Skip to content

lab_log.md

Latest commit

 

History

History
639 lines (424 loc) · 31.7 KB

lab_log.md

File metadata and controls

639 lines (424 loc) · 31.7 KB
Raw

Lab log

Chronological record of changes, test runs, failures, and limitations.

Bootstrap

  • Initial repository skeleton with Podman-first workflow.
  • Engine: original Rust UCI implementation with full movegen, negamax search, and verifier tooling.
  • ./scripts/podman/ci: PASS (fmt, clippy, tests, verify-smoke, tournament-smoke).

CI run 2026-06-20T08:37:48Z

  • Result: FAIL
  • Failures: cargo fmt --check cargo clippy
  • Command: ./scripts/podman/ci

CI run 2026-06-20T08:38:26Z

  • Result: PASS
  • Command: ./scripts/podman/ci

CI run 2026-06-20T10:25:49Z

  • Result: FAIL
  • Failures: cargo fmt --check
  • Command: ./scripts/podman/ci

CI run 2026-06-20T10:26:20Z

  • Result: PASS
  • Command: ./scripts/podman/ci

Sprint 1 — Hard verification (2026-06-20)

  • Result: PASS
  • Command: ./scripts/podman/verify-deep
  • Perft depth 1–6 (startpos, kiwipete, all EPD fixtures)
  • 200-game random fuzz (seed=42)
  • Legality oracle: 26 positions
  • cozy_crosscheck + shakmaty_crosscheck depths 1–4 on startpos

Sprint 2 — Gauntlet (2026-06-20)

  • Smoke: PASS — ./scripts/podman/gauntlet --smoke (8 games, 0 illegal moves)
  • Full: PASS — ./scripts/podman/gauntlet --games 100 — 200 games aggregate (4×50), 0 illegal moves/crashes; log: docs/gauntlet/gauntlet_20260620T105634Z.log

Sprint 3 — Human play (2026-06-20)

  • Docs: user_manual.md, human_play_checklist.md, play-uci.sh
  • UCI: id version 0.2.0 in handshake
  • Automated protocol QA: PASS (uci_protocol_tester, gauntlet TC)
  • Manual 10-game GUI checklist: pending operator (see human_play_checklist.md)

Sprint 4 — Ops / release (2026-06-20)

  • Result: PASS
  • Command: ./scripts/podman/release
  • Version: 0.2.0
  • SHA256: 63b8173f183aaf0bfa1080216a8ebd0b5f68b4ed71a217e51665c50849f6167e
  • CHANGELOG.md, submission_package.md, panic logging to stderr

Sprint 5 — Lichess bot (2026-06-20)

  • Bridge (live): lichess-bot + host target/release/labzero
  • Dev stub: lichess_bot/bot.py, ./scripts/podman/bot --dry-run
  • Dry-run: PASS — 20 plies, zero illegal moves
  • Live 5+ games: pending — see lichess_bot_setup.md

Sprint 6 — Public candidate (2026-06-20)

  • GHA: .github/workflows/ci.yml — smoke on push/PR; verify-deep + gauntlet-smoke weekly
  • README: badges, quickstart, links to user manual + submission pack
  • Submission pack: docs/submission_package.md (reviewer-ready except live Lichess URL)

Release v0.2.0 2026-06-20T10:57:00Z

  • Binary: .cargo-target/release/labzero
  • SHA256: 63b8173f183aaf0bfa1080216a8ebd0b5f68b4ed71a217e51665c50849f6167e
  • Version: 0.2.0

Release v0.2.0 2026-06-20T10:56:43Z

  • Binary: .cargo-target/release/labzero
  • SHA256: 63b8173f183aaf0bfa1080216a8ebd0b5f68b4ed71a217e51665c50849f6167e
  • Version: 0.2.0

CI run 2026-06-20T12:25:21Z

  • Result: PASS
  • Command: ./scripts/podman/ci

Host benchmark round 2 (2026-06-20)

  • SF UCI Elo: 2000, 1+0, 32 games
  • Score (labzero): 2–23–7 (17.2%)
  • Performance Elo (approx): 1727
  • Illegal / errors: 0
  • Artifacts: docs/strength/benchmark_20260620T192641Z.txt, .pgn

Beta v0.3.0-beta ladder (2026-06-20)

