PaperlayerView exact revision on arXiv

Source-linked AI summary

FAER: Auditable Utility-Aligned Trajectory Replay for Language Model Post-Training

Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Tianshu Fu, Daren Zha, Jun Xiao

arXiv:2610.00385v1cs.LGcs.AIcs.CLstat.ML

TL;DR

Replay selection can optimize cache-level correctness without aligning with downstream learner utility. FAER establishes an auditable full-trajectory replay protocol and a learner-aware selector. In the matched study, the completed learner-aware row reaches 0.6624, while the fixed selector exceeds uniform and format-feedback on held-out quality.

  • Problem

    Cache-level correctness and downstream learner performance after optimization are distinct measures, and connecting them requires stable task identities, a defined information boundary, and matched update budgets.

    Source: §1 · p. 1

  • Method

    FAER freezes selector-visible fields and selected IDs before evaluator labels are joined, then estimates trajectory utility with a learner-aware selector calibrated on disjoint blocks.

    Source: §1 · pp. 1–2

  • Results

    The completed FAER-UTILITY row reaches 0.6624 quality; FAER-Fixed reaches 0.6329 held-out quality versus 0.5482 for uniform and 0.6037 for format-feedback.

    Source: §4.2 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    Format-feedback has higher frozen-cache correctness than FAER-Fixed, 0.6953 versus 0.3594, while FAER-Fixed has higher held-out quality, 0.6329 versus 0.6037.

    Source: §6 · p. 8

    Evidence covers GSM8K/Qwen2.5-1.5B-Instruct and two cache families; gradient alignment is local, and producer-field semantics remain empirical for the evaluated packages.

    Source: §5 · p. 8

Abstract

from arXiv · show

Replay selectors often rank cached trajectories by format feedback, confidence, freshness, or response length, although cache-level correctness and downstream learner utility are distinct objectives. We formalize this selection-to-learning gap and introduce FAER as an auditable full-trajectory replay framework. Its training-free fixed selector is a protocol baseline; FAER-UTILITY is the learner-aware selector fitted on disjoint calibration blocks. The normalized gradient alignment is reported as a baseline, while a disposable optimizer-aware virtual update supplies a magnitude-aware utility surface. The audit contract freezes observed fields and replay traces before evaluation labels are joined. On GSM8K with Qwen2.5-1.5B-Instruct, the matched learner study reports quality 0.6329 for the fixed selector, compared with 0.5482 for uniform and 0.6037 for format-feedback under 128 updates. Metadata-only cross-fitted calibration reaches $0.6476\!\pm\!0.0139$ over eight seeds (median 0.6481; paired 95% interval $[+0.079,+0.122]$) at 63,276 target-run tokens; its recorded full cost is 189,642 tokens and 3.48 GPU-hours including calibration. The completed FAER-UTILITY row reaches 0.6624 at 62,844 target-run tokens and 4.26 GPU-hours. Format-feedback selects records with correctness 0.6953, compared with 0.3594 for the fixed selector, despite the different downstream ranking. The completed comparison surfaces report the learner-aware ablation, same-seed gap, policy-optimization rows, and strict zero-shot transfer.

1 Introduction

FAER frames replay as selecting complete trajectories for downstream learner utility, not merely cache-level correctness. It pairs a frozen information boundary with learner-aware utility estimation and a training-free feedback-aware baseline.

1 Introduction

  • Complete trajectories retain the generated context consumed by an update and expose feedback that can guide reuse.The paper presents them as a natural replay unit as post-training reuses generated trajectories to allocate limited optimization budgets.Source: §1 · p. 1
  • Replay must be evaluated both by correctness of selected cache records and by held-out learner performance after optimization.Connecting these requires stable task identities, a defined information boundary, matched update budgets, and explicit producer-field lineage.Source: §1 · p. 1
  • FAER freezes selector-visible fields and commits selected IDs before evaluator labels are joined, then estimates downstream utility using disjoint calibration blocks.Producer lineage, learner checkpoints, measured age, and compute telemetry bind the audit and learner comparison.Source: §1 · p. 1
  • FAER formalizes the selection-to-learning gap with rank inversion and regret estimands and reports matched learner, same-seed, cost, transfer, and policy-optimization comparisons.Its contributions also include row-level provenance, immutable traces, post-freeze evaluation, and a cross-fitted learner-aware utility selector with a first-order gradient-alignment connection to calibration-loss reduction.Source: §1 · pp. 1–2

2 Related Work

Related work studies replay signals, data valuation, post-training, agent memory, and reproducibility. FAER connects complete-trajectory selection to matched learner outcomes while recording the information available at each stage.

