Experiment ·
SEEDVAR
How much do the recommended deep configuration's (DT512 arm H: cut 6 + Qwen3.6-35B + thinking 512) V-para numbers move with the sampling seed, adding seed 44 to s42/s43 (comparison: T, D, K two seeds; v0.2 existing rows)?
From the record
- What was tested
- SEEDVAR
- Result
- raw mean 0.550, SD 0.020 (0.533 / 0.572 / 0.544); floor-adjusted mean 0.432, SD 0.025 (0.412 / 0.460 / 0.425); 0 wins each seed
- Verdict
- measurement: measures a property; no lever verdict
- Detail
- measurement only (no verdict rule)
- Set
- V-para (V-para); v0.2 (bench-v02) existing rows
- Judge
- Claude Opus
- n
- 51 judged V-para ids (58 generated), 102 Opus calls; v0.2 110 ids existing rows
- Seeds
- 3 for H (s42/s43/s44); 2 for T, D, K
- Pre-registered
- yes, sha256-frozen
Linked facts · 2
- Three seeds of the recommended deep configuration show how much one run can move.Three seeds of the recommended deep configuration show how much one run can move.raw mean 0.550, SD 0.020 (0.533 / 0.572 / 0.544); floor-adjusted mean 0.432, SD 0.025 (0.412 / 0.460 / 0.425); 0 wins each seedSetV-paraStatusdev-styleJudgeClaude Opus (judge.v1, both orders)n51Seeds3 (42–44)Date2026-10-10Sourceresearch record
reports/BAR-TABLE.md:40research recordreports/RULE-CHANGES.md:16Open the data file - Seed noise: the same deep configuration run with two random seeds differs by more than many tested "levers".Seed noise: the same deep configuration run with two random seeds differs by more than many tested "levers".s42 vs s43, same configuration: Δ −0.039 (−0.072…−0.011) raw; house rule since 2026-10-16: a one-seed deep-tier Δ below 0.05 on V-para is seed noise (raised from 0.03)SetV-paraStatusdev-styleJudgeClaude Opus (judge.v1, both orders)n51Seeds2Date2026-10-16CaveatThe public boar-research README.md:77 still says 0.03.Sourcenot in the public repositories yetOpen the data file