  • Engine: qsearch, TT, null move, LMR, SEE, tapered eval
  • 1320: 28–0–4 (93.8%) — benchmark_20260620T214301Z
  • 1800: 14–13–5 (51.6%) — benchmark_20260620T221648Z
  • 2000: 10–20–2 (34.4%) — benchmark_20260620T230310Z
  • Gauntlet smoke: PASS, 0 illegal

CI run 2026-06-20T20:37:54Z

  • Result: FAIL
  • Failures: cargo fmt --check cargo clippy
  • Command: ./scripts/podman/ci

CI run 2026-06-20T20:40:58Z

  • Result: FAIL
  • Failures: cargo clippy cargo test build engine verify smoke tournament smoke
  • Command: ./scripts/podman/ci

CI run 2026-06-20T20:41:18Z

  • Result: FAIL
  • Failures: cargo clippy cargo test build engine verify smoke tournament smoke
  • Command: ./scripts/podman/ci

CI run 2026-06-20T20:41:47Z

  • Result: PASS
  • Command: ./scripts/podman/ci

CI run 2026-06-20T21:42:59Z

  • Result: FAIL
  • Failures: cargo clippy
  • Command: ./scripts/podman/ci

CI run 2026-06-20T21:43:34Z

  • Result: PASS
  • Command: ./scripts/podman/ci

CI run 2026-06-21T05:37:09Z

  • Result: FAIL
  • Failures: cargo fmt --check cargo clippy
  • Command: ./scripts/podman/ci

CI run 2026-06-21T05:37:40Z

  • Result: PASS
  • Command: ./scripts/podman/ci

Phase C — gamma v0.4.0 (2026-06-21)

  • Search: depth cap 64, aspiration, PV ordering, root make/unmake, check qsearch evasions; TT ordering-only under movetime (score cutoffs reverted after ladder blunders)
  • Time: soft stop before new ID, panic reserve, wtime/increment allocation
  • UCI: info depth score cp nodes nps time, options Hash, Threads, OwnBook/BookFile
  • Eval: pawn structure, rook files, king safety enabled
  • SMP: Lazy SMP (Threads 1–8, shared thread-safe TT)
  • Book: optional opening lines (OwnBook off by default)
  • CI: ./scripts/podman/ci PASS (13 unit tests)
  • Ladder: anchor rows recorded in docs/strength/ladder.md (superseded by v0.5.0 re-measure)

Phase D — strengthening v0.5.0 (2026-06-21)

  • D1a: verifier/positions/tactical.epd (10 positions); fixed-depth tactical tests in search.rs
  • D1b: TT complete flag; score cutoffs enabled for go depth only — movetime remains ordering-only (complete-guard cutoffs still regressed SF@1320 under 1s)
  • D1c: LMR reduction move_idx/10 (was /8); aspiration windows from depth 5 (was 4)
  • D2: UCI protocol matrix; ucinewgame clears stop flag; wtime benchmark fix in host-benchmark.sh
  • D3: eval weight tune — doubled pawn penalty 10→8, isolated 8→6, rook file 12→10
  • CI: ./scripts/podman/ci PASS (16 unit tests); gauntlet smoke 0 illegal
  • Ladder (1+0 anchor): SF@1320 14–0–2 (93.8%); SF@2000 9–17–6 (37.5%, 32-game confirm, ≈1911 perf) — 0 illegal; see docs/strength/ladder.md
  • Spot wtime 3+2 @ SF2000 (8 games, pre-harness-fix): 0–8–0 — superseded; see below (benchmark_20260621T063138Z)

Host benchmark — gamma anchor (2026-06-21)

Post–TT-fix v0.5.0, TC_MODE=movetime TC_SEC=1 THREADS=1, 16 games each:

SF_ELO Score % Artifact
1320 15–1–0 93.8% benchmark_20260621T063529Z
1900 5–6–5 46.9% benchmark_20260621T064836Z
2000 5–8–3 40.6% benchmark_20260621T071050Z
2100 3–10–3 28.1% benchmark_20260621T073101Z

Invalidated: benchmark_20260621T062854Z (1–15 @ 1320) — TT score cutoffs under movetime (pre-fix).

