relower: paired per-item noise floor over >=3 prompted repeats; panel gain clears the noise of the sum

weksa commented

2026-11-13 02:33:03 +00:00

Member

Follow-up golf-loop-paired-floor. A one-pair floor misread a prompted repeat as Regressed(Meaning) (ja-JP floor meaning 0.000 from two means that met). The floor is now FLOOR_Z=3 standard errors of the paired per-item difference, pooled over three or more floor arms (paired_floor); decide_panel's gain test requires the sum of composite deltas to clear the root of the floors squared (a prompted repeat read Accepted 40% of trials without it). decide stays the one-language owner. The cycle's floor (floor_of, two replicate means) is unchanged: round records hold no per-item scores.

Calibration on runs/distil-base (8 independent prompted draws after Soul s5 found p7 overlapping a degraded arm and redrew it; 14 en + 14 ja Validation items; 28 pairs; 95% bootstrap over draws): prompted repeat NoGain 92% (80-100), Regressed 8% (0-20), Accepted 0% (bounds false-accept only to ~10%); Yes. arm Regressed 100%; mild degradation 100%; subtler degradation 75% (50-100). Z was chosen on this data (in-sample); no better-than-base arm exists, so gain power is unmeasured. Soul verdict weksa-lowering:verdict:cut-golf-loop-edit-accept.s5.

🤖 Generated with Claude Code

Follow-up golf-loop-paired-floor. A one-pair floor misread a prompted repeat as Regressed(Meaning) (ja-JP floor meaning 0.000 from two means that met). The floor is now FLOOR_Z=3 standard errors of the paired per-item difference, pooled over three or more floor arms (`paired_floor`); `decide_panel`'s gain test requires the sum of composite deltas to clear the root of the floors squared (a prompted repeat read Accepted 40% of trials without it). `decide` stays the one-language owner. The cycle's floor (floor_of, two replicate means) is unchanged: round records hold no per-item scores. Calibration on runs/distil-base (8 independent prompted draws after Soul s5 found p7 overlapping a degraded arm and redrew it; 14 en + 14 ja Validation items; 28 pairs; 95% bootstrap over draws): prompted repeat NoGain 92% (80-100), Regressed 8% (0-20), Accepted 0% (bounds false-accept only to ~10%); Yes. arm Regressed 100%; mild degradation 100%; subtler degradation 75% (50-100). Z was chosen on this data (in-sample); no better-than-base arm exists, so gain power is unmeasured. Soul verdict weksa-lowering:verdict:cut-golf-loop-edit-accept.s5. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

weksa added 1 commit

2026-11-13 02:33:03 +00:00

relower: paired per-item noise floor over three or more prompted repeats; panel gain clears the noise of the sum dcbe369f99

A one-pair floor read Regressed(Meaning) on a prompted repeat when two means met by chance. The floor is now FLOOR_Z standard errors of the paired per-item difference, pooled over >=3 floor arms (paired_floor); decide_panel's gain test requires the sum of composite deltas to clear the root of the floors squared. Calibrated on runs/distil-base: prompted repeat NoGain 92%, degraded arms Regressed.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

weksa added 1 commit

2026-11-13 03:01:42 +00:00

relower floor: pin FLOOR_Z and pooling, drop dead guards, recalibrate on clean draws ca4e50d2b5

Soul s5: draw p7 shared 13 answers with a degraded arm; redrawn, plus p8. Numbers in docs and FLOOR_Z doc now from 8 independent draws.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

weksa merged commit 353dd2138f into main

2026-11-13 03:01:53 +00:00

weksa referenced this pull request from a commit

2026-11-13 03:01:55 +00:00

relower: paired per-item noise floor over >=3 prompted repeats; panel gain clears the noise of the sum (#70)

Rows
Columns

relower: paired per-item noise floor over >=3 prompted repeats; panel gain clears the noise of the sum #70