PaperlayerView exact revision on arXiv

Source-linked AI summary

Auditability Is Not One Property: Rule Overlap, Behavioural Agreement, and Composition in Reinforcement Learning

Liu Hung Ming

arXiv:2609.28581v1cs.LGcs.AI

TL;DR

This paper asks how independently trained RL policies can be described and composed when checkpoints are opaque and logs do not explain what was learned. It builds an auditable rule-fusion protocol, then shows that symbolic overlap does not establish behavioral agreement and that composition results remain bounded by task and evaluation conditions.

  • Problem

    Opaque policy checkpoints and training logs provide limited evidence about what independently trained agents learned or how their behaviors compare.

    Source: §1 · p. 2, §1.1 · p. 2

  • Method

    The paper turns trained policies into rule banks anchored to replay-verified, hash-bound traces and uses offline arbitration to compose banks.

    Source: §7.6 · p. 33, §8 · p. 36

  • Results

    On CartPole-v1, rule fusion exceeded the post-hoc-selected single-actor comparator by +42.31 return points, an exploratory result that does not establish a general performance win.

    Source: §6.13 · p. 28, §8 · p. 36

  • Takeaways & Limitations

    The audit layer supports attribution in composition settings, not discovery of new behavior or skills.

    Source: §7.6 · p. 34

    The comparison uses eight seeds, a post-hoc-selected actor baseline, and a partly saturated CartPole task.

    Source: §7.6 · p. 31

Abstract

from arXiv · show

Reinforcement-learning (RL) policies are often distributed as opaque neural checkpoints, while training logs show that a run occurred without explaining what the policy learned. We study whether independently trained policies can be represented and composed through auditable discrete behavioral rules. We define auditability as six separately testable predicates: trace integrity, lossless coding, rule coverage, behavioral agreement, composition quality, and value-model reliability. Our protocol uses a shared frozen symbolizer, passive rule extraction, an append-only hash-bound ledger, exact environment replay, and offline confidence-ranked arbitration with an explicit blind-spot fallback. The results place strict limits on this description layer. Rule-set overlap does not imply behavioral agreement: policies may share symbolic rules while choosing near-chance-matching actions on fresh states. The fused policy therefore selects among existing rules rather than generating a new skill. On a conflict-dominated task, an apparent fusion failure is traced to an induction/deployment mismatch: rules induced from sampled actions were evaluated under argmax actions, and deployment-consistent re-induction reverses the arbitration ordering. A fitted-Q generalized-policy-improvement diagnostic also fails in both environments, limiting claims that rule fusion is superior to value-based composition. One exploratory comparison favors rule fusion, but its comparator is post hoc, the task is partly saturated, and the fused policy remains below the strongest held-out actor. We contribute an evidence-bounded audit and composition protocol, not a claim of universal interpretability or autonomous skill generation. Future work must add temporally extended skills, cross-skill interfaces, composition search, and independent novelty audits.

1 Introduction

The paper asks how independently trained RL policies can be described and composed when checkpoints are opaque and logs do not explain learned behavior. It proposes an auditable rule-level protocol while showing that rule overlap, behavioral agreement, and composition quality are distinct claims.

1.1 Auditability is not a single property

  • Auditability is six separately testable predicates, and a pass on one predicate does not license a claim about another.The paper distinguishes predicates from verification mechanisms and warns that collapsing auditability into one score hides separate negative results.Source: §1.1 · p. 2
  • Rule overlap does not imply behavioral agreement; the reported 0.999 mean symbol-level agreement on CartPole-v1 is a quantizer artifact, not evidence of a shared policy core.The intersection of all eight banks contains 38 conditions on CartPole-v1 and two on Acrobot-v1; on Acrobot-v1, neither condition receives the same action from all eight.Source: §1.1 · pp. 2, 4
  • The protocol describes policies as condition–action rule banks over a shared discretized state space, anchored to execution logs whose fidelity is operationally verified.Its components are a frozen shared symbolizer, passive rule observer, replay verifier and hash-bound ledger, and offline fusion with a blind-spot fallback.Source: §1.1 · p. 3
  • Earlier protocol attempts failed through lossy compression, training distortion, blind spots, log expansion, and a contaminated ledger schema.The authors document these failures as evidence that shaped the final protocol.Source: §1.1 · p. 3
  • On CartPole-v1, rule fusion exceeded the post-hoc-selected single-actor comparator by +42.31 return points, but the comparison is exploratory and fusion remained below the strongest held-out actor.The task exhibits ceiling effects, and the selected comparator was chosen post hoc by training return.Source: §1.1 · p. 3

2 Preliminaries and Problem Statement

The paper frames auditing as separate questions about trace fidelity, rule description, and composition. Its scope explicitly excludes stronger inferences from replay, symbolic overlap, or operational checks.