CI run 2026-06-21T06:25:39Z

  • Result: FAIL
  • Failures: cargo fmt --check cargo clippy verify smoke
  • Command: ./scripts/podman/ci

CI run 2026-06-21T06:27:18Z

  • Result: PASS
  • Command: ./scripts/podman/ci

CI run 2026-06-21T08:04:32Z

  • Result: PASS
  • Command: ./scripts/podman/ci

Host benchmark — paper confirm (2026-06-21)

Post–TT-fix v0.5.0, TC_MODE=movetime TC_SEC=1 THREADS=1:

Phase SF_ELO Score % Perf Elo Artifact
Regression 1320 14–0–2 93.8% — benchmark_20260621T084207Z
Confirm 2000 9–17–6 37.5% ≈ 1911 (95% CI ≈ 1790–2030) benchmark_20260621T091006Z

0 illegal, 0 errors on both runs.

Prior 16-game probes (same protocol): SF@1900 46.9%, SF@2100 28.1% — see docs/strength/ladder.md.

Next: SMP spot THREADS=8 @ SF@2000 — 2–9–5 (28.1%, perf ≈ 1837); no gain vs T=1 confirm (benchmark_20260621T095930Z).

SMP spot — Threads=8 (2026-06-21)

TC_MODE=movetime TC_SEC=1 THREADS=8, 16 games vs SF@2000:

Score % Perf Elo (approx) vs T=1 confirm Artifact
2–9–5 28.1% ≈ 1837 (CI ≈ 1660–2020) −9.4 pp / ≈ −74 Elo benchmark_20260621T095930Z

0 illegal, 0 errors. CIs overlap — treat as negative spot result, not proven SMP regression.

Spot blitz — wtime 3+2 @ SF@2000 (2026-06-21)

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=1, vs SF@2000:

Games Score % Perf Elo (approx) Artifact Notes
16 (probe) 7–6–3 53.1% ≈ 2022 benchmark_20260621T132942Z superseded by 32-game confirm
32 (confirm) 10–11–11 48.4% ≈ 1989 (95% CI ≈ 1860–2115) benchmark_20260621T140403Z paper-grade blitz row

0 illegal, 0 errors on both runs. Supersedes pre–host-benchmark.sh clock fix 0–8–0 (benchmark_20260621T063138Z). Not comparable to 1+0 anchor (37.5%, ≈1911).

Sharded TT + SMP spot — 3+2 (2026-06-21)

Change: engine/src/tt.rs — global Mutex<Vec<_>> replaced with 64-shard table (per-shard mutex); API unchanged.

TC_MODE=wtime TC_SEC=3 TC_INC=2, 16 games vs SF@2000:

Threads Score % Perf Elo (approx) vs 32g blitz confirm Artifact
1 5–8–3 40.6% ≈ 1934 −7.8 pp (noise) benchmark_20260621T151817Z
4 2–9–5 28.1% ≈ 1837 −20.3 pp benchmark_20260621T154803Z

0 illegal, 0 errors. T=1 in line with 48.4% confirm; T=4 still regressed (same score band as pre-shard THREADS=8 @ 1+0). Conclusion: lock contention was not the sole bottleneck.

Zobrist hash + SMP re-measure — 3+2 (2026-06-21)

Change: engine/src/zobrist.rs — deterministic Zobrist tables; make_unmake.rs incremental XOR on make (no full-board recompute); compute_hash() kept as oracle.

TC_MODE=wtime TC_SEC=3 TC_INC=2, 16 games vs SF@2000:

Threads Score % Perf Elo (approx) vs post-shard T=4 Artifact
1 10–6–0 62.5% ≈ 2089 +21.9 pp vs shard T=1 benchmark_20260621T162932Z
4 7–7–2 50.0% ≈ 2000 +21.9 pp vs shard T=4 benchmark_20260621T165748Z

0 illegal, 0 errors. Decision: keep Zobrist hash; T=4 now even on 3+2 (was 28.1%). Next SMP work: helper depth offset / split diversification for further gain.

Lazy SMP v2 — helper start-depth diversification (2026-06-21)

Change: search_with_info_from_depth(start_depth) in search.rs — helpers skip shallow ID plies; aspiration gated on prior completed score (fixes cold-start when helper begins at depth 5). smp.rs — helpers cycle start depths 3, 4, 5; main thread unchanged at depth 1.

TC_MODE=wtime TC_SEC=3 TC_INC=2, 16 games vs SF@2000:

Threads Score % Perf Elo (approx) vs post-Zobrist Artifact
1 3–8–5 34.4% ≈ 1888 −28.1 pp (noise; main path unchanged) benchmark_20260621T181146Z
4 7–5–4 56.2% ≈ 2044 +6.2 pp vs post-Zobrist T=4 benchmark_20260621T184359Z

0 illegal, 0 errors. Decision: keep Lazy SMP v2 — T=4 9/16 W-equivalent (≥ keep threshold); beats post-Zobrist 7–7–2.