Replay and data selection

  • Replay research samples transitions by learning signal with importance weights [Schaul et al., 2016; Fu et al., 2021], and later work examines replay ratio, buffer composition, sequence sampling, and actor–learner lag.Offline RL and sequence modeling also make the training-data contract important when distributions are fixed or curated.Source: §2 · p. 2

Learning-aware data valuation

  • Influence functions and gradient-based valuation connect example gradients to changes in a held-out objective [Koh & Liang, 2017].FAER applies this local connection to complete trajectories and measures multi-step utility under a matched learner rather than treating the first-order score as an end-to-end guarantee.Source: §2 · p. 2

Language-model post-training

  • Post-training work makes correctness, confidence, and response length natural selection signals, while replay methods study age, problem-level statistics, verified trajectories, or stale-data updates.FAER links frozen-cache selector diagnostics to matched learner outcomes with shared provenance.Source: §2 · p. 2

Agent memory and feedback

Reasoning evaluation and reporting

  • Reasoning evaluation and model-reporting work emphasize generated traces, evaluation contracts, model and data identity, and explicit sampling and outcome joins.FAER records the observed projection, random draw, trace digest, and post-freeze evaluator join as one auditable contract.Source: §2 · p. 2

Freshness, provenance, and evaluation

  • Artifact lineage, provenance documentation, reproducible-RL reporting, and task documentation motivate recording freshness and evaluation context.Recorded row order supports an index-age stress test, whereas measured freshness requires timestamps and version provenance for every trajectory.Source: §2 · pp. 2–3

3 Method

FAER separates selector-visible cache fields from post-freeze evaluation labels, then compares replay rules through matched learner updates. Its design combines a training-free fixed baseline with cross-fitted learner-aware utility selection and trace-bound audits.

3.1 Problem and Information Boundary

  • The selector uses declared observed fields while the evaluator joins correctness labels only after selected IDs and the trace digest are committed, separating cache correctness from held-out learner quality.The cache audit also checks selector serialization and priorities under taint substitution; derived-field provenance is tracked separately.Source: §3.1 · p. 3

3.2 Priority and Audit Sampling

  • FAER compares seven priority rules, sampling remaining records proportional to priority without replacement, with uniform as the equal-weight control.The feedback-aware rule combines verifier confidence, format feedback, surprise, and uncertainty with fixed offsets; the audit commits an ordered trace digest.Source: §3.2 · pp. 3–4, Figure 1 · p. 4

3.3 Cross-Fitted Learner-Aware Utility Replay

  • The learner-aware selector adds a gradient feature computed at a frozen reference checkpoint and fits parameters on disjoint calibration blocks, while normalized alignment remains a comparison baseline.A disposable virtual update adds optimizer geometry and a quadratic correction; its state is discarded, and the completed learner rows use normalized alignment.Source: §3.3 · pp. 4–5

3.4 From Selection to a Matched Learner

  • A common learner receives ordered replay batches under fixed configuration and evaluation conditions; cache and learner rankings differ, with τSL = 0.33, Spearman ρSL = 0.54, and ISL = 5 of 15 pairs.The supplied run uses seeded cyclic replay of 32 records for 128 updates, with equal token loss weights and greedy decoding.Source: §3.4 · pp. 5–6

3.5 Paired Outcomes and Replication

  • Replication seeds serve as blocks binding producer cache, initialization, trace, checkpoint, and sealed evaluation IDs for paired learner comparisons.Full pairing and randomization details are reported separately.Source: §3.5 · p. 6

4 Experiments

The experiments compare replay policies on matched Qwen2.5-1.5B-Instruct learners, then examine whether cache selection and downstream learner quality agree. The learner-aware selector achieves the highest reported quality on the completed comparison surface, while the policy with highest cached-record correctness differs from the fixed selector’s downstream ranking.

