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 judges every claim with real probes, and verified claims are measured for contextual advantage (does giving the claim back to the actor change play?). Everything was produced between 2026-08-30 and 2026-09-03 by src/knowledge_learning/loop/ of ziyan-wang98/knowledge-learning-e2k (docs/E2K_LOOP_V0.2_PLAN.md). 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 (loop/iter_0/knowledge/ at an 8,192-token output budget, the 6 cases that produced nothing re-run at 16K in knowledge_16k/)
verification (120 claims, all at the intervention tier)	81 verified (68 %) / 32 partially verified (27 %) / 7 refuted (6 %); ARC3 (58 claims): 34 / 20 / 4; DiG (62): 47 / 12 / 3
dominant ARC failure	scope over-generalisation, caught by active intervention probes, never by trace reading
contextual advantage	ARC only (the paired design needs a deterministic replay; DiG's server assigns seeds); 34 verified claims × 8 seed-matched pairs × 20 decisions; j_levels is at the floor (0 on every episode), 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 with validation outcome, blind ids only), text_trace.json/.md (the allow-listed trace the extractor saw); summary.json per pool. knowledge_16k/ is the 16K-budget re-run of the 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 / abstained, with checks, counterexamples and confidence) and the probe transcripts. 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, with the dataset root unreadable to the agent (blinding is mechanical).
• loop/iter_0/teachability/<case>/ — paired episodes (baseline vs knowledge arm on one sampling seed) per claim, graded outcomes per episode, per-shard summaries; wave 2 adds partially-verified and refuted claims as comparison and negative-control groups.
• 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 / status.html, rebuilt deterministically by loop.report.
• loop/iter_1/knowledge/ — a second self-extraction pass over the 40 cases (2026-09-02/03) with the updated text-trace budget; never verified.
• retention/smoke/ — the retention-diagnostic smoke run (base vs distilled Qwen3-1.7B on ARC3_002).
• opsd/ — logs of the OPSD smoke on Qwen3-1.7B (teacher = student); the 23 GB Megatron smoke checkpoint and its HF export are not included.
• 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 file. Released under CC BY 4.0 (observations of third-party environments remain under their own terms). Ziyan Wang, King's College London, 2026; compute: Isambard-AI (BriCS, University of Bristol).