Skip to content

Research: publish authenticated RHT-CQER Stage B result - #3

Merged
Labeeb2339 merged 22 commits into
mainfrom
codex/v03-auto-policy
Jul 26, 2026
Merged

Research: publish authenticated RHT-CQER Stage B result#3
Labeeb2339 merged 22 commits into
mainfrom
codex/v03-auto-policy

Conversation

@Labeeb2339

@Labeeb2339 Labeeb2339 commented Jul 26, 2026

Copy link
Copy Markdown
Owner

Summary

  • adds the exact-byte adaptive row-packing research path and records the rejected intermediate hypotheses instead of hiding them
  • freezes, authenticates, and evaluates RHT-CQER-32 on the preregistered 32-task Experiment 009 Stage-B development window
  • publishes two evidence-derived README graphs, a compact release manifest, the result report, and the immutable-artifact verification receipt
  • fixes the strict loader's post-write canonical-manifest ordering assumption without modifying or rerunning the original result
  • makes generated-asset checks work in a clean clone while retaining strict raw-artifact validation for regeneration
  • makes the real tiny-Qwen GPU test compare against an explicit CUDA device index

Authenticated Stage-B result

At the same 2,564,096 packed-state bytes and 2,711,552 selector-aware resident bytes:

  • macro aligned excess NLL: 0.323944 -> 0.153129 (52.73% lower)
  • aggregate local state SSE: 36,409.363073 -> 15,345.844948 (57.85% lower)
  • paired task improvement: 0.170815, frozen 95% bootstrap CI [0.116082, 0.229438]
  • strict task outcomes: 27 wins, 0 ties, 5 losses
  • all eight frozen development advancement gates passed

This is authenticated development evidence on one pinned model and task window. It is not held-out confirmation, novelty, speed, state-of-the-art, deployment, or breakthrough evidence.

Verification

  • full local test suite passed across 632 collected cases
  • GitHub's CPU test job passed with its expected platform-specific skips
  • Ruff lint passed
  • raw-present generated-asset check passed
  • clean-clone manifest-backed generated-asset check passed
  • frozen v0.2 confirmation verifier passed
  • wheel and source distribution built; strict Twine metadata checks passed
  • 61 local documentation links checked with no missing targets
  • staged diff and whitespace checks passed

The 168 MB raw Stage-B JSON remains outside Git history. After merge, its byte-identical ZIP and checksum receipt will be attached to the prerelease tag experiment009-stage-b-cdc603b.

Merge requirement: use a merge commit, not squash, so the exact frozen Stage-A, identity, and evaluator commit SHAs remain ancestors of main and continue to authenticate.

@Labeeb2339
Labeeb2339 marked this pull request as ready for review July 26, 2026 06:47
@Labeeb2339
Labeeb2339 merged commit 9101e04 into main Jul 26, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant