ziyan98/E2K-loop-v0.2

E2K loop v0.2 — one iteration of the self-extraction / verification / teachability loop

What this is. Iteration 0 (plus a second extraction pass) of the v0.2 loop that followed E2K-40: the same Qwen3.8-27B actor that played the 40 trajectories extracts one atomic knowledge claim per sample from an allow-listed text trace, an independent verifier (headless Claude Code) judges every claim, 119 of the 120 with real probes, and verified claims are measured for contextual advantage (does giving the claim back to the actor change play?). The loop data (loop/, retention/, the OPSD smoke logs, the Slurm logs) was produced between 2026-08-30 and 2026-09-03 by src/knowledge_learning/loop/, scripts/run_claim_verification.py and the verifier skills under skills/ of ziyan-wang98/knowledge-learning-e2k (docs/E2K_LOOP_V0.2_PLAN.md; the repository is private at the time of this release, available on request). report/ is the progress report of 2026-09-11/12 and opsd/qwen3_1_7b_student_smoke/ an unrelated Qwen3-1.7B inference smoke of 2026-08-25. Nothing here was written into the E2K-40 dataset root; episodes produced no recordings.

Headline numbers (iteration 0)

stage	result
self-extraction	4 samples per case; 120 valid primary claims over all 40 cases from 184 samples (98 of 160 at the 8,192-token output budget — all 62 failures hit the output limit — plus 22 of 24 for the 6 empty cases re-run at 16K, knowledge_16k/)
verification (120 claims)	81 verified (68 %) / 32 partially verified (27 %) / 7 refuted (6 %); ARC3 (58 claims): 34 / 20 / 4; DiG (62): 47 / 12 / 3. 119 claims reached the intervention tier; one DiG refutation (DIG_010/001-575c5032) rests on the trace alone because its probe command was denied by the harness
verifier	headless Claude Code; 63 claims were judged by claude-opus-5, 57 by claude-fable-5 (each ledger's verifier.model says which), and the two did not judge alike — opus refuted about 2.5× as often — so the pool is a mixed instrument
dominant ARC failure	scope over-generalisation: 17 of the 20 partially-verified ARC claims carry an intervention-probe counterexample (the other 3 were downgraded because fresh-seed reproduction hit a probe bug); caught by active probes, never by trace reading alone
contextual advantage	ARC only (the paired design needs a deterministic replay; DiG's server assigns seeds); wave 1 = the 34 verified ARC claims × 8 seed-matched pairs × 20 decisions. j_levels is at the floor (0 on all 544 wave-1 episodes); the secondary outcomes j_effective_action_rate (11 advantage / 2 harm / 11 identical / 10 below margin) and j_valid_action_rate discriminate
selection rules	all-knowledge 113 rows, correctness-gated 81, metric-routed 11, human-selected 0 (no human review was recorded)

Every rate in the reports carries its numerator and denominator; an effect below 2·sd/√n is reported as below detectable margin, never as "no effect".

Contents

path	files	size
loop/iter_0	1792	14.3 MiB
loop/iter_1	121	15.8 MiB
opsd	9	1.6 MiB
report	14	575.8 KiB
retention	52	117.0 KiB
slurm-logs	2	339.7 KiB
total	1990	32.7 MiB

