement. Open reserve routes before tail latency collapses only if the measured congestion rule warrants it; the ant result in C-035 is a lead, not the routing algorithm.
9. Execute with local backpressure. A module consumes its local compute, memory, and bandwidth budget. Exhaustion triggers deferral, a cheaper approximation, or escalation; it never silently drops a safety check.
10. Reconcile prediction and cost. Update output, uncertainty, and the measured resource ledger. If the risk bound is unmet and deadline and budget remain, return to acquisition valuation. The loop has a fixed maximum iteration count.
11. Emit an attributable result. Return prediction or action together with confidence calibration version, routes used, evidence provenance, elapsed time, and measured or allocated energy. Runtime telemetry enters later maintenance analysis; it does not directly rewrite slow weights.

Control topology

flowchart TB
    subgraph fast["Event loop"]
        direction LR
        event["Timestamped event"] --> predict["Predict change · uncertainty · risk"]
        predict --> decide{"Exit or acquire?"}
        decide -->|"exit"| output["Attributed output"]
        decide -->|"acquire"| work["Expert · memory · sensor · tool"]
        work --> update["Update result + measured cost"]
        update --> decide
        decide -->|"budget exhausted"| fallback["Safe fallback / explicit failure"]
    end
    subgraph context["Rate-limited context loop"]
        direction LR
        global["Operating observations"] --> encoder["Fast + slow encoder"]
        encoder --> receivers["Receiver-local filters + gains"]
    end
    subgraph resource["Slower resource loop"]
        direction LR
        telemetry["Queues · bytes · energy · temperature"] --> controller["Resource controller"]
        controller --> prices["Capacity · prices · reserve · backpressure"]
    end
    receivers -.-> decide
    receivers -.-> work
    work --> telemetry
    prices -.-> decide
    prices -.-> work

Authority follows information and remaining capability

Power-grid protection supplies a hard engineering version of local reflex and global escalation. Local digital relays can act quickly inside a declared zone (C-186), while wide-area schemes add broader but delayed and failure-prone evidence (C-190). Fast response is still constrained by current, headroom, energy, duration, and interacting controls (C-196). The regional/global physiology boundary sharpens the same rule: a useful local correction must still expose shared resistance, exported load, trigger prevalence, and fallback cost (C-1493).

The held translation assigns controller $i$ an admissible action set

where $\tau_i$ is observation age in seconds, $q_i$ is integrity state, $m_i$ is operating-mode state, $b_i$ is remaining physical or compute headroom, $c_i$ is coordination availability, $\mathcal E_i$ is a certified set-valued map, and $u_i$ is the proposed action. The vector components retain their own units, timestamps, uncertainty, and provenance; the notation does not make them interchangeable.

Stale evidence, lost integrity, exhausted reserve, or lost coordination should shrink authority toward an independently enforced fallback. A wider action requires validated handoff and a checked postcondition. The candidate loses if adaptive protection, gain scheduling, constrained control, barrier functions, or runtime assurance reproduce the same frontier. Candidate 012 tests this under saturation, hidden failure, attack, communication loss, partition, second events, and staged recovery.

Context is a constrained interface

The context candidate sends a small quantized code at declared fast and slow rates. Receivers expand it locally through distinct filters and bounded transfer functions. This is useful only when receiver-local structure reconstructs module-specific responses more cheaply than transmitting them.

Candidate 002 tests that proposition against instantaneous and shared-filter FiLM, memoryless and recurrent gates, bottleneck and standard global tokens, low-rank hypernetworks, fixed multirate receiver banks, and gain-scheduled supervisory control. Its primary channel is two 8-bit components at different cadences, with an average logical rate of $825\ \mathrm{bit\,s^{-1}}$. That number is an experiment setting, not a proposed universal bus rate.

A bounded interface still needs a back-action test

For the same upstream input, initial state and random seed, characterize the producer first without a consumer and then after attaching one. Keep producer trajectory distortion separate from delivery latency, consumer fidelity, shared-resource contention and any intended feedback. A connection is not harmful merely because it consumes work, and a producer is not insulated when its consumer receives no usable service.

The biological evidence supplies several scoped warning cases: downstream binding pools can alter transcriptional transients, signalling targets can change the operating regime or bandwidth of an upstream cycle, and fast intermediate processes can attenuate some load effects (C-1550–C-1555). The systems translation starts with stronger ordinary controls:

1. an ideal immutable-copy path with no finite-interface backpressure;
2. a tuned bounded ring buffer and explicit backpressure;
3. asynchronous snapshots, copy-on-write and expiry;
4. load-aware admission and resource reservation;
5. unrelated CPU or memory contention without a producer connection; and
6. explicit filtering or intended feedback where temporal shaping is useful.

Only connection-specific distortion that survives those controls is called a retroactivity analogue. Fixture F-027 and its mathematical contract freeze that comparison. The first executable slice is a public diagnostic and has no performance or energy authority.

Deficit travels; feasibility remains local

Plant nitrogen acquisition supplies a precise three-part allocation loop: nitrogen-starved roots send CEP deficit signals upward, shoot-derived CEPD signals return global context, and high-affinity uptake increases only where a root also encounters nitrate (C-207). This is not evidence for a new controller. Backpressure and primal–dual allocation are the strongest nulls.

