[DevomB]

Summary

Until now the training loop published nothing: new networks stayed in runs/<run>/models/ on the training machine. Now the loop tracks a champion (the best network so far) and can publish each new one as a GitHub Release, so the newest model is always one download away:

https://github.com/DevomB/Gofer/releases/latest/download/gofer-9x9.onnx

How it works

• Who becomes champion. This check runs on anchor generations, every anchor_every laps. If the new network is rated above the current champion on the Elo ladder, it plays the champion directly: up to champion_games games at champion_visits, with defaults of 200 games and 200 visits. It's promoted only if the match's SPRT (champion_sprt, default 0,50) accepts H1. An undecided match promotes nothing. The first network rated at least champion_min_elo (default 1500) becomes the first champion, so the early, weak generations never get released. Training itself still has no gating: the newest network always plays the next self-play.
• What's kept locally. Each champion goes to runs/<run>/champion/gen-NNNN/: the network, champion.json (rating, champion match, anchor results, games and positions so far, SHA-256) and notes.md. champion/gofer-9x9.onnx always holds the current champion.
• Publishing. With --release-repo OWNER/REPO (or GOFER_RELEASE_REPO), the loop runs gh release create model-<run>-gNNNN … --latest with the network as gofer-9x9.onnx plus champion.json, and release notes covering the match result, rating, anchors, network size and checksum.
• Failures. If publishing fails (no gh, bad token, no network), the release stays pending. It's retried on the next generation and at startup, or by hand with python -m training.release --run runs/9x9 --repo OWNER/REPO. If a release already exists it counts as published, and only the newest pending champion is published, so an outage doesn't produce a burst of stale releases.
• Unchanged. The committed models/gofer-9x9.onnx, which the website uses, isn't touched automatically.

The PR also documents the setup in docs/training.md ("Champions and releases") and the README, adds a CHANGELOG entry, and names the MIT license in the README and the site footer.

SPRT fix (second commit)

The arena's SPRT had three defects that could promote a network that isn't actually better:

• It accepted H1 after two games. It estimated variance from the games so far, so after a 2–0 start the variance was 0. The 1e-3 floor then pushed the log-likelihood ratio to about 66 against a bound of 2.94.
• It tested games in the order they finished, so quick results such as resignations were seen first.
• Games finishing after a decision re-ran the test and could overwrite the decision.

The new test counts opening pairs that one side wins twice, takes them in pair order, and runs Wald's SPRT on them. Under the logistic Elo model, P(A sweeps | decisive pair) = 1/(1 + 10^(-2d/400)) whatever the color or opening advantage, because those terms cancel. So Wald's inequality bounds false promotion by alpha/(1-beta) = 5.3% at any game limit, and the decision is taken once, at the first crossing.

TestSPRTErrorRates checks this by simulating 4000 matches per case, with a color advantage and random per-opening advantages:

True situation	Event	Measured rate
Equal strength	false promotion	4.1%
Challenger 50 Elo weaker	false promotion	0.0%
Challenger 50 Elo stronger	false rejection	4.5%
Challenger 150 Elo stronger	promoted within 200 games	99.95%

On the real engine, the final network against the random generation 0 was accepted after exactly 12 sweeps, the theoretical minimum. The network against itself went 35–25 by chance and stayed undecided.

Setup on the training machine

gh auth login          # or: export GH_TOKEN=<fine-grained token, Contents: read and write on DevomB/Gofer>
python -m training.loop --run runs/9x9 --gofer bin/gofer --release-repo DevomB/Gofer

Test plan

• go test ./... (including the SPRT simulation) and pytest training: 21 passed (7 new)
  • The win decision: only SPRT H1 promotes
  • The first champion needs champion_min_elo
  • A lower-rated generation never plays the champion
  • A challenger that loses isn't promoted
  • A promotion writes the files and the state
  • Publishing through a fake gh: correct tag, assets, --repo and --latest; older pending releases superseded; failures stay pending; existing releases count as published; no repository means no publishing
• End-to-end with the real engine: a tiny 4-generation run with a fake gh on PATH
  • Gen 2 became the first champion and was published.
  • Gen 4, rated higher, played the champion with -sprt 0,50, lost 3–6, and was correctly not promoted.
  • python -m training.release reports the state.
• First real release from a training machine with gh authenticated

https://claude.ai/code/session_01JyK2tUFLPsLyWjy2h2erVw

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
gofer	Ready	Preview	Oct 24, 2026 8:31pm UTC