Source 516da1a9ef8c07310f8bc987854df3ab67c35cf6; run voxelshogi-phase5-policy-diagnosis1-control; 12 committed matches of the unchanged12 target, 13 actual observed neural queries. Pipeline complete: True. Both arms use identical G2 FP32 weights bf4fba2699d9d8578a821766b0fcaf861d56aab3effc296d87ec3a4e9408c8a4, all three previously fixed mate-offset2/3/4 positions, seeds, color swaps, one additional ply and Rust M1 depth3/nodes20000/TT4096. Only model policy differs: legal-argmax-v1 versus puct32-v1 (32 simulations, depth64, cpuct1, no root noise or temperature sampling, fresh tree each move and smallest-ID ties). This results-led development diagnosis follows the fully retained150-game comparison, whose protected-input overlap4 was reported. Three positions share one tactical family; no selected stronger model, training or Web adoption is claimed.

Usage and schema

Read either Parquet with pyarrow.parquet.read_table and to_pylist. The development split has arm, protocol_sha256, case_id, model_side, outcome and match_json strings. match_json preserves each original committed UTF-8 file including whitespace/newline; verify against cohort_receipts. The observations split has query_index:int64 and arm,role,case_id,side,physical_id,input_id,vsfen,record_json strings. Each row records a real model call, including off-trajectory nonterminal PUCT nodes; side is the queried side to move, which can differ from the root model color. Repeated calls remain repeated observations. record_json retains the complete Rust branch history; these are observations, not supervised training samples. Partial attempted queries are not completed inference or games.

Verification

Every committed match and every observed query history was reconstructed through Rust, canonical records/VSFEN and physical/side-canonical IDs checked, original match hashes checked and both typed Parquet tables roundtrip verified. When present, input_audit enumerates all M4 splits, all P5-2 published raw samples and protected IDs against all played continuations/terminals plus all actual nonterminal neural queries, including PUCT branches. False or missing audit status is not a clean result. Both full arms and all12 games, fixed conditions and the negative/positive/inconclusive decision are retained. Natural wins/losses/draws and cutoff remain distinct; one-ply cutoff does not adjudicate a loss. Publication does not prove later resource release or population strength.

Access and license

Synthetic evaluation data and observations, CC-BY-4.0, VoxelShogi contributors; attribute immutable source/HF revisions. Public/manual download approval and Community disabled. Model/state weights remain separately gated Apache-2.0. Embedded execution code retains Apache-2.0, with license text in manifest.json. No tensors, optimizer state, provider credentials, HF token, account/billing data, raw host logs or human games are included.

Downloads last month: 101

Total file size:

408 kB