For module $i$, let $d_i(t)$ be unmet task demand per second, $a_i(t)\in[0,1]$ be local capability availability, $u_i(t)$ be allocated compute-seconds per second or watts, and $B(t)$ be the total budget in the same unit as $u_i$:

The possible advantage is communication structure: compress deficits upward, return a small scarcity code downward, and keep the expensive feasibility decision at the receiver. It should lose when complete state is fresh, capability is uniform, or backpressure already carries the required deficit and feasibility information more cheaply. C-1488 therefore keeps aggregate supply, local arrival, usable uptake, and export separate: a global balance cannot certify receiver feasibility. Candidate 013 tests moving demand and resource patches, delay, topology churn, strategic over-reporting, reversible allocation, slow growth, and second events.

Electrochemical depletion adds a geometry-sensitive failure case. A healthy global mean can conceal a receiving boundary whose local carrier supply is approaching zero; once the support assumptions change, positive feedback can amplify a protrusion rather than merely serve more demand (C-1536). The artificial analogue must therefore bind every growth or allocation action to a local support estimate and expose an unknown or fallback state when that estimate is not observable. Ordinary queue, backpressure, and robust load controllers remain the nulls, and the local estimator's sensing, messages, throttling, and false alarms are charged.

Engineering null models

Biological language is removed when an established method explains the result at the same cost.

The equations and unit requirements for innovation, value of information, CUSUM, allocation, and energy are recorded in the engineering-analogue audit.

Adaptive resolution must be target- and regime-qualified

Fluid dynamics is a hostile test for “allocate more compute where prediction error is high.” The governing equations can be known while the useful state is still limited by unresolved scales, boundary and forcing uncertainty, discretization, closure error, partial observation, chaotic amplification, regime change, intermittency, and rare extremes. The audited evidence in C-881–C-925 therefore tightens the allocator contract.

An allocation signal is valid only relative to a declared target and measurement identity:

• variable and support: which physical or latent variable, spatial region, temporal window, and averaging kernel define the target;
• filter and detector: which scale, threshold, reference frame, event detector, and observation operator define a residual or coherent event;
• signed relation: whether energy, information, error, or another invariant should move toward larger or smaller scales under the current regime;
• model boundary: which part is resolved dynamics, closure, discretization, observation error, and out-of-support model-form discrepancy;
• quantity of interest: whether the reduced state must preserve average field reconstruction, control authority, transition timing, mixing, or an extreme tail; and
• regime and history: geometry, forcing, ramp direction, dwell, disturbance, prior occupancy, sensor drift, actuator condition, and current uncertainty.

A mean residual can be small while the signed scale flux is wrong, the tail is miscalibrated, or the rare event that matters is missed. A high-energy reduced basis can discard a weak direction that controls transition or actuation. A mesh can use fewer cells at one instant while spending more on regridding, subcycling, synchronization, load imbalance, data transfer, and failed solves. These are different failures and stay separate in the outcome vector.

flowchart LR
    truth["Reference plant or simulation<br/>equations · geometry · boundaries · forcing"] --> observe["Versioned observation operator<br/>filter · kernel · support · latency · covariance"]
    regime["Hidden regime and history<br/>forcing · ramp · disturbance · dwell"] --> truth
    truth --> unresolved["Resolved state + unresolved scales<br/>discretization · closure · model-form error"]
    observe --> assimilate["Estimate and calibrate<br/>EnKF · 4D-Var · moving horizon · learned estimator"]
    unresolved --> infer["Forecast and represent<br/>DNS/RANS/LES · ROM · operator surrogate"]
    assimilate --> infer
    infer --> allocate{"Allocate scarce resolution?"}
    allocate --> mesh["Mesh and compute<br/>uniform · AMR · goal-oriented"]
    allocate --> sensors["Sensors and bandwidth<br/>fixed · optimal · adaptive"]
    mesh --> truth
    sensors --> observe
    infer --> act["Act or abstain<br/>passive · robust · MPC · learned control"]
    act --> truth
    infer --> firewall["Outcome firewall<br/>field · signed flux · tails · coherent events<br/>closure · ROM · refinement · assimilation · sensors<br/>control · mixing · transition · extremes · measurement · energy"]
    truth --> firewall
    nulls["Reference/DNS · RANS/LES · POD/DMD/resolvent<br/>AMR · EnKF/4D-Var · adjoint placement<br/>LQG/H∞/MPC · passive control · learned baselines"] --> compare{"Equal information and lifecycle budget"}
    firewall --> compare
    meter["Complete ledger<br/>cell-steps · samples · bytes · seconds · person-hours · joules"] --> compare
    compare --> decision["Retain measured residual<br/>or retire the composition"]

Editable source: regime-qualified-flow-inference-control.mmd.

