v0.3.0 - statistical gate preprint and code

arx v0.3.0

Source snapshot: c546a707ef059f26fc570ca40a3697206e184b7e

This release packages the 6-page manuscript arx: A Statistical Gate for
Autonomous Research Loops with the matching source state.

What the evidence supports

• The loop produced zero findings that survived its statistical gate in the
  reported runs.
• The useful result is the separation between broken experiments, statistical
  refusals and apparently successful work.
• The repository documents five offline gate examples and a reproducible claim
  ledger.

Verification captured before publication

• Python 3.12 and 3.13 tests: success.
• examples and audit-example: success.
• The README records 374 tests, with 371 passing and 3 environment-skipped in
  CI.

Limits

• This is an arXiv-style preprint, not a peer-reviewed publication.
• The reported headline is a negative result, not evidence that the loop
  discovers valid findings.
• The case study uses 2019-era encoder checkpoints, and the published judge
  identities are taken on trust, as documented in the repository.

Asset

• arx-v0.3.0-preprint.pdf - 6 pages
• SHA-256:
  902d555985ec9061239b902445c32b9a0eb1b7b9b2d99e679eb4f4fa23d40468
• arx-v0.3.0-SHA256SUMS.txt - standalone checksum file attached to the release

Pending publication command

cd /home/liuzewen/gh-release-drafts-20261028
python3 preflight.py --repo arx
sha256sum -c assets/arx-v0.3.0-SHA256SUMS.txt
gh release create v0.3.0 \
  --repo aidless/arx \
  --target c546a707ef059f26fc570ca40a3697206e184b7e \
  --title "v0.3.0 - statistical gate preprint and code" \
  --notes-file notes/arx-v0.3.0.md \
  assets/arx-v0.3.0-preprint.pdf \
  assets/arx-v0.3.0-SHA256SUMS.txt