32-game confirm — Lazy SMP v2 (2026-06-22)

TC_MODE=wtime TC_SEC=3 TC_INC=2, vs SF@2000:

Threads Score % Perf Elo (approx) Notes Artifact
4 13–12–7 51.6% ≈ 2011 paper-grade T=4 confirm; ~even vs SF@2000 benchmark_20260622T120949Z
8 11–15–6 43.8% ≈ 1956 diagnostic; worse than T=4 benchmark_20260622T131945Z

0 illegal, 0 errors. Headline: ≈2010 on this protocol (32g T=4 confirm); 16g T=4 spot ≈ 2044 was high-variance.

Null-move EP correctness (2026-06-22)

Change: NullUndo in board.rs — null move clears ep_square and EP hash XOR; unnull_move restores prior EP state. Fixes stale EP rights / wrong Zobrist key in null-move subtrees.

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4, 32 games vs SF@2000:

Score % Perf Elo (approx) vs SMP v2 anchor Artifact
15–9–8 59.4% ≈ 2066 +7.8 pp vs 13–12–7 benchmark_20260622T144847Z

0 illegal, 0 errors. Decision: keep — correctness fix; tests + smoke pass.

Timed TT score cutoffs re-enabled (2026-06-22)

Change: search.rs — remove is_timed() gate around tt_cutoff; safeguards unchanged (complete, depth, mate exclusion, partial-node no-store).

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4, 32 games vs SF@2000:

Score % Perf Elo (approx) vs SMP v2 anchor (13–12–7) Artifact
18–9–5 64.1% ≈ 2100 +12.5 pp benchmark_20260622T160120Z

0 illegal, 0 errors. Decision: keep — 20.5/32 W-equivalent (≥ 16/32 threshold).

Headline validation — SF@2100 direct (2026-06-22)

Protocol: TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4, vs limited Stockfish:

SF_ELO Games Score % Perf Elo (approx) Decision Artifact
2100 32 19–11–2 62.5% ≈ 2189 Keep headline ≈2100 (>60%) benchmark_20260622T172332Z
2200 16 4–8–4 37.5% ≈ 2111 probe (below 50% vs SF@2200) benchmark_20260622T181751Z

0 illegal, 0 errors on both runs.

Qsearch-in-check fix (2026-06-22)

Change: search.rs qsearch — compute in_check before stand-pat; skip stand-pat in check; search all legal evasions (not noisy-only) until QSEARCH_MAX.

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4, 32 games vs SF@2000:

Score % Perf Elo (approx) vs anchor (18–9–5) Artifact
16–7–9 64.1% ≈ 2100 20.5/32 W-equiv (≥ 18/32) benchmark_20260622T185217Z

0 illegal, 0 errors. Decision: keep.

SEE capture-ordering sign fix (2026-06-23)

Change: engine/src/see.rs — see_capture_value returns positive for materially favorable captures and negative for losing ones (early return when no recapture; invert defender-stop path). Move-ordering callers unchanged. No qsearch pruning in this patch.

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4:

SF_ELO Games Score % Perf Elo (approx) vs anchor Artifact
2000 32 18–9–5 64.1% ≈ 2100 20.5/32 W-equiv (keep) benchmark_20260623T044106Z
2200 16 5–3–8 56.2% ≈ 2245 probe (≥ 7.5/16) benchmark_20260623T055140Z
2200 32 15–14–3 51.6% ≈ 2211 ~even vs SF@2200 benchmark_20260623T062210Z

0 illegal, 0 errors. Decision: keep — SF@2000 gate passed; SF@2200 improved vs pre-fix probe (4–8–4). README headline stays ≈2100.

Deferred: qsearch SEE pruning — prior attempt regressed to 5–22–5 (benchmark_20260622T205652Z, rolled back) because pre-fix SEE sign was inverted.

Qsearch SEE pruning retry (2026-06-23)

Change (rolled back): conservative qsearch SEE prune — margin -250, skip first qsearch ply (qs_depth == 0) and all in-check evasions; promotions/EP always kept.

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4, 32 games vs SF@2000: 15–11–6 (56.2%, 18/32 W-equiv). 0 illegal, 0 errors. Rollback — below 19/32 gate (anchor 20.5/32, benchmark_20260623T044106Z). Artifact: benchmark_20260623T073759Z.

History gravity + quiet malus (2026-06-23)

Change: search.rs — bounded history updates (HISTORY_MAX=16384, history_bonus(depth) capped at 2048, gravity formula); on quiet beta cutoffs, positive gravity on cutoff move and negative gravity on earlier quiet non-cutoffs in the same ordered list. Killers unchanged.

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4:

SF_ELO Games Score % Perf Elo (approx) W-equiv Artifact
2000 32 20–6–6 71.9% ≈ 2163 23/32 (keep) benchmark_20260623T090318Z
2200 16 7–7–2 50.0% ≈ 2200 8/16 (probe) benchmark_20260623T100753Z
2200 32 12–7–13 57.8% ≈ 2257 18.5/32 (strong keep) benchmark_20260623T104424Z

0 illegal, 0 errors. Decision: keep — SF@2000 +2.5 W-equiv vs SEE-fix anchor; SF@2200 32g 18.5/32 (≥ 17/32 strong tier). Headline revised to ≈2200 (README, ladder) from SF@2200 32g perf ≈ 2257.

Eval v2: passed pawns + mobility (2026-06-23)

Change: engine/src/eval.rs — original passed-pawn bonuses by relative rank (MG/EG tables), protected-passer bonus, and pseudo-attack mobility for knights/bishops/rooks/queens; phase-tapered. Four unit tests in eval.rs.

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4:

SF_ELO Games Score % W-equiv Artifact
2300 16 8–4–4 62.5% 10/16 (probe) benchmark_20260623T131324Z
2200 16 12–1–3 84.4% 13.5/16 (keep) benchmark_20260623T134940Z
2300 32 19–5–8 71.9% 23/32 (headline) benchmark_20260623T150655Z

0 illegal, 0 errors. Decision: keep — SF@2200 keep gate passed (≥ 8/16); SF@2300 32g 23/32 (≥ 16/32 headline gate). Headline revised to ≈2300 (README, ladder).

Direct SF@2400/2500 measurement (2026-06-23)

Change: none (measurement-only on eval v2 baseline).

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4:

SF_ELO Games Score % W-equiv Artifact
2400 16 5–4–7 53.1% 8.5/16 (probe) benchmark_20260623T163900Z
2500 16 3–3–10 50.0% 8/16 (probe) benchmark_20260623T171507Z
2400 32 11–9–12 53.1% 17/32 (headline) benchmark_20260623T175115Z

0 illegal, 0 errors. Decision: SF@2400 32g ≥ 16/32 headline gate → ≈2400 (README, ladder). No SF@2500 32g (2400 32g 17/32 < 18/32 trigger).

Eval v3: king pressure + hanging threats (2026-06-23)

Change: eval.rs — king-pressure ring attackers + hanging-piece threat penalties (reverted).

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4:

SF_ELO Games Score % W-equiv Artifact
2300 16 5–3–8 56.2% 9/16 (keep) benchmark_20260623T191448Z

0 illegal, 0 errors. Rollback — below 10/16 SF@2300 keep gate. Full revert on codex/eval-v3-king-pressure.

Direct SF@2500 32g measurement (2026-06-24)

Change: none (measurement-only on eval v2 baseline).

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4:

SF_ELO Games Score % W-equiv Artifact
2500 32 8–12–12 43.8% 14/32 (baseline) benchmark_20260624T035947Z

0 illegal, 0 errors. Decision: below 16/32 headline gate — keep ≈2400; row is PVS baseline.

PVS v1: null-window search (2026-06-24)

Change: search.rs — PVS for non-first quiet moves at depth >= 2, skip in check (reverted).

TC_MODE=wtime TC_SEC=3 TC_INC=2 THREADS=4:

SF_ELO Games Score % W-equiv Artifact
2400 16 7–7–2 50.0% 8/16 (keep) benchmark_20260624T052632Z
2500 32 11–14–7 45.3% 14.5/32 (headline) benchmark_20260624T055916Z

Startpos depth-8 nodes: 198070 pre-PVS → 292922 post-PVS (+48%). 0 illegal, 0 errors. Rollback — headline gate miss (14.5/32 < 16/32) and node regression; SF@2500 vs baseline within noise only.

