Experiment ·
LITE5-180
Pack retrieval reach for the lite tier (HEADPASS-X2b on Qwen3-1.7B Q4_K_M; LITE5 → LITE5-180)
From the record
- What was tested
- Pack retrieval reach for the lite tier
- Result
- +0.071 (+0.048…+0.094)
- Verdict
- adopted: shipped or recommended by its frozen rule
- Detail
- **LEVER (Haiku, dev 180)** (coordinator to review)
- Set
- 180 public self-tests (dev), 1 seed, both arms regenerated on the MacBook, Haiku 5.5-low N = 4 both orders, 720 calls, 0 Opus, 0 paid; criterion 7fb8926d… frozen before any run
- Judge
- Claude Opus, Claude Haiku (L1)
- Pre-registered
- yes, sha256-frozen (7fb8926d…)
- Reopen if
- Opus pass or phone prefill cost if it goes on a card
Linked facts · 3
- R79 X2b on lite 1.7B: 180 HaikuR79 X2b on lite 1.7B: 180 Haiku+0.071 (+0.048…+0.094)Kinda record, not a measurement (set, status, judge, n, seeds, date: not recorded)Caveatwhy closed: LEVER (registered, not a null)Sourcenot in the public repositories yetOpen the data file
- Large levers survive multiplicity correctionLarge levers survive multiplicity correctionX2b z ≈ 5.5, LITE5-180 z ≈ 6, DEEPTHINK z ≈ 4.2Date2026-10-16Kinda record, not a measurement (set, status, judge, n, seeds: not recorded)Caveatcitation fix; ≈ 60 judged comparisons, Bonferroni over 100; the bigger risk is selection on the same dev setsSourcenot in the public repositories yetOpen the data file
- X2b on the 1.7B (LITE5-180)X2b on the 1.7B (LITE5-180)+0.071; 1.7B+X2b − 4B plain −0.007Setthe 180 devStatusdevJudgeHaiku L1n180Seedsnot recorded95 % CI(+0.048…+0.094) / (−0.031…+0.018); n 180Date2026-10-16 00:56Caveat"= 4B" holds only on pack-matched dev; on the kit 1.7B −0.111 and X2b inertSourcenot in the public repositories yetOpen the data file