• loop/iter_0/knowledge/<case>/ — samples.jsonl (raw extractor samples: the model was prompted with the blind id only, but each row also records the real case_id, the model's reasoning, the raw response and the parse/validation outcome), text_trace.md (the allow-listed trace the extractor saw) and text_trace.json (its character/token-budget record); summary.json per pool. knowledge_16k/ is the 16K-budget re-run of the 6 empty cases.
• loop/iter_0/verification/<case>/<claim>/ — the verifier's ledger.json (five-way verdict schema loop/verify_schema.py: verified / partially_verified / refuted / unknown / not_checkable, the last two being abstentions; with evidence, counterexamples, limitations, confidence and tier_reached), the headless-agent prompt.txt / transcript.txt (plus attempt-NN.* for failed attempts) and, for DiG claims, the fresh-session probe transcripts probes/sessionN.jsonl (ARC probe output is embedded in the ledger's evidence entries; the ARC probes/ directories are empty). ARC claims were probed by the arc-rule-verifier skill (offline, deterministic replay of the cached game at engine seed 0, then interventions and fresh seeds); DiG claims by the dig-rule-verifier skill (budgeted fresh bench sessions). The driver was scripts/run_claim_verification.py over headless Claude Code, whose permission system limited the agent to the loop output tree and the two probe commands (the recording reached it only through the probe, which strips actor reasoning).
• loop/iter_0/teachability/<case>/ — paired episodes (baseline vs knowledge arm on one sampling seed) per claim, graded outcomes per episode, per-shard summaries under summaries/ (the tree-wide summary.json is a stale stub of the failed first pilot: 7 claims, 30 decisions, no episodes — ignore it). Wave 2 (2026-09-01) adds the 20 partially-verified and 4 refuted ARC claims as comparison and negative-control groups at a 40-decision budget (wave 1: 20), so the two waves are not directly comparable; wave 2 is incomplete (11 of the 20 partially-verified claims stopped at 2–7 pairs) and is not reflected in reports/. Across all 860 episodes a single wave-2 knowledge-arm episode (ARC3_011/003, pair 0) cleared one level.
• loop/iter_0/selection/ — <rule>.knowledge_set.jsonl + <rule>.manifest.json (every excluded claim is listed with the gate that excluded it; refuted is excluded by every rule).
• loop/iter_0/opsd/<rule>/ — student/teacher context pairs for on-policy self-distillation (opsd_dataset.jsonl, opsd_slime.jsonl, manifest with disjoint seed bands). This training path was later dropped (see the RL dataset); the data is kept as built.
• loop/iter_0/reports/ — report.md / report.json, rebuilt deterministically by loop.report (built 2026-09-01 07:20, before teachability wave 2), plus status.html, a hand-written progress snapshot of 2026-08-31 23:40 UTC (teachability then at 70/544 episodes) that loop.report does not regenerate.
• loop/iter_1/knowledge/ — a second self-extraction pass over the 40 cases (2026-09-02/03) with the v2 extractor prompt (self_extractor_v2: strategy as well as mechanism claims, up to three claims per response, a 20,480-token output budget and a 45,056-token trace budget): 99 valid claims of 160 samples, 3 ARC cases empty; never verified.
• retention/smoke/ — the retention-diagnostic smoke run (base vs distilled Qwen3-1.7B on ARC3_002).
• opsd/smoke/ — logs and latest_checkpointed_iteration.txt of the OPSD smoke on Qwen3-1.7B (teacher = student, 2026-09-01); opsd/smoke_hf_metadata/ — config, README and provenance of the HF export (the 23 GB Megatron checkpoint and the 3.8 GB HF weights are not included; the README there is Qwen's own Apache-2.0 model card, reproduced unchanged); opsd/qwen3_1_7b_student_smoke/ — an unrelated 2026-08-25 inference smoke of the Qwen3-1.7B checkpoint from the Crafter trials.
• report/ — the HTML progress report of 2026-09-11/12 with its figures and board renders.
• slurm-logs/ — teachability, extraction, retention and OPSD job logs.

Reading the ledgers

A claim is a hypothesis until independently checked. The verdicts above are the verifier's; the review UI of the code repository (loop.review_ui) was built to collect human agreement, but no human review was recorded in this iteration, so the verifier's accept-precision is unmeasured.

Integrity and licence

MANIFEST.sha256 / MANIFEST.json hash every data file (1,990 files; the README and the manifests themselves are not listed). Released under CC BY 4.0 (observations of third-party environments remain under their own terms), with one exception: the three files under opsd/smoke_hf_metadata/ that are the upstream Qwen/Qwen3-1.7B model card and configuration (README.md, config.json, generation_config.json; Qwen Team, Apache-2.0) are reproduced unchanged and not relicensed here. Ziyan Wang, King's College London, 2026; compute: Isambard-AI (BriCS, University of Bristol).