2.1 Markov decision process and agents

  • The setup models finite-horizon MDPs and trains two-hidden-layer, 64-unit agents with REINFORCE, using eight independent seeds throughout.A policy maps states to action distributions, and its value is expected return from the initial-state distribution.Source: §2.1 · p. 4

2.2 The audit problem

  • The audit problem asks whether traces can be certified, compressed into comparable rules across agents, and fused without inheriting source disagreements.The trace records state, action, reward, next state, termination, and truncation.Source: §2.2 · p. 4
  • Replay verifies recorded transitions but cannot recover policy weights, optimizer state, gradients, or the random-number state that produced them, nor establish recorder honesty.Trace fidelity, policy description, behavioral agreement, provenance, and causal explanation are distinct evidentiary layers.Source: §2.2 · p. 5

2.3 Scope and non-claims

  • The study limits positive claims to two benchmark tasks, eight seeds per task, one shared quantizer per task, and one software/hardware configuration.Operational-health checks certify the artifact, not the policy.Source: §2.3 · p. 5

3 Related Work

Related work spans policy description, value-based composition, symbolic communication, and audit trails. The paper distinguishes its rule-arbitration objective from behaviorally equivalent extraction, theorem-backed GPI, learned communication, and new cryptographic contributions.

Policy description and interpretability of RL

  • Unlike interpretability methods that produce saliency maps or behaviorally equivalent extracted policies, this study evaluates trace anchoring, cross-seed comparability, provenance, and whether rule overlap predicts behavior.VIPER uses imitation-style extraction guided by a neural policy and value estimates [6]; the paper does not claim its rule bank replaces the neural policy.Source: §3 · p. 5

Value-based composition

  • Rule fusion ranks candidate actions by empirical confidence and support, not source-policy action values, so it has no GPI-style policy-improvement guarantee.The fitted-Q comparator is a separate baseline and its failure is treated as an implementation validity diagnostic, not evidence against value-based composition generally.Source: §3 · pp. 5–6

Discrete symbolic communication and auditability

  • The study uses a fixed calibration-derived abstraction rather than emergent communication or a learned codebook, and makes no claim about learned symbol systems.Its causal-symbol test is a positive demonstration of answering a causal question by intervention in a designed task.Source: §3 · p. 6

Auditability of training records

  • The ledger contribution is a hash-bound ledger over canonical payloads plus exact environment replay, not new cryptography.The work reports practical reproducibility findings, including serialization size inflation and grammar-style codecs that can fail to compress traces.Source: §3 · p. 6

4 Method

The method fixes a shared state abstraction before training, then passively induces provenance-linked rules from traces. Admission separates rule support and confidence from coverage, with explicit safeguards against blind spots.

4.1 Shared state symbolizer

  • A single symbolizer fitted from pooled random-policy calibration rollouts is frozen so all agents and experimental arms within a task use the same four-bin-per-dimension vocabulary.The resulting vocabularies have 256 symbols for CartPole-v1 and 4096 for Acrobot-v1.Source: §4.1 · pp. 6–7
  • A shared symbolizer makes labels comparable, but symbol overlap is not behavioral equivalence because policies can disagree within a symbol's raw-state region.The paper reports a maximum observed edge difference of 0.157 observation units under per-seed fitting, motivating the single frozen symbolizer.Source: §4.1 · pp. 7–8

4.2 Rule bank and admission

  • A rule is a condition–action pair derived from logged transitions, and admission deterministically compresses trace statistics without sampling.The induced action is the most-counted action for a condition, with ties going to the smallest index; support and confidence are reported separately from coverage.Source: §4.2 · p. 8
  • Each admitted rule retains the multiset of source step identifiers and its digest, linking the rule to its provenance.Source: §4.2 · p. 8
  • At τc = 0.90, strict admission sent all held-out decisions to fallback and produced a 100% blind-spot collapse, so confidence is not safety or generalizability.At τc = 0.70 with τs = 8 and an explicit fallback, coverage failure becomes observable rather than silent.Source: §4.2 · p. 8

4.3 Audit ledger and replay verification

The audit checks trace fidelity through exact environment replay, a hash-bound ledger, and per-decision provenance resolution. These checks are operationally scoped to the recorded artifacts and do not certify training honesty or policy semantics.

