The evolved-NNUE-leaf experiment is audited and launched, and a depth-6 gauntlet game joins the leaderboard by import

What was audited before a single leased seed was read

The request was to run the registered evolutionary-NNUE experiment to a verdict, with a reproducible teacher and a baseline set of weights that later work can improve on, using the Rust solver and the whole machine. Before launching, the theory (TH-20260825-evolved-nnue-leaf-d3-0f47e46c), the frozen experiment (EX-20260825-nnue-evolution-d3-bca7f330), both seed leases, the August 25 smoke-run record, and all 2,800 lines of the nnue-evolution crate were read against the methodology and benchmark contract. The audit found no blocking defect. The points worth writing down:

• The deployed policy is depth 3. The depth-5 search is the teacher only: its sibling-complete root values are distilled into the leaf, and the candidate is then the ordinary d3s7 fair search with that leaf. Anyone reading "depth 5 plus NNUE" should expect a depth-3 player that has seen a depth-5 teacher, not a depth-5 player.
• Labels are search values, not outcomes. That is what makes a per-game move cap on the teacher stage safe: capping a heavy-tailed teacher game truncates the late-game state distribution but changes no label. The frozen protocol left the cap open; it was fixed at 500 moves before launch.
• Common random numbers hold by construction. Every chance scenario in the search is keyed on the public state and the fixed policy seed, so every candidate in a generation, and both controls, face identical futures on identical seeds. Fitness comparisons within a generation are paired.
• The information boundary is structural. The Rust engine draws a covered disc's value at reveal time, so the board bytes the NNUE reads contain no hidden value to leak; the gate that disguises score, level and move count then confirms the network is blind to everything else.
• The 64-game screen can only see a large effect. The prior CMA-ES screen at this cohort size had a detection floor near 40,000 points; a real but smaller improvement will read as a non-measurement. The gate is preregistered and stays as written, but the result must be read against its floor.
• The crash-recovery fix of August 27 is in the binaries that were rebuilt today from this branch (cargo test --release, 9 unit tests).
• make research-validate still reports the eight pre-existing errors from the deleted August 25 run artifacts noted on 2026-08-27; nothing today adds to them.

positive

All eleven CHECK gates pass on the rebuilt binaries

11/11 gates, 160 s, probe block only

Feature determinism and bounds, information-boundary blindness, reflection consistency on eleven asymmetric mirrored roots (one symmetric board excluded per finding-13), fresh-searcher and 1-vs-4-worker determinism, legality and completed depth under random, all-zero and saturated weights, finiteness fuzz, teacher argmax equivalence at depth 3 and depth 5, and serialisation round-trip all passed against 220 harvested probe states from the already-opened block 0xa5277000. Record: runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/gates.log.

positive

The teacher's labels are bit-identical at 16 and 32 threads, and SMT is worth 1.29x

7,830 roots identical; 1.020 vs 1.577 s per root

A seed-free scaling preflight played the same 64 probe games (0xa5277800, depth 4, seven strata) through the corpus generator at 16 and then 32 threads. Every root record and every game record was identical between the two runs once the wall-clock field was ignored, which is the thread-count independence the corpus stage relies on. In the fully concurrent phase a root cost 1.020 s per thread at 16 threads and 1.577 s at 32, so the second hardware thread on each of the 16 physical cores adds about 29% to total throughput. The frozen record declared 16 threads; it was amended to 32 before any leased seed was read, with the preflight as the reason. Records: runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/preflight-d4-t{16,32}.log.

Two of those 64 probe-block games, played by the fair leaf at depth 4, ended above 1.5 million points. They are on already-read development seeds, in a timing preflight, with no comparator: a demonstration that the tail exists, not evidence about anything.

proposed

Stage A is running: 512 depth-5 teacher games on 32 threads

owner: claude-fable-5-1 (Claude Code)

The training lease SL-20260825T063000Z-a52e0300 was marked opened at 04:24 UTC (2026-09-02) and the teacher corpus started on seeds 0xa52e0300-0xa52e0500 with the move cap at 500, a 46,800 s sub-budget after which no new game starts, and per-game part files so an interruption loses nothing finished. Every stage is launched through scripts/pipeline.sh, which hard-codes the leased sub-blocks and records wall, CPU and peak-RSS usage per stage. The supervised warm start, the 60-generation evolution, the 128-game elite re-selection and the one-shot 64-game screen follow in order; the screen lease stays reserved until the frozen candidate's SHA-256 is on disk. The approach page now carries a step-by-step, plain-language account of the evolutionary loop.