4.1 Experimental Setup

  • The study uses complete Qwen-generated GSM8K trajectories, matched learner updates, stable non-overlapping task splits, and separate target-run and calibration cost accounting.The matched learner contract specifies 128 updates; preregistered runs remain included except for predeclared execution or artifact failures and matched-budget violations.Source: §4.1 · p. 6

4.2 Main Learner Results

  • 0.6624 quality for FAER-UTILITY exceeds the fixed selector’s 0.6329, uniform’s 0.5482, and format-feedback’s 0.6037 on the matched learner study.The metadata-only calibration reaches 0.6476 ± 0.0139 over eight seeds; its full recorded cost is 189,642 tokens and 3.48 GPU-hours, while FAER-UTILITY uses 62,844 target tokens and 4.26 GPU-hours.Source: Table 2 · p. 7, §4.2 · p. 7

4.3 Same-Seed Selection-to-Learning Gap

  • Same-seed blocks bind each replay arm’s consumed trajectories to its resulting checkpoint and sealed evaluation IDs, with rank statistics recomputed within bootstrap resamples.This design tests whether disagreement between cached correctness and learner quality persists when both arise from the same replay intervention.Source: §4.3 · pp. 7–8, Table 3 · p. 8

4.4 Learner-Aware Utility Ablation

  • The learner-aware ablation compares metadata-only, alignment-only, full, and negative-control selectors, including shuffled alignment and a norm-matched random-gradient control.The controls distinguish trajectory-specific alignment information from altered priority dispersion or alignment scale.Source: Table 4 · p. 8, §4.4 · p. 8

4.5 Selection Composition and Priority Analysis

  • Format-feedback selects records with 0.6953 correctness versus 0.3594 for FAER-Fixed, despite their different downstream learner ranking.The complete selection table, scatter, and audit packages are reported in Appendix D.12 and Appendix D.Source: §4.5 · p. 8

5 Limitations

The reported evidence is bounded to GSM8K with Qwen2.5-1.5B-Instruct, two cache families, and a 128-update horizon; gradient alignment is local, and producer-field semantics remain empirical.

5 Limitations

  • The evaluated scope is GSM8K/Qwen2.5-1.5B-Instruct, two cache families, and a 128-update horizon.The evidence includes matched three-seed runs and five- and eight-seed aggregates.Source: §5 · p. 8
  • Gradient alignment is a local signal, not an end-to-end guarantee of multi-step learner utility.FAER-UTILITY is distinguished from the training-free FAER-Fixed protocol baseline.Source: §5 · p. 8
  • Producer-field semantics remain an empirical property of the evaluated packages, despite the audit contract recording information flow and artifact identity.Source: §5 · p. 8

6 Conclusion

The conclusion frames replay selection as learner-utility estimation under an explicit audit boundary. Its reported comparisons include distinct cache-correctness and learner-quality rankings, a selection-to-learning gap, and completed policy and transfer surfaces.

6 Conclusion

  • FAER-Fixed reaches 0.6329 held-out quality versus 0.5482 for uniform and 0.6037 for format-feedback, while FAER-UTILITY reaches 0.6624.The policy-optimization row reaches 0.6575 final quality.Source: §6 · p. 8
  • Format-feedback’s frozen-cache correctness is 0.6953 versus 0.3594 for FAER-Fixed, while the six-arm descriptive gap has Kendall τSL = 0.33, Spearman ρSL = 0.54, and five pairwise inversions.Metadata-only calibration reaches 0.6476 ± 0.0139 with full cost reported separately.Source: §6 · pp. 8–9
  • Same-seed, zero-shot transfer, policy-optimization, and optimizer-aware diagnostic surfaces remain tied to their declared learner, trace, and evaluation boundaries.Source: §6 · p. 9

AI-use

The authors report using generative AI for language polishing and reference summaries, but not for experimental results, synthetic data, mathematical claims, proofs, or research decisions.

AI-use

  • Generative AI tools were used for language polishing and summarizing references.Source: AI-use · p. 9
  • The authors state that generative AI tools were not used to generate experimental results, create synthetic datasets, formulate mathematical claims, provide proofs, or decide research conclusions.Source: AI-use · p. 9
  • The authors say they designed and verified the methodology, experimental setup, analysis, and interpretation, and take responsibility for the final content and artifacts.Source: AI-use · p. 9

ETHICS STATEMENT

The work uses model-generated trajectories and public mathematical reasoning benchmarks, without human subjects, personal data, user studies, or deployment affecting individuals.

ETHICS STATEMENT

  • The audit separates selector-visible metadata from post-freeze evaluation labels and uses fixed task splits, bounded compute, and reproducible offline evaluation.The supplementary release excludes model weights and source artifacts whose licenses restrict redistribution.Source: ETHICS STATEMENT · p. 9

REPRODUCIBILITY STATEMENT

The supplementary materials provide code and run records intended to reconstruct the reported tables and figures.

