Files
rUv e7005d4d25 docs(gaia): submission reproducibility package (stable 34/53 config) (#2213)
Adds docs/benchmarks/submission/ — the engineering-first reproducibility
package for ruflo's GAIA L1 stable configuration (iter 63, convergence
layer, commit 3ef6e175d).

Package contents:
- predictions.json: 53 per-question answers from iter63 run (HAL schema)
- config.json: exact harness config — model, tools, convergence thresholds
- metadata.md: methodology narrative, component attribution table, honest
  positioning vs HAL (64.2% vs ~82%), rollback discipline rationale
- reproduction.md: exact clone/build/run steps with expected score range
- variance-analysis.md: all measured runs, per-Q flip analysis (47%),
  std dev rationale for submitting mean not peak
- witness.json: Ed25519 signature over file-hash tree (noble/ed25519 v2)

Status: DRAFT — pending n=3 confirmation (iter63b empty, run in flight).
Headline: n=1 only = 34/53 (64.2%).

Ready to submit to HAL pending: (a) n=3 mean ≥ 35, (b) user authorization.

Ref: #2156
2026-05-28 21:02:02 -04:00
..