Basilisk Development Workflow Guide

How to drive the improvement plan with an AI coding model and know what to run, what to report, and when a decision is yours. Read alongside PLAN.md, which holds the technical rationale.

Current Checkpoint

CURRENT STATE (2026-07-08): Releases 1.5.0 / 1.6.0 / 1.7.0 / 🎉 1.8.0 RELEASED 2026-07-08 (Phases 5+6+7 bundled), bench 12,661,251. 1.8.0 ≈ +93 Elo over 1.7.0 at 3+0.03, and an LTC field gauntlet (colosseum 10+0.1) confirmed it holds at long TC — ~+40 Elo head-to-head vs 1.7.0 (gains compress at LTC, as expected for eval refits). Phase breakdown: Phase 5 (TM) clock-at-go +2.95; Phase 6 (search) 6.1 TT-bound eval +7.18 → 6.10 bundle +9.14; Phase 7 (eval) 7.1 SF@60k distill +6.75 + five on-policy self-play cycles (+21.02, +19.51, +18.29, +15.32) ≈ +80.9. 7.4 histshape SPSA was SKIPPED (low EV — family SPSA 0-for-2 at maturity, LTC-compresses to ~+2); **7.5 = the LTC gauntlet

release.** ⚠️ Key lesson from Phase 7: holdout-MSE-delta badly under-predicts Elo for on-policy self-play refits — the linear eval barely moves each cycle yet gains ~+15-20 (the king-safety funnel keeps re-shaping); trust the SPRT, and expect ~half the fast-TC gain to survive at LTC. ▶ NEXT (post-1.8.0): Phase 8 (feature menu, EV≈0, skip unless evidence) or Phase 9 (NNUE, the ceiling-breaker); another on-policy self-play cycle is still mildly positive (~+7 LTC) if a 1.9.0 is wanted. The bullets that follow are the Phase-1 historical record.

Phase 0 harness is complete.
SPSA and SPRT now share the representative clock TC: tc=3+0.03. Use .\tools\sprt.ps1 -TC "10+0.1" for LTC confirmation and -MoveTime 0.1 only for the optional old fixed-movetime sanity check.
Phase 1 search-parameter plumbing is complete.
SearchParams exists and is exposed as tune-gated UCI options with -DTUNE=ON.
Default-equivalence has been recorded: bench 13 = 4,972,548 nodes, full CTest passed.
Phase 1 pruning SPSA was SPRT-accepted: +18.87 +/- 8.81 Elo, 2930 games, H1 accepted.
Phase 1 LMR SPSA completed: 4034 iterations, 129,088 games.
Phase 1 LMR SPRT was accepted: +15.63 +/- 8.02 Elo, 3714 games, H1 accepted.
Narrowed combined Phase 1 polish was rejected: -0.40 +/- 3.20 Elo, 23,210 games, H0 accepted.
Harness-corrected retry audit: the combined-polish retry folds into the relocated Step 7.4 histshape SPSA (after the eval refit); the post-LMR deeper re-search became 6.4 (done, exposure-only).
Phase 1 validation completed:
- vs Basilisk 1.4.9/defaults: 2000 games, 53.90%, approx +27.16 Elo.
Phase 1 is complete. The accepted search-constant head is tools\test_engines\basilisk-