Experiment ·
LENGTHSHARE
Measurement: how much does the 3-dim floor-adjusted ratio of H (deep + thinking 512, s42) and K (4B cut 6, s42) move on V-para when the same answers are re-judged against the frontier+web references condensed to <= 80 words, vs the original ~280-word references (control = original judgments).
From the record
- What was tested
- LENGTHSHARE
- Result
- deep + think 512 floor-adjusted 0.412 → 0.514 (0.456–0.578) with ≤ 80-word references; paired +0.102 (+0.068…+0.144); ≈ 17 % of the gap is length, 83 % substance
- Verdict
- measurement: measures a property; no lever verdict
- Detail
- closed (descriptive measurement done; no LEVER/NO verdict by design, criterion stage 4; CLOSED.md:11 lists it as a measurement)
- Set
- V-para (V-para); v0.2 (bench-v02, Part 1 dimshare only)
- Judge
- Claude Opus
- n
- 51 V-para judged ids, both orders, 2 arms (H, K)
- Seeds
- 1 (s42; re-judge only)
- Pre-registered
- yes, sha256-frozen
Linked facts · 1
- Part of the gap to the reference is length: shortening the references to 80 words raises the deep tier's score, mostly through completeness and usefulness.Part of the gap to the reference is length: shortening the references to 80 words raises the deep tier's score, mostly through completeness and usefulness.deep + think 512 floor-adjusted 0.412 → 0.514 (0.456–0.578) with ≤ 80-word references; paired +0.102 (+0.068…+0.144); ≈ 17 % of the gap is length, 83 % substanceSetV-paraStatusdev-styleJudgeClaude Opus (judge.v1, both orders)n51Seeds1 (42)Date2026-10-11CaveatDiagnostic with altered references; not a bar reading.Sourceresearch record
reports/BAR-TABLE.md:98research recordreports/BAR-TABLE.md:97-98research recordreports/BAR-TABLE.md:97Open the data file