4.3 Audit ledger and replay verification

  • Exact replay compares regenerated state, reward, termination, and truncation against each recorded transition under its reset seed.Equality is exact float32 equality; replay checks the executed trace, not whether the policy is good.Source: §4.3 · p. 8
  • The SHA-256 ledger chains canonicalized training events and checkpoint hashes, so a mismatch invalidates the run and links optimizer steps to weight states.The update hash must cover optimization-relevant tensors, not metadata; normalizing the core makes optimizer-equivalent runs byte-identical.Source: §4.3 · p. 9
  • Hash integrity, environment replay, and provenance resolution are separate proofs, respectively checking record alteration, transition regeneration, and decision-to-rule traceability.They must not be collapsed into a single audit score.Source: §4.3 · p. 9
  • Certification is relative to the stored artifact, environment version, reset seeds, and verifier; it does not establish honest training or faithful semantic explanation.It also does not certify the training harness, optimizer state, random-number state, or neural policy mechanism.Source: §4.3 · p. 9
  • Replay-and-provenance audit passes only when every transition replays, every decision resolves to an admitted source rule, and the hash chain recomputes exactly.The predicate is scoped to the reported eight-seed, two-environment study and does not assert that the training-time logger was honest.Source: §4.3 · p. 9

4.4 Lossless trajectory codec

The trajectory codec is an exact, bounded-rule grammar for reconstructing logged traces, not the state–action rule bank or an explanation of the neural policy.

4.4 Lossless trajectory codec

  • The Re-Pair-style codec uses at most 96 grammar rules and asserts exact decoding by checking that the decoded payload hashes to the input hash.It compresses the trace in addition to the rule bank, with provenance bytes reported separately.Source: §4.4 · p. 10
  • The two-part description length adds the grammar description to the trace encoded under that grammar, with provenance bytes reported separately.Terminal counts use Elias-gamma coding and non-terminal references use fixed width.Source: §4.4 · p. 10
  • The induced grammar reconstructs the observed sequence exactly but is neither the policy’s rule bank nor an explanation, and the method claims no optimal MDL model selection.In practice, the codec does not beat generic gzip on size.Source: §4.4 · p. 10

4.5 Training arms

The study compares unconstrained training, passive rule observation, and rule-constrained training, then fuses rule banks offline through confidence-ranked arbitration with an explicit blind-spot fallback.

4.5 Training arms

  • Each seed trains an unconstrained REINFORCE baseline, a byte-identical audit-only observer, and a constrained arm that masks actions to rules extracted from earlier traces.The constrained arm tests the design intuition that rules should steer training and quantifies its cost.Source: §4.5 · p. 10

4.6 Rule-level policy fusion

  • At a state with no matching rule, fusion uses the argmax of mean source-actor logits; otherwise it selects the candidate with highest confidence, then support, mean reward, and source index.The selected rule is an existing rule, with its provenance recorded.Source: §4.6 · p. 10
  • When candidate rules disagree, the arbiter chooses by confidence, support, mean reward, and source index, then logs the outcome and updates the chain digest.Seeds count as conflicting when they recommend different actions for the same condition.Source: §4.6 · pp. 10–11
  • Fusion records agreement, single-source, conflict, blind-spot, or fallback decisions because each class conditions interpretation of the result.A fused return without coverage, conflict, blind-spot, fallback, source-distribution, and arbitration reporting is incomplete.Source: §4.6 · p. 11

4.7 Analysis instruments

The analysis separates fresh-state behavioral agreement, symbolic rule overlap, and fitted-Q generalized-policy-improvement diagnostics as distinct instruments.

4.7 Analysis instruments

  • Behavioral agreement measures how often two actors choose the same action on fresh states sampled from their own occupancies.The protocol also reports an episode-cluster-balanced estimate and bootstrap interval.Source: §4.7 · p. 11
  • Rule-set overlap uses Jaccard similarity between exact state–action sets and separately counts shared conditions with conflicting recommended actions.Source: §4.7 · p. 11
  • The fitted-Q GPI diagnostic fits action-value regressors over pooled replay transitions, with a reservoir cap of 25,000 transitions per seed and γ = 0.99.It is approximate; bounded value-estimation error is not controlled, so critic failure is not evidence about exact GPI.Source: §4.7 · pp. 11–12

4.8 Matched-support causal symbol intervention

The communication experiment estimates the local effect of replacing a message on receiver action probability under a matched-support intervention. Composition evaluations use paired held-out episodes and percentile-bootstrap intervals.

