← older: fast__qwen3.6-35b-a3b__20260710-213727all runs · task boardnewer: smoke__qwen3.5-122b-a10b__20260712-135546

medic__qwen3.6-35b-a3b__20260710-232538

medic__qwen3.6-35b-a3b__20260710-232538 — medic recovery over the fast set (P0)

RULE-RELAXED recoverability axis — NOT a baseline/leaderboard number. The medic is told the verifier failed and shown the exact failing-check output. Baseline scoring stays the standard number; this is a separate column.

Phase 1 (baseline): runs/fast__qwen3.6-35b-a3b__20260710-213727 (MinimalPi, MEDIC_CAPTURE=on, fast set default caps). 5/9 pass, 4 fail. Phase 2 (this run, MedicPi, signal=full log_access=dump workspace=restore, per-session cap default=900): all 4 failed workspaces restored + failing pytest injected, then re-verified.

Recoverability verdict: 1/4 recovered (25%)

task baseline medic verdict
cancel-async-tasks fail (cut@3m ph1) pass RECOVERED (6 passed, 14s)
filter-js-from-html fail fail still-fails (stopped on its own, 2m09s)
query-optimize fail fail still-fails (stopped on its own, 4m41s)
regex-log fail (0/0 tok ph1 — wire-format-leak empty session) fail inconclusive — cut@15m by the 900s medic cap (exit 124); churned to 142k output tokens (looped)

Suite line: baseline 5/9 · +1 nudge 6/9 · recoverability 1/4 = 25% (of the 3 that got a full attempt: 1 recovered, 2 clean still-fails; regex-log cut).

cancel-async-tasks recovery is a genuine verifier pass (reward=1.0, 6 passed).

Notable: cross-run variance on cancel-async-tasks

The earlier smoke-run medic on the SAME task (runs/medic__…-190552) was a still-fails near-miss (fixed the target test, regressed another); here it fully recovered. Same P0 config — the difference is the base agent's phase-1 broken draft (different each run) + temp-0.6 sampling. This is the K=1 caveat made concrete: single-trial recoverability is noisy; a K>1 pass would average it.

How this compares to fast2

fast2 recovered 3/5 (60%); fast recovers 1/4 (25%). Recoverability is strongly task-dependent — fast2's failures (git-multibranch, mailman) were one-round repairs, while the fast set's (filter-js-from-html, query-optimize) resisted, and regex-log's phase-1 session was EMPTY (wire-format leak), so the medic had to solve from near-scratch rather than repair a draft — a harder regime and it looped to the cap.

Caveats / honesty

💬 4 analyst comments inline below (from runs/medic__qwen3.6-35b-a3b__20260710-232538/COMMENTS.md).

Run details

modelllama-local/qwen3.6-35b-a3bagentharnesses.medic_pi:MedicPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials4 of 4 — 1 pass · 3 failmean reward0.25tokens (job total)3,489,126 in / 206,418 outstarted / finished2026-07-10T23:25 / 2026-07-10T23:56wall clock30m31s

Tasks

cancel-async-tasks — 1/1 passed

💬 analyst comment

Medic RECOVERED (baseline fail → pass, 6 passed in 14s). NB the earlier smoke-run medic on this exact task (runs/medic__…-190552) was a still-fails near-miss — same config, different phase-1 broken draft + temp-0.6 = cross-run variance (K=1 caveat).

#resulttotalagentin/out tokflags
1PASS3m38s2m51s1019015/18691
🔍 view

filter-js-from-html — 0/1 passed

💬 analyst comment

Clean still-fails: the medic stopped on its own (2m09s) without fixing the check.

#resulttotalagentin/out tokflags
1FAIL5m41s2m09s411848/18838
long reasoning (13,733 chars) ×2
🔍 view

query-optimize — 0/1 passed

💬 analyst comment

Clean still-fails: stopped on its own (4m41s). A real not-recovered datapoint.

#resulttotalagentin/out tokflags
1FAIL5m26s4m41s1810040/26367
a bash command timed out ×2
🔍 view

regex-log — 0/1 passed

💬 analyst comment

Inconclusive: the phase-1 base session was EMPTY (0/0 tokens — the known wire-format-leak), so the restored workspace was ~unchanged and the medic had to solve from near-scratch, not repair a draft. It looped to 142k output tokens and was CUT at the 900s cap (exit 124). Worst case for the medic.

#resulttotalagentin/out tokflags
1FAIL15m44s15m00s248223/142522
fast-timeout cut at 15mlong reasoning (16,300 chars)
🔍 view