CI run 2026-06-21T17:51:27Z

  • Result: PASS
  • Command: ./scripts/podman/ci

Gauntlet baseline_sf2500 (2026-06-24)

  • Result: complete — 18-7-7 (67.2%), perf ≈ 2624
  • Opponent: SF UCI_Elo=2500, TC 3+2, 32 games
  • Artifact: docs/strength/baseline_sf2500.txt

Gauntlet gate_sf2500_smoke (2026-06-24)

  • Result: complete — 6-3-7 (59.4%), perf ≈ 2666
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2500_smoke.txt

Gauntlet confirm_sf2600_32g (2026-06-24)

  • Result: complete — 10-12-10 (46.9%), perf ≈ 2578
  • Opponent: SF UCI_Elo=2600, TC 3+2, 32 games
  • Artifact: docs/strength/confirm_sf2600_32g.txt

Gauntlet gate_sf2500_32g (2026-06-24)

  • Result: complete — 11-10-11 (51.6%), perf ≈ 2511
  • Opponent: SF UCI_Elo=2500, TC 3+2, 32 games
  • Artifact: docs/strength/gate_sf2500_32g.txt
  • SPSA s2 rollback: worse than baseline_sf2500 (67.2%); do not ship tuned params in v0.6.0

Release v0.6.0 — superhuman-band sprint S1 (2026-06-24)

  • Branch: codex/superhuman-band merged to main
  • Keep: magic bitboards, mailbox, movegen speed, gauntlet/SPSA/NNUE tooling (NNUE off by default)
  • Headline: ≈2600 — baseline_sf2500 32g @ SF@2500 3+2 T=4 → 18–7–7 (67.2%, perf ≈2624)
  • Rollback: SPSA smoke + s2 eval params (failed 32g keep gates vs baseline)
  • Next: NNUE sprint on branch or v0.7.0 bet

Gauntlet gate_sf2500_v060_16g (2026-06-25)

  • Result: complete — 6-3-7 (59.4%), perf ≈ 2566
  • Opponent: SF UCI_Elo=2500, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2500_v060_16g.txt

Gauntlet gate_sf2500_v060_32g (2026-06-25)

  • Result: complete — 15-8-9 (60.9%), perf ≈ 2577
  • Opponent: SF UCI_Elo=2500, TC 3+2, 32 games
  • Artifact: docs/strength/gate_sf2500_v060_32g.txt

Sprint S3 — ID time-depth fix (2026-06-25)

Change: search.rs — last_iter_ms is now per-iteration wall time for TimeBudget::should_start_depth (was cumulative elapsed). Branch codex/id-time-depth from 731af6d.

Gate SF@2600: 16g 7–5–4 (56.2%, 9/16 W-equiv); 32g 13–9–10 (56.2%, 18/32 W-equiv — KEEP), perf ≈ 2644, 0 illegal/errors. Direct ≈2600 claim validated (gate_sf2600_idtime_32g).

Gauntlet gate_sf2600_idtime_16g (2026-06-25)

  • Result: complete — 7-5-4 (56.2%), perf ≈ 2644
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_idtime_16g.txt

Gauntlet gate_sf2600_idtime_32g (2026-06-25)

  • Result: complete — 13-9-10 (56.2%), perf ≈ 2644
  • Opponent: SF UCI_Elo=2600, TC 3+2, 32 games
  • Artifact: docs/strength/gate_sf2600_idtime_32g.txt

Gauntlet gate_sf2700_baseline_16g (2026-06-26)

  • Result: complete — 2-4-10 (43.8%), perf ≈ 2656
  • Opponent: SF UCI_Elo=2700, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2700_baseline_16g.txt

Gauntlet gate_sf2600_lmrhist_16g (2026-06-26)

  • Result: complete — 7-6-3 (53.1%), perf ≈ 2622
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_lmrhist_16g.txt

Gauntlet gate_sf2700_lmrhist_16g (2026-06-26)

  • Result: complete — 2-5-9 (40.6%), perf ≈ 2634
  • Opponent: SF UCI_Elo=2700, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2700_lmrhist_16g.txt

Gauntlet gate_sf2600_cm_16g (2026-06-26)

  • Result: complete — 4-6-6 (43.8%), perf ≈ 2556
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_cm_16g.txt

Gauntlet gate_sf2600_tt2_16g (2026-06-26)