4.8 Matched-support causal symbol intervention

  • Replacing a sender’s message with a shuffled one measures its local effect on receiver action probability while holding task state and receiver context fixed as implemented.The designed game rewards correct identification, so ignoring the message cannot be optimal; the intervention does not establish human-interpretable meaning or generalize to arbitrary neural policies.Source: §4.8 · p. 12
  • The communication study’s causal claims are gated through pre-registered checks from diversity and positive control through necessity, generalisation, causality, and channel exclusivity.The evidence is bounded to the redesigned game and depends on common-support and no-private-channel checks.Source: §4.8 · p. 12

4.9 Statistical conventions

  • Composition comparisons use 100 paired held-out episodes per policy, with 95% percentile-bootstrap intervals over 5,000 resamples and no cross-seed variance claim.The seed pool of eight is too small for cross-seed variance estimates; deterministic replay equality and hash validity have structurally zero variance.Source: §4.9 · p. 12

5 From Intuition to Protocol: a Failure-Driven Design History

The section documents failures that shaped the protocol, separating fidelity from description and recording where rule constraints, audit artifacts, arbitration, and task design fell short.

5.1 Compression, replay, and explanation are different claims

  • A faithful trace is not itself a behavioral description; the protocol therefore keeps trace fidelity separate from rule-bank compression.A CartPole trace can contain about 97,646 transitions, beyond practical human inspection.Source: §5.1 · pp. 12–13

5.2 First-generation prototype failures

  • The first prototype used lossy sequence compression, conflated per-decision rules with sequence patterns, and checked aggregate replay statistics instead of exact transitions.The protocol dropped grammar compression, retained only per-decision condition–action rules, and changed replay verification to exact float32 per-transition equality.Source: §5.2 · p. 13

5.3 Rule-constrained training interferes with behavior

  • Rule-constrained training had heterogeneous CartPole returns, while Acrobot remained at its −500 step limit in both arms.The mask could discard gradient signal and re-admitted rules from mask-generated observations; the hybrid and network-only evaluations also answered different questions, motivating offline composition rather than a general claim that rules are harmful.Source: §5.3 · pp. 13–14, Table 2 · p. 15
  • Coverage rose in all eight seeds while covered-state action agreement fell, so the constraint narrowed policy occupancy rather than rule-bank reach.The arms were measured on their own state occupancies, not matched states; the threshold-relaxation test rejected low bank coverage as the cause of damage.Source: §5.3 · pp. 13–14, Figure 3 · p. 17

5.4 Confidence thresholds and the 100% blind-spot collapse

  • A strict confidence threshold left fusion falling back on every held-out decision because no condition was covered.The protocol used τc = 0.70, τs = 8, and an explicit blind-spot fallback; confidence and coverage do not validate rule correctness.Source: §5.4 · p. 15

5.5 Serialization and ledger-schema failures

  • The trace ledger remained 8.4% larger on CartPole and 9.9% larger on Acrobot than gzip-compressed raw traces, despite exact decoding.A schema repair limited update hashes to a normalized core record so non-core annotations could not change hashes between equivalent runs.Source: §5.5 · p. 15

5.6 Arbitration sensitivity and the random-selection anomaly

  • On Acrobot, random arbitration beat tested deterministic variants under sampled-bank induction, but greedy-label re-induction reversed the ordering.The deployed confidence-ranked fusion scored −141.77 versus −234.26 for random selection after re-induction, identifying an induction/deployment mismatch.Source: §5.6 · p. 16

5.7 Communication-task redesign

  • The original parity game collapsed to constant actions, and an early redesigned run showed chance-level accuracy after only 79 updates before 12,500 updates succeeded.The abandoned private-target design also exposed payoff, masking, codebook-index, and continuous-path problems that prompted redesign.Source: §5.7 · pp. 16–17

6 Experiments and Results

Across two control benchmarks, the experiments separate artifact integrity, rule structure, behavioral agreement, and composition quality, with outcomes that depend on task and induction protocol. Rule overlap did not establish agreement on fresh states, and the apparent Acrobot arbitration failure reversed under deployment-consistent re-induction.

