v1.3.1 — champion restored to the direct head-to-head winner

A direct 300-game head-to-head between the v1.2.0 champion and the v1.3.0 champion (512 sims per move, human-opening book) came out +29 Elo in favor of the v1.2.0 net, which also holds the raw external benchmark record. This release restores that net as champion.pt so the latest release always ships the strongest known weights. The v1.3.0 net remains in service where its human-opening profile is strongest.

champion.pt attached (TorchScript, 20 input planes, auto-detected by webserve).

v1.3.0 — prod12 first promotion (human-opening gate)

New champion from the prod12 training run — the first net trained with the human-opening gate (promotion requires holding a floor on a human-games opening book in addition to the standard arena bar).

Versus the external Ka benchmark (3000 nodes, exhaustive top-20 human openings, 512 sims): 72.5% raw, 96.4% count-weighted, 0.830 sqrt-weighted — the weighted scores are the strongest we have ever measured. Head-to-head it was promoted over the v1.2.0 champion in the internal arena and now serves on the public webapp.

champion.pt attached (TorchScript, 20 input planes, auto-detected by webserve).

Claustrophobia v1.2.0

New strongest champion — a step-change over v1.1.0's net.

Champion (champion.pt): wins the head-to-head match against the previous champion (15-12-3 at equal search) AND sets a new record on our external certification: 77.5% raw score over the exhaustive human-opening suite vs Ka at full search (previous best 65%). Product of an improved teacher-student distillation recipe: doubled self-play corpus plus recalibrated value targets.

Training remains 100% self-play. Same engine as v1.1.0 — drop-in replacement for webapp/champion.pt.

Claustrophobia v1.1.0

Strongest certified champion to date plus a month of engine work.

Champion (champion.pt): our broad-strength best — 65% raw score over the exhaustive 20-line human-opening certification vs Ka at full search (best result to date), and winner of direct head-to-head evaluation against every other champion lineage we have trained.

Engine:

• Monte-Carlo Graph Search (transposition-sharing search, Czech et al. style) alongside the batched tree search
• Gumbel completed-Q policy targets for self-play training (env-gated)
• Human-opening-book promotion gate: champions must now hold up on the openings people actually play
• Search speed stack: 1.8-2.9x per-search, bit-identical results
• Measurement suite: de-degenerate cert metric, per-cell replay baselines, pentanomial reporting, cross-plane evaluation harness
• Teacher-student distillation pipeline (v4) used to seed new training generations

Training remains 100% self-play — no human or external-engine data in any corpus.

Claustrophobia v1.0.0

First public release.

champion.pt — the current strongest network (20 blocks x 256 filters, ~25.5M parameters), trained purely from self-play. Beats the open-source Ka engine at 80000 nodes with pure intuition (2 nodes) and Zero-Ink at maximum thinking time.

Setup: save as webapp/champion.pt, then follow the README quickstart. TorchScript trace — loads on any platform PyTorch supports, GPU or CPU.