REPRODUCIBILITY STATEMENT

  • Code covers trajectory selection, audit checks, sampling, learner evaluation, statistical summaries, and figure generation.Manifests, selected-ID traces, checkpoint identifiers, result mappings, and digest-bound artifact records link reported visuals to their packages.Source: REPRODUCIBILITY STATEMENT · p. 9

A Task, Sampling, and Method Details

The methods define replay over complete trajectory records using declared observed fields, with selection traces frozen before correctness labels are joined. They specify policy priorities, sampling and audit procedures, and matched learner comparisons.

A.1 Selection unit and field contract

  • The audit samples 32 unique records without replacement from declared priorities, serializes ordered IDs and a trace digest, then joins labels after freezing the trace.The complete trajectory is the sampling unit, identified by a stable task/cache ID and accompanied by generated text and producer-side fields.Source: §A.2 · p. 12, §A.1 · p. 12

A.3 Policy interpretation

  • Uniform, format-feedback, reliability, surprise, uncertainty, feedback-aware, and token-efficiency use distinct cached-field priorities, with feedback-aware multiplying four factors.The verifier-dependent factor g(ρ) increases through ρ = 0.75 and decreases afterward; the analytic maximum is independent of dataset and seed.Source: §A.3 · p. 12, §A.5 · p. 13

A.4 Sampling variation and finite-population estimands

  • Eight-seed averages and sample SD summarize selected-record rates conditional on fixed rows and priorities; paired comparisons require joint inclusion probabilities or validated coupling.The eight fixed-cache seeds assess draw variation, while independent task prompts, generations, producers, and training runs are separate replication units.Source: §A.4 · p. 13

A.6 Audit metrics and interpretation

  • The cache audit keeps selected-record correctness, format status, verifier confidence, cached token length, and Jaccard overlap separate, with distinct denominators and directions.Family A has 64 calibration/train and 128 held-out/test rows; family B has 256 calibration/train and 1,319 held-out/test rows, with disjoint train/test IDs within each family.Source: §A.6 · p. 14, §B.1 · p. 14

B.2 Observed fields and provenance stages

  • The fixed-cache audit reports complete declared-field coverage, zero online oracle-key hits, and invariant serialization and priority across 1,344/1,344 target-field mutations.This establishes selector-side exclusion; the semantic lineage of upstream fields such as verifier confidence and format status is reported separately.Source: §B.2 · p. 14, Table 5 · p. 14

B.4 Freshness metadata

  • Measured per-record age is absent from all primary cache rows, so the existing age-stress grid uses row position as a declared proxy.The independent-producer package has populated source fields for 1,767/1,767 rows; lineage records specify producer inputs, versions, access policy, and row digests.Source: §B.4 · p. 15, §B.5 · p. 15

B.6 Evaluator validation surface

  • Evaluator validation reports 95/96 agreement (98.96%) with independent adjudication, with one changed label retained.The stratified sample includes exact matches, equivalent numeric forms, malformed answers, multiple candidate answers, and no-answer cases.Source: §B.6 · p. 15

C.1 Matched learner configuration

  • The learner configuration compares additional reliability and uncertainty-PER controls with feedback-aware under a shared initial checkpoint, task IDs, and update budget.Feedback-aware has higher mean quality than both controls; direct paired intervals against feedback-aware require the paired seed records.Source: §C.1 · p. 16, Table 8 · p. 16

C.2 Experiment identities and comparison axes

The experiment packages bind cache identity, replay traces, and learner evaluation to shared comparison rules. Their diagnostics distinguish selected-record correctness from held-out learner quality and report cache, age, and seed sensitivities separately.