6.3 Audit integrity / 6.12 Lossless coding results

  • Hash chains, ledgers, replay, and lossless codec round-trips passed, but the grammar–gzip archive was 8.4% larger than gzip on CartPole-v1.The codec is presented as a fidelity tool for exact replay-against-compressed-form checks, not as a storage win.Source: §6.3 · p. 19, §6.12 · pp. 27–28

6.1 Setup / 6.2 Registered predictions / 6.4 Rule structure across seeds

  • Across eight seeds, CartPole-v1 banks shared 38 unanimous conditions and mean Jaccard overlap 0.56, whereas Acrobot-v1 shared only two conditions, both conflicting, with mean overlap 0.096.The eight-seed experiments used a shared frozen symbolizer and held-out matched episodes; predictions were unregistered, and the CartPole best-single-actor verdict depended on comparator choice.Source: §6.1 · p. 17, §6.2 · p. 18, §6.4 · p. 19

6.5 Composition on CartPole-v1

  • 205.96 mean held-out return for CartPole-v1 rule fusion versus 163.65 for the post-hoc training-return-selected actor, a paired gain of +42.31 [23.98, 61.93]; fusion was 294.04 points below the best held-out actor.Four individual actors reached the 500-step ceiling; the reported gain is exploratory and comparator-dependent.Source: §6.5 · p. 20, Table 7 · p. 21, Table 8 · p. 21

6.6 Acrobot-v1: the conflict regime and the failure of confidence arbitration

  • On Acrobot-v1, all source actors returned −500 while fusion averaged −443.37, but 89.2% of decisions conflicted and random rule arbitration substantially outperformed confidence-ranked arbitration.Confidence-ranked fusion returned −443.37; random arbitration returned −187.02 on the primary evaluation.Source: §6.6 · p. 21, Table 9 · p. 23

6.7 Protocol-consistent induction recovers Acrobot fusion

  • Greedy-label re-induction reversed Acrobot arbitration: confidence-ranked fusion reached −141.77 versus −234.26 for random arbitration, while coverage and conflict rates remained nearly unchanged.The reported gain is attributed to induction-protocol consistency; on CartPole-v1, the sampled bank remained marginally better, 205.96 versus 191.41.Source: §6.7 · pp. 22–23, Table 11 · p. 24

6.8 Behavioral agreement: rule overlap does not transfer / 6.9 Opportunity baseline: what a single rule bank already covers

  • CartPole-v1 banks with Jaccard overlap 0.58 agreed on only 51.38% of actions across 10,000 fresh states, essentially the 51.32% action-frequency baseline; symbol-level agreement of 0.999 was a quantizer artifact.A single bank accepted an average 59.7% of the observed CartPole vocabulary and 34.8% of the Acrobot vocabulary; the all-bank intersection was 38 conditions on CartPole and two conflicting conditions on Acrobot.Source: §6.2 · p. 18, §6.8 · pp. 23–24, §6.9 · p. 26

6.10 Causal symbol intervention as a separate sub-study

  • In the separate redesigned communication task, receiver accuracy was 0.998923 with the true message, 0.503292 with shuffled messages, and 0.500 without a message; the causal claim is limited to that task.The study does not establish that symbols are human-interpretable or generalize to other RL agents.Source: §6.10 · p. 27

6.11 Fitted-Q GPI baseline failure / 6.13 Summary of verdicts

  • Fitted-Q GPI failed in both environments, so its diagnostic does not support a claim that rule fusion is superior to value-based composition; the summary distinguishes identification-bearing results from diagnostics and health checks.The headline CartPole comparison remains exploratory and reverses against the best held-out actor.Source: §6.11 · p. 27, §6.13 · p. 28

7 Discussion

The discussion separates what the rule-fusion protocol establishes from what it does not: shared rules do not guarantee behavioral agreement, and fusion outcomes depend on task and induction protocol. It treats the results as bounded evidence for an audit and composition layer, not generative skill learning.