  • Result: complete — 4-5-7 (46.9%), perf ≈ 2578
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_tt2_16g.txt

Gauntlet gate_sf2600_policy_16g (2026-06-26)

  • Result: complete — 1-11-4 (18.8%), perf ≈ 2345
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_policy_16g.txt
  • Note: policy v1 with move-ordering bug (TT/killer quiets demoted into policy band); invalid for strength claims

Gauntlet gate_sf2600_policy_fix_16g (2026-06-26)

  • Result: complete — 7-3-6 (62.5%), perf ≈ 2689
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_policy_fix_16g.txt

Gauntlet gate_sf2700_policy_fix_16g (2026-06-26)

  • Result: complete — 2-11-3 (21.9%), perf ≈ 2479
  • Opponent: SF UCI_Elo=2700, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2700_policy_fix_16g.txt

Gauntlet gate_sf2600_policy_soft_16g (2026-06-26)

  • Result: complete — 9-4-3 (65.6%), perf ≈ 2712
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_policy_soft_16g.txt

Gauntlet gate_sf2700_policy_soft_16g (2026-06-26)

  • Result: complete — 2-7-7 (34.4%), perf ≈ 2588
  • Opponent: SF UCI_Elo=2700, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2700_policy_soft_16g.txt
  • Note: below 8/16 bar; policy sprint paused, keep policy off by default

Policy sprint status (2026-06-26)

  • Policy ordering infrastructure kept on codex/policy-train-v1; off by default.
  • Soft gates are diagnostic only (SF2600 9-4-3, SF2700 2-7-7); no further policy retrains.
  • data/policy/policy.lzp is experimental — not recommended for play.
  • Next bet: residual NNUE on codex/residual-nnue-v1 (classical + scaled correction).

Residual NNUE v1 infra (2026-06-26)

  • Branch codex/residual-nnue-v1: NnueMode replace/residual via LABZERO_NNUE_MODE, scale via LABZERO_NNUE_SCALE (default 50).
  • search_eval: classical-only TLS for data gen; residual blend classical + clamp(net,±300)*scale/100.
  • CLI: clasieval, nnueforward, residualdata (10k games play d4 label d8, resumable).
  • Scripts: host-nnue-residual-train.py, host-sprint-residual-nnue.sh, residual parity in host-nnue-verify.sh.
  • Verify: cargo test 59 pass, podman smokes pass, replace + residual parity OK.
  • NPS (movetime 3s): off ~1.69M; replace ~771k; residual ~724k (~6% slower than replace; any loaded NNUE ~2× slower vs classical).
  • Data gen in tmux resnnue → data/nnue/residual.txt; train + gates pending completion.

Gauntlet gate_sf2600_resnnue_s50_16g (2026-06-27)

  • Result: complete — 4-8-4 (37.5%), perf ≈ 2511
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_resnnue_s50_16g.txt

Repetition And Progress Patch v1 (2026-06-27)

  • v0.6.2 draw fix (on codex/v0.6.2-draw-fix): search returns draw 0 on is_draw() before TT cutoff in negamax/qsearch; board.repetition_count_current(); UCI clears stop at each go.
  • Root rank helpers land in search.rs (tested) but are not wired in search_root — enabling them caused UCI opening blunders; follow-up branch.
  • Gate passed: gate_sf2600_v062_draw_16g — 8-6-2 (56.2%, 9 W-equiv, perf ≈2644), illegal=0.
  • INVALID gates (do not record in superhuman-band):
    • gate_sf2600_rep_patch_16g — interrupted 0-11-2 @ 13/16; move-1 blunders (h3/g5); polluted by policy branch + broken rep patch.
    • benchmark_rep_patch_16g.log — 0-15-0 (first run had LABZERO_NNUE leak; restart also failed).
    • benchmark_root_rank_32g.run.log / benchmark_root_rank_fix_32g — wired root rank; 0-16 then 0-3-1; reset to v0.6.2.
    • gate_candidate_sf2500_wtime_16g — interrupted @ 1-5-3/9; real-clock miss vs ≥8/16 bar.