Review fix: the amended protocol becomes a successor record

Greptile's review of PR #17 flagged that the two pre-launch amendments had been written into the frozen August 25 experiment record with its hash recomputed, which the repository rule on frozen records does not permit even though earlier records had done the same. The fix follows the P-SOL-1 to P-SOL-2 precedent: the original record is restored byte for byte, marked superseded with one note, and keeps its historical hash; a successor EX-20260902-nnue-evolution-d3-v2-49c18bc2 carries the identical scientific protocol with the operational parameters (move cap, sub-budgets, 16 epochs, 256 probe roots, 32 threads, one-shot screen) fixed in its body. The opened training lease, the run record and the contribution record now reference the successor's preregistration hash, and the evolve and screen binaries were rebuilt with the successor's identifier before stage C so their artifacts name the record they ran under; the corpus binary and its running process were untouched. Two script findings from the same review were also fixed: the screen stage now performs the held-out lease transition itself, so no invocation path can read the block while the registry still says reserved, and the corpus wall default matches the protocol.

---

Importing a game instead of re-playing it

On 27 August the operator played rust-fair-d6-s7 through gauntlet-01 on the 16-core workstation with just play-round, and the resulting runs/BENCH-* directory was copied to the research laptop today. The goal was to put that game on the public competition leaderboard, which lives in the DynamoDB ledger rather than in web/data/leaderboard.json.

The ledger's seeding command (npm run competition -- seed) was built for the fast TypeScript policies: it plays each policy live on the immutable round, replays the chosen columns independently, packs them, and does a conditional insert. Used as designed it would have re-run the whole depth-6 search here, and the run artifact records the original game at 9,925,584 ms of wall time on the larger machine. So the command gained a --replay option. It reads the bench replay file, checks that it names the competition's round and a registered policy, replays the recorded columns from the opening position, and requires the replayed score, move count, censor flag, every per-move board, and the sixteen-hex trajectory checksum to match the recording exactly. The policy itself never runs. The bench checksum moved into an exported helper in src/bench/runner.ts so the bench and the importer compute it the same way; two tests cover acceptance and rejection (another round, a changed column, an inflated score, illegal play, a wrong checksum).

positive

The recorded depth-6 game replays exactly and is now in both ledgers

428,292 points in 120 moves

The imported replay (runs/BENCH-20260827T150335Z-rust-fair-d6-s7--gauntlet-01/, generated 2026-08-27T17:49:01Z) reproduced under the independent competition replay: 428,292 points, 120 moves, game over rather than censored, no illegal decisions, longest chain 11, 248 discs cleared, 139 covered discs revealed, trajectory checksum df69d802b6e3b377. The same record was inserted into the production and dev competition ledgers under global#2026-08-v1 with submission id 8ffae14844e9767ac989aee2197186afceef33c8bee7f6a52e644a053af004e5, and both items were read back and compared field for field with the dry-run preview. It enters the production leaderboard behind the top human entry and above every previously seeded policy. That is a playground demonstration on one scripted future and says nothing about policy strength; scripted rounds are never tier evidence.

Two records differ deliberately from a live seed. The source revision is declared by the operator rather than observed by the command: 4bebccdff26d9335c25740161fc101ded535e22c, the empty-frontier fix committed minutes before the run started, which a depth-6 game needs to survive near a board-filling rise. The command verifies that the named commit exists in this repository and nothing more. The dirty-worktree flag is omitted for imports instead of being asserted, and the ledger timestamps are the import time, not the 27 August completion time.

Open: nothing has shown that the current decide binary at HEAD reproduces this trajectory, only that the recorded columns are legal and score as claimed on the fixed round. The runs/ directory holding the original artifact is gitignored and local to this laptop, so the ledger record and this entry are the durable references to the game. No research seeds were opened and no research record changed.