7.1 Why the mechanism may work, and where it breaks

  • On Acrobot-v1, greedy-label re-induction reverses the arbitration ordering: confidence-ranked fusion returns −141.77 versus −234.26 for random arbitration, correcting a sampled-action/argmax deployment mismatch.The earlier 22.5% Monte-Carlo agreement came from a two-actor continuation probe and does not characterize the deployed eight-actor arbiter.Source: §7.1 · pp. 28–29

7.2 Prediction outcomes

  • Rule overlap still fails to predict fresh-state behavioral agreement (51.4% versus chance); opposing near-uniform action preferences across actors’ occupancies are one proposed mechanism, not a unique explanation.The four-bin symbolizer remains a possible confounder because policies can disagree within a shared symbol.Source: §7.2 · p. 29

7.3 What the protocol does not establish

  • Replay and provenance do not establish causal explanation or neural-mechanism recovery, and fitted-Q GPI failure does not establish rule fusion’s superiority; inspectability is the stated trade-off over an actor-logit ensemble.The ensemble scores 500.00 versus 205.96 for rule fusion on CartPole-v1, but cannot report per-decision disagreement or blind spots.Source: §7.3 · pp. 29–30

7.4 Same framework, task-specific regimes

  • Both tasks use the same provenance-preserving fusion skeleton, but task-specific banks and conflict regimes differ; protocol-consistent greedy induction moves Acrobot-v1 from −443.37 to −141.77.The changed result reflects aligned induction and deployment, not a different fusion algorithm.Source: §7.4 · p. 30

7.5 Theoretical and practical implications

  • The six audit predicates should be reported separately: trace integrity can pass while behavioral agreement fails, and composition can improve on CartPole yet collapse into conflict on Acrobot.Replay, hashes, and provenance provide evidence for selected predicates but are not substitutes for composition quality.Source: §7.5 · p. 31

7.6 Limitations

  • The evidence is bounded by eight seeds, cross-episode rather than cross-seed uncertainty, post-hoc actor selection, CartPole ceiling effects, and a description layer that lacks timing and cross-task skill interfaces.The paper frames temporal abstraction, cross-task interfaces, composition search, and independent novelty checks as future work, not solved capabilities.Source: §7.6 · pp. 31–34

7.7 Steelman

  • The authors defend the work as auditable, bounded evidence rather than a claim that fusion wins: the CartPole gain is exploratory, while Acrobot’s failure is reported under the same measurement discipline.The matched CartPole comparison’s bootstrap interval excludes zero, but its comparator was selected post hoc and the task is partly saturated.Source: §7.7 · p. 35

7.8 Replicability note

  • Replication depends on the selector stream, reset seeds, and symbolizer calibration; sharing a frozen symbolizer, pairing reset seeds, and repeating random selection mitigate identified risks.The random-selection condition was repeated across five independent streams.Source: §7.8 · pp. 35–36

Reproducibility Statement

The reproducibility statement identifies the experiment settings and released records needed to inspect or reproduce the work. It distinguishes archived artifacts and checksums from rerunning training.

Reproducibility Statement

  • Experiments use eight fixed control seeds, 100 matched reset seeds for composition comparisons, and six communication runs; rule induction uses four quantile bins, minimum support 8, and minimum confidence 0.70.Analyses not pre-registered are labeled exploratory.Source: Reproducibility Statement · p. 36
  • The code, compact hashed result package, and manuscript sources are public, while 584 files totaling 4.7 GB of training traces, checkpoints, rule banks, and related data are archived.The archive includes a SHA-256 checksum for verifying the downloaded copy.Source: Reproducibility Statement · p. 36
  • The compact result package records each headline number’s artifact path, byte count, and SHA-256 so reported values can be traced to archived files without rerunning training.Source: Reproducibility Statement · p. 36

8 Conclusion

The conclusion reports an operationally verified rule-level description and composition protocol, with results that differ across tasks. Its central boundary is that rule overlap does not establish behavioral agreement, and the audit layer verifies composition rather than generating novel skills.

8 Conclusion

  • On CartPole-v1, fusion exceeded the post-hoc-selected single-actor comparator by +42.31 return points (95% interval [23.98, 61.93]), an exploratory comparison under ceiling effects.Source: §8 · p. 36
  • On Acrobot-v1, confidence-ranked arbitration was anti-selective under the evaluated conflict and occupancy regime, and uniformly random arbitration outperformed it.A fitted-Q GPI baseline collapses in both environments.Source: §8 · pp. 36–37
  • The protocol is useful in low-conflict regimes as an auditable composition layer, but composition remains environment-dependent and the present work does not claim novel-skill generation.Source: §8 · p. 37

A Per-Seed AIM Causal-Intervention Appendix

The appendix reports a per-seed matched-token intervention on a designed private-target communication game. Its rows expose variation across runs rather than relying only on an aggregate.

A Per-Seed AIM Causal-Intervention Appendix

Loading 2609.28581v1…