Root-Rank Wiring RCA: tie-break regression (2026-06-28)

  • Symptom: wiring pick_root_move into search_root made the engine play move-1 flank junk (a2a4, h2h4) and lose 0-16 vs SF2600 (benchmark_root_rank_32g.run.log, INVALID), despite all unit tests + the single-thread startpos_depth_twelve test passing.
  • Root cause: old search_root selected best with if score > best_score (keeps the first / best-ordered move on ties). pick_root_move used max_by_key/max_by, which return the last element on ties. Aspiration narrowing makes trailing root moves fail low and return fail-soft bounds that tie the PV score, so last-on-tie handed the pick to a late, badly-ordered move. At startpos root_static≈0 < ROOT_AHEAD_THRESHOLD, so ranking is bypassed and only this tie-break mattered.
  • Fix: pick_root_move now keeps the first candidate on ties in both the raw-best and ranked paths (strict > replacement), matching the original semantics. Post-fix UCI: startpos → e2e4, e7e5 reply; 60 tests pass; clippy clean.
  • Follow-up run INVALID: benchmark_root_rank_fix_32g — stopped @ 0-3-1/4; openings OK but middlegame king-side collapses when ahead (g3?? hxg5). Likely ROOT_PROGRESS_BONUS when root_static >= 150. Reset to v0.6.2 tag (root rank unwired); wired ranking deferred to codex/root-rank-v2.

Gauntlet gate_sf2600_v062_draw_16g (2026-06-27)

  • Result: complete — 8-6-2 (56.2%), perf ≈ 2644
  • Opponent: SF UCI_Elo=2600, TC 3+2, 16 games
  • Artifact: docs/strength/gate_sf2600_v062_draw_16g.txt
  • Release: tagged v0.6.2 on main after merge.

v0.6.2 stabilization sprint (2026-06-28)

  • Harness: TC_MODE=wtime real decreasing clocks; freshclock preserves old synthetic protocol; DEBUG_MOVES TSV telemetry; MAX_PLIES truncation; SIGINT → status: interrupted footer; scripts/host-benchmark-analyze.py.
  • Search (v3 root rank, wired): first-on-tie pick_root_move; immediate-draw penalty even when not ahead; repeat/progress bonuses only when root_static >= 150; progress = quiet passed-pawn pushes only; lichess_draw_fen_avoids_queen_shuffle test.
  • Gold standard: SF@2600 32g wtime real-clock (production); legacy gate_sf2600_idtime_32g relabeled freshclock synthetic in superhuman-band.
  • Obsolete branch: codex/root-rank-v2 (progress bonus disabled) superseded by v3 on main.
  • No v0.6.2 retag: real-clock confirm gates below ≥8/16 bar; harness + root-rank v3 remain on main @ 6b390cd.
  • v0.6.2 retag (2026-06-28): public ≈2600/≈2400 headline withdrawn; legacy gauntlet = freshclock synthetic; real-clock INVALID rows tracked in superhuman-band + ladder.

Gate results (wtime real-clock)

  • Smoke PASS: benchmark_20260628T080455Z — SF2400 4g → 1-0-3 (2.5/4 W-equiv); 0 illegal/errors; 0 threefold; openings OK.
  • Candidate INVALID: gate_candidate_sf2500_wtime_16g — stopped @ 1-5-3/9 (3.5/9 W-equiv); cannot reach ≥8/16 bar; do not run SF2600 gold.
  • SF2400 16g INVALID: gate_sf2400_wtime_16g — stopped @ 0-4-1/5 (0.5/5 W-equiv); real-clock collapse at same elo smoke was 1-0-3/4.

Build/NPS Sprint (2026-06-28)

  • Default release profile: workspace Cargo.toml now uses thin LTO, one codegen unit, and panic abort. Package-local release profiles are avoided because Cargo ignores them inside this workspace.
  • Native opt-in: LABZERO_NATIVE=1 ./scripts/build-host-engine.sh adds -C target-cpu=native for local host binaries only.
  • NPS tool: scripts/host-nps-bench.sh runs a fixed FEN suite with NNUE/policy env cleared and LABZERO_ROOT_POLICY=raw, recording TSV artifacts under docs/perf/.
  • A/B evidence: generic vs optimized release at same commit:
    • nps_20260628T174048Z vs nps_20260628T174058Z: Threads=1 depth=8 optimized median NPS ≈ +30%.
    • nps_20260628T174330Z vs nps_20260628T174338Z: Threads=4 depth=9 optimized median NPS ≈ +3.4% with stable bestmoves.
  • Decision: keep optimized profile as default production polish. Do not present it as a strength claim; use host benchmarks for strength and NPS probes for speed only.