Experiment ·
HEADPASS-X2b
Does appending the retrieved article's neighbouring chunks plus up to 2 dense-vector candidates (≤ +600 prompt tokens) raise the 4B's score on the 180 pack-matched self-tests?
From the record
- What was tested
- neighbour chunks + budgeted dense candidates appended after the keep step
- Result
- Opus +0.066 (+0.043…+0.091); Haiku +0.056; fact chunk present in context 35 → 102 of 180
- Verdict
- adopted: shipped or recommended by its frozen rule
- Set
- 180 pack self-tests (dev)
- Judge
- Claude Opus (L2) and Claude Haiku (L1)
- Pre-registered
- yes, sha256-frozen
Linked facts · 2
- The one retrieval change that converted: append the retrieved article's neighbouring chunks plus a few vector-search candidates, under a 600-token budget (HEADPASS-X2b).The one retrieval change that converted: append the retrieved article's neighbouring chunks plus a few vector-search candidates, under a 600-token budget (HEADPASS-X2b).Opus +0.066 (+0.043…+0.091); Haiku +0.056; fact chunk present in context 35 → 102 of 180Set180 pack self-testsStatusdev (designed on the same set)JudgeClaude Opus (judge.v1, both orders)n180Seeds1Date2026-10-14CaveatLEVER on pack-matched dev data only; inert on the claim kit (ceiling ≈ 0.011) and not a bounty-bar lever.Sourceresearch record
arms/headpass-st/HEADPASS-ST.md:57research recordarms/headpass-st/HEADPASS-ST.md:89research recordreports/REPORT-PACKS-2026-10-13.md:21research recordarms/headpass-st/HEADPASS-ST.md:57,89research recordarms/headpass-stOpen the data file - HEADPASS-X2b (X1 + budgeted dense candidates, +600 tokens per item)HEADPASS-X2b (X1 + budgeted dense candidates, +600 tokens per item)fact present 102/180; Opus +0.066; Haiku +0.056Setthe 180 devStatusdevJudgeOpus + Haikun180Seeds195 % CI(+0.043…+0.091) / (+0.036…+0.076); n 180; 1 seedDate2026-10-14 15:30CaveatLEVER by frozen rule on pack-matched dev only; designed on the same 180Sourceresearch record
arms/headpass-st/HEADPASS-ST.md:57,89Open the data file