From bbe59be7e371ec54dab30e4b9587b6fe48e6b7fa Mon Sep 17 00:00:00 2001 From: wassname <1103714+wassname@users.noreply.github.com> Date: Tue, 30 Jun 2026 17:36:06 +0800 Subject: [PATCH] Record MFQ2 path failure --- docs/spec/20260630_authority_steer_pipeline.md | 1 + 1 file changed, 1 insertion(+) diff --git a/docs/spec/20260630_authority_steer_pipeline.md b/docs/spec/20260630_authority_steer_pipeline.md index 89ab02f..11ca2a4 100644 --- a/docs/spec/20260630_authority_steer_pipeline.md +++ b/docs/spec/20260630_authority_steer_pipeline.md @@ -132,3 +132,4 @@ Out: - 2026-06-30: Killed steering-lite pueue task 399 after calibration reached `C=39.758`, `kl_p95=0.547` but then spent 34 minutes in the final calibration snapshot with no eval outputs. Killed task 400 because the runner did not forward CLI overrides, so it would have ignored `--fixed-C`. - 2026-06-30: Pushed steering-lite `eafae69` so `scripts/run_authority_only_score60_showcase.sh` forwards `"$@"`. Queued pueue task 401 with `--fixed-C 39.758 --out outputs/20260630_authority_only_qwen3_14b_score60_sspace_mfv_mfq2_n8_fixedC39758`, then killed it because the pre-unbuffered runner gave no live stdout/files after five minutes. - 2026-06-30: Pushed steering-lite `899d932` to set `PYTHONUNBUFFERED=1` in the authority runner. Queued pueue task 402 with output `outputs/20260630_authority_only_qwen3_14b_score60_sspace_mfv_mfq2_n8_fixedC39758_unbuf`; log confirms `fixed C=+39.7580 (iso-KL skipped)`. +- 2026-06-30: Killed pueue task 402 after MFQ-2 gave a mixed result: small `c` had the right local sign (`authority` C base `30.80`, `+0.5` `33.17`, `-0.5` `29.72`) but `+1` doubled back to `27.23` while pmass stayed `1.000`. Queued pueue task 403 as focused MFV-only check at `c=0.5,1`: `outputs/20260630_authority_only_qwen3_14b_score60_sspace_mfv_c05_c1_fixedC39758`.