C.2 Experiment identities and comparison axes

  • The experiment packages separate selected-set outcomes from learner endpoints, and their run records bind configurations, traces, telemetry, checkpoints, and evaluator records.The comparison of prior replay methods organizes them by selection unit, age handling, update treatment, and verifier timing.Source: §C.2 · p. 16, §C.3 · p. 17, §C.4 · p. 17

C.5 Typed record interfaces

  • The fixed-cache release passes identity, schema, taint, trace, oracle-join, and release gates, while learner and independent-producer rows use the same gate names.The typed records connect cache manifests, trajectory rows, frozen selection traces, and post-freeze evaluator joins.Source: §C.5 · p. 18, §C.5.1 · p. 18, §C.5.2 · p. 19

D.1 Primary train and test slices

  • Seed-13 test results show format-feedback, reliability, and token-efficiency each select 22/32 correct records, while rankings shift between train and test caches.Across two cache families, format-feedback exceeds token-efficiency by 0.0625 in family A, but token-efficiency exceeds format-feedback by 0.0742 in family B.Source: Table 14 · p. 19, §D.1 · p. 19, §D.3 · p. 20

D.2 Telemetry and correctness deltas

  • Format-feedback selects test-slice traces with 1.0000 format-ok and 0.8500 mean verifier confidence, while uniform records 0.3438 and 0.4234, respectively.These are producer metadata and cached response-length telemetry; inference-time cost is reported on the matched learner surface.Source: §D.2 · p. 20

D.5 Paired without-replacement inference

  • The feedback-aware-minus-format-feedback paired interval is negative under the stated cache and sampling design, and every reported 95% interval in the paired table excludes zero.The comparison couples random streams, records ordered draws and inclusion probabilities, and reports mean cached-token length as a separate trade-off.Source: §D.5 · p. 20, Table 15 · p. 20

D.7 Matched learner study

  • The matched learner design fixes model, optimizer, precision, update schedule, budgets, and task splits across arms, using independent seeds 13, 17, and 23.It retains checkpoints, progress traces, generated-token counts, synchronized latency, and evaluator digests for reporting quality and cost at matched tokens and compute.Source: §D.7 · p. 21, Table 17 · p. 21, Table 18 · p. 21

D.8 Runtime preflight

  • The runtime preflight reports every arm at 0/16 held-out correctness, while the reconstruction report covers cache rows, ordered traces, and numeric cells.Evaluator validation is adjudicated independently, and digest mismatches are localized by pipeline stage.Source: Table 19 · p. 22, §D.9 · p. 22, Table 20 · p. 22

D.10 Freshness and Age Sensitivity

  • On the frozen test cache, freshness-policy correctness ranges from 0.1562 to 0.3047, compared with 0.6562 to 0.7031 for format-feedback.All rows lack age fields, so this four-setting stress grid uses the declared index-age fallback; freshness selected-set Jaccard ranges from 0.1377 to 0.6274.Source: §D.10 · p. 22, Table 21 · p. 22

D.13 Extended quality, loss, and cost results

The extended comparisons report package-specific learner quality, test selector components and sensitivity to training horizon, freshness decay, and field changes, and extend replay comparisons to policy optimization and transfer. They also describe calibration, compute, and reproducibility conditions for interpreting those results.

D.13 Extended quality, loss, and cost results

  • 0.6497 task quality and 62,948 mean generated tokens are reported for the end-to-end package; the age/family package reaches 0.6418 with complete timestamp integrity.The paired summaries also report +0.0847, +0.0768, and +0.0831 for their respective comparisons; all three packages reach 128/128 updates.Source: §D.13 · p. 25, Table 28 · p. 26

D.14 Component and calibration ablations

  • Removing uncertainty lowers quality from 0.6329 to 0.6164, while removing surprise lowers it to 0.6088 under the reported mechanism experiment.The monotone and calibrated verifier alternatives reach 0.6205 and 0.6261; the reported treatment intervals distinguish their evidence from the full product.Source: §D.14 · p. 26, Table 29 · p. 26

D.15 Budget and age-scale sensitivity

  • Task quality rises from 0.5787 at 32 updates to 0.6089 at 64 and 0.6329 at 128; measured-age quality peaks at 0.6418 for λ = 0.04.The measured-age endpoints are 0.6214 at λ = 0 and 0.6306 at 2λ, so stronger decay does not monotonically improve the endpoint; decay also changes replay mass.Source: §D.15 · pp. 26–27, Table 30 · p. 26

D.16 Counterfactual field controls

  • Verifier permutation changes trace Jaccard to 0.3614 and task quality to 0.5946, while format-status permutation yields 0.4278 and 0.6072.Age permutation leaves Jaccard and quality at 1.0000 and 0.6329, as expected because the fixed product has no age input; freshness is tested separately with measured-age decay.Source: §D.16 · p. 27, Table 31 · p. 27