Fixture F-005 crosses ten adversarial tracks: signed multiscale transfer, closure portability, target-qualified reduced state, adaptive resolution, assimilation, sensor placement, closed-loop control, mixing, path-dependent transition, and extreme prediction. Its decisive comparator is a complete composition of reference simulation, classical RANS/LES closures, POD/DMD/resolvent or balanced reduction, goal-oriented AMR, EnKF/4D-Var and moving-horizon estimation, adjoint/Fisher/Gramian sensor placement, robust/MPC/passive control, and learned operators or policies.

The flow mathematics binds every result to geometry, equation, boundary, forcing, solver, grid, measurement, filter, detector, data lineage, actuator, target, regime, and resource identity. An adaptive mechanism earns credit only when it improves its preregistered target beyond that complete stack without reversing flux, losing stability, undercovering uncertainty, missing the sealed natural tail, or producing negative net lifecycle energy. Otherwise the ordinary method remains and the proposed composition is retired.

Efficiency mechanism

Sparse work has four independent levers

The runtime can save resources by:

1. suppressing unchanged events;
2. exiting at a cheaper validated depth;
3. activating fewer modules or memory paths; and
4. avoiding movement by keeping repeated work near its state.

These levers are not interchangeable. An early exit can reduce arithmetic while leaving input and cache traffic unchanged. Sparse experts can reduce active weights while increasing all-to-all communication. Local placement can reduce bytes moved without reducing operations.

For module gates $g_i(x)\in[0,1]$ and declared operation costs $c_i(x)$, define the operation-weighted active fraction

Also report byte-weighted activity $\rho_{\mathrm{bytes}}$ from measured memory and network traffic. Neither quantity is an energy estimate.

Charge every control path

Over the same measurement interval,

Every term is in joules. The runtime produces an efficiency gain only when

at matched quality, risk, latency class, batch opportunity, and hardware boundary. Report board, node, and facility energy separately when measured; do not infer one from another.

State movement and controller work must also amortize. For a placement, compilation, or cache migration completed $N$ times, let $E_{\mathrm{runtime}}$ be incremental runtime energy in joules per completed use; let $E_{\mathrm{discover}}$, $E_{\mathrm{migrate}}$, and $E_{\mathrm{validate}}$ be one-time energies in joules; and let $N$ be a dimensionless completed-use count. Then

$\bar E_{\mathrm{use}}$ is joules per completed use. If $N$ is not observed, report the break-even reuse count rather than claiming a saving.

Resource supply is a feedback problem

For one optimization instance, choose a single rate basis—requests per second, tokens per second, or bits per second—and let $x_i$ be module $i$'s allocation in that unit. Let $U_i(x_i)$ be a nondecreasing concave utility normalized to a common dimensionless scale, let $R$ be a dimensionless resource-incidence matrix, and let each component of $c$ be capacity in the same chosen rate unit. A conventional allocation null is

The slower resource plane observes queues, deadlines, utilization, energy, and temperature, then adjusts capacity or prices. Any learned controller must beat a tuned PI or primal/dual implementation on settling time, overshoot, constraint violations, tail latency, and joules per control update. “Local metabolism” is not a substitute for that comparison. Interacting loops with gain and delay must also be tested for oscillation and stability rather than accepted from their mean allocation (C-1492). Regime-dependent mediator supply must retain starvation, drag, leakage, failure, and pumping terms (C-1499); mean load also cannot clear a state-delay instability such as stick--slip (C-1501).

Evidence status

The integrated runtime is speculative until its gates are tested separately and then recombined under one measurement boundary. A combined win cannot identify which control path caused it.

Speculative extensions

Residual queues instead of global lockstep

Each module could own a queue of unresolved residuals and wake only when local work, context, or a deadline changes its priority. The comparator is an efficient batched scheduler with the same queue and dispatch overhead. The extension is abandoned if asynchronous launches reduce utilization or increase p95 latency enough to erase skipped work.

Predictive placement

Frequently reused expert weights, key–value state, and memories could move toward their expected consumers. The hypothesis is about reuse-distance and migration amortization, not artificial mitochondria. Static placement, LRU, and NUMA-aware policies remain required baselines; every migration is charged in bytes, seconds, and joules.

Controlled recovery probes

A maintenance observer could inject small bounded perturbations and measure recovery time, overshoot, and restoring margin before ordinary quality metrics fail. This is specified in Candidate 003. Recovery diagnostics do not themselves stabilize the runtime, and their probe energy and task disturbance remain part of the cost.

Cross-modal acquisition brokerage

A common valuation layer could choose between deeper internal computation, a memory lookup, another sensor view, a physical action, a tool call, or a human query. Each option needs a calibrated outcome model and a common decision contract. Entropy reduction alone is not sufficient because an observation can be surprising yet irrelevant to the pending action.

Failure modes

Measurable predictions

These are experiment commitments. Directional improvements require paired uncertainty intervals and a predeclared practical margin; a lower theoretical operation count is not a pass.

Mechanism ablations are interpreted selectively. Removing the exit gate should increase depth without changing routing identity; removing sparse routing should increase active modules and traffic; removing the context fast path should selectively damage transient response; removing the slow path should selectively damage sustained response; freezing resource control should worsen burst recovery rather than semantic accuracy. If every ablation merely lowers capacity and hurts everything, the architecture has not isolated its claimed control loops.