Representative adaptive performance

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.