D.17 Policy-Optimization Replay

  • The policy-optimization comparison uses matched update, generation, and optimization settings, with FAER-Utility fitted on disjoint calibration groups and frozen before target optimization.The utility selector leads the completed surface in final quality and reward while using the fewest generated tokens.Source: §D.17 · p. 27, Table 32 · p. 27

D.18 Replication, Transfer, and Stronger Replay Baselines; D.19 Strict Zero-Shot Selector Transfer

  • The replication and transfer surfaces compare independent learner runs, an existing replay baseline, and cross-task or model-family transfer; strict transfer freezes GSM8K/Qwen2.5-1.5B-Instruct selector parameters before target evaluation.Target task and model outcomes do not contribute calibration, temperature updates, gradient normalization choices, or stopping decisions in the strict test.Source: §D.18 · pp. 27–28, §D.19 · p. 28, Table 34 · p. 28

D.20 Verifier shape and selection composition; D.21 Cross-fitted utility and diversity controls

  • Verifier calibration and replay utility are distinct: Brier score evaluates probability calibration, whereas matched-budget learner quality evaluates replay utility.The cross-fitted utility procedure fits weights, temperature, and checkpoint choices on remaining calibration blocks, then freezes them for the target block; its controls compare utility with correctness calibration and diversity.Source: §D.20 · p. 29, §D.21 · pp. 29–30, Table 38 · p. 29

D.22 Utility-estimator fidelity and optimizer awareness; D.23 Compute-matched and replay-regime sensitivity; E.2 Earlier pilot records; E.4 Serialization and release contract

  • The utility-fidelity test compares predicted scores with one-step calibration-loss improvements after a disposable virtual update, separating estimator fidelity from long-horizon learner effects.Compute comparisons distinguish target-budget matching from total-compute matching, which includes calibration; earlier pilot records use distinct checkpoints and aggregation units from the mature study. The release contract binds runs and replay traces to manifests and digests, with evaluation labels joined after trace commitment.Source: §D.22 · p. 30, §D.23 · p. 30, Table 42 · p. 31

E.5 Additional Analysis

The additional analysis reports observed-metadata selector audits and provenance checks, then describes how split integrity, stage-level records, and diagnostic schemas support reproducibility and localized result updates.

E.5 Additional Analysis

  • The observed-only product selects 14/32 correct test records, versus 22/32 for format-feedback, reliability, and token-efficiency; the matched learner package adds held-out task outcomes and compute-normalized cost.In eight fixed-cache draws, feedback-aware remains below format-feedback and token-efficiency in mean correctness.Source: §E.5 · p. 33

E.6 Observed-Metadata Provenance Audit

  • Required selector fields have 100% valid coverage, with zero online oracle-key hits and unchanged priorities in all 1,344/1,344 target-field taint trials.The private target_answer column is excluded from online serialization; measured-age coverage is 0/192 in the current cache and populated in the independent producer package.Source: §E.6 · p. 33

E.7 Reproducibility Checklist

  • The reproducibility checklist keeps train and test rows separate, rejects duplicate IDs, validates manifests, and joins the oracle only after each policy trace is frozen.The evidence records zero online oracle-key hits and 1,344/1,344 taint-invariant priorities, while learner outcomes remain on a dedicated matched-budget surface.Source: §E.7 · pp. 33–34

E.8 Reporting stages; E.8.1 Stage-level records

  • The reporting matrix maps each computational stage to its unit, artifact, and primary metric, linking table cells to the raw records that generated them.For the fixed cache, it records 32 IDs per policy across eight seeds and joins correctness using the same task/cache IDs after trace commitment.Source: §E.8 · p. 34, §E.8.1 · p. 34, Table 46 · p. 34

E.8.2 Selection and learner diagnostics

  • The diagnostic taxonomy separates field, trace, evaluator, and update events, each with an identity, expected contract, observed value, and resolution field.This lets discrepancies be localized before they change a learner comparison; the completed selector audit, trace ledger, and correctness summary populate distinct diagnostic rows.Source: §E.8.2 · pp. 34–35, Table 47 · p. 34

E.8.3 Release update rule

  • Release updates retain the original trace and evaluator boundaries: selector-visible fields are fixed before sampling, correctness joins after the trace digest, and learner outcomes stay on the matched-budget surface.A changed result updates the arm record and associated figure while fixed-cache summaries remain bound to their original cache bytes.Source: §E.7 · p. 33, §E.8.3 · p. 35
Loading 2610.00385v1…