2026-10-10T13:01:54Z
https://api.figshare.com/v2/oai
oai:figshare.com:article/33970177
2026-09-23T04:16:35Z
category_28864
item_type_3
month_year_09_2026
AIM Auditability and Rule-Level RL Fusion: Full Training Traces and Checkpoints
hung Liu (25093672)
Artificial intelligence not elsewhere classified
Reinforcement Learning Learning
Auditability (AI)
Provenance Challenge
Rule Induction
policy composition
<h3 dir="ltr">Track 2 — archived dataset (<code>runs/</code>)</h3><table><tr><td><p dir="ltr">Directory</p></td><td><p dir="ltr">Size</p></td><td><p dir="ltr">Role</p></td></tr><tr><td><p dir="ltr"><code>runs/shared_multienv_gpi/</code></p></td><td><p dir="ltr">3.1 GB</p></td><td><p dir="ltr"><b>Primary.</b> The 8-seed, 2-environment experiment behind every headline number</p></td></tr><tr><td><p dir="ltr"><code>runs/grammar_confidence_070/</code></p></td><td><p dir="ltr">584 MB</p></td><td><p dir="ltr">Confidence-threshold rule-bank sweep</p></td></tr><tr><td><p dir="ltr"><code>runs/grammar_audit_pilot/</code></p></td><td><p dir="ltr">340 MB</p></td><td><p dir="ltr">Pilot audit run</p></td></tr><tr><td><p dir="ltr"><code>runs/repeatability_full/</code></p></td><td><p dir="ltr">164 MB</p></td><td><p dir="ltr">Deterministic-repeatability verification</p></td></tr><tr><td><p dir="ltr"><code>runs/private_target_replicates/</code></p></td><td><p dir="ltr">153 MB</p></td><td><p dir="ltr">Multi-replicate private-target-game runs</p></td></tr><tr><td><p dir="ltr"><code>runs/grammar_merge_070_final/</code>, <code>_070/</code>, <code>_070_v2/</code></p></td><td><p dir="ltr">364 MB</p></td><td><p dir="ltr">Seed-grammar merge at confidence 0.70 (final, base, superseded variant retained for audit trail)</p></td></tr><tr><td><p dir="ltr"><code>runs/grammar_merge_090_final/</code>, <code>_090/</code></p></td><td><p dir="ltr">165 MB</p></td><td><p dir="ltr">Seed-grammar merge at confidence 0.90</p></td></tr><tr><td><p dir="ltr"><code>runs/private_target_extended/</code>, <code>_batch8/</code>, <code>private_target/</code></p></td><td><p dir="ltr">44 MB</p></td><td><p dir="ltr">Private-target-game variants</p></td></tr><tr><td><p dir="ltr"><code>runs/grammar_intersections/</code></p></td><td><p dir="ltr">36 KB</p></td><td><p dir="ltr">Pairwise rule intersections</p></td></tr><tr><td><p dir="ltr"><code>runs/seed-0_20260922T013053Z/</code>, <code>_012930Z/</code></p></td><td><p dir="ltr">15 MB</p></td><td><p dir="ltr">Single-seed AIM baseline runs cited in <code>paper/AIM_Auditability_Research_Log.md</code></p></td></tr><tr><td><p dir="ltr">loose files</p></td><td><p dir="ltr">~1.6 MB</p></td><td><p dir="ltr"><code>replication_summary.json</code>, <code>replication_summary_interim.json</code>, <code>grammar_mdl_seed11_070.json</code>, <code>objective_diagnostic.json</code>, <code>seed-0_full_baseline_console.log</code></p></td></tr></table><p dir="ltr">Total: 584 files, 4.7 GB uncompressed.</p><p dir="ltr">Per-directory role notes: <code>shared_multienv_gpi/</code> holds per-step trajectory traces, training-update ledgers, actor and training checkpoints, shared calibration, induced rule banks, and FQE critic checkpoints. The legacy directories hold the single-seed AIM baseline, the grammar-merge and calibration sweeps, the private-target-game runs, and the repeatability verification.</p><p dir="ltr"><b>Ten files exceed GitHub's 100 MB per-file limit</b> (largest 391.5 MB); this is the reason for the two-track split. The largest offenders are <code>trajectory.jsonl</code> traces and <code>lossless_trace_grammar.json</code>, which compress by roughly 6.4x.</p>
2026-09-23T04:16:35Z
Dataset
Dataset
10.6084/m9.figshare.33970177.v1
https://figshare.com/articles/dataset/AIM_Auditability_and_Rule-Level_RL_Fusion_Full_Training_Traces_and_Checkpoints/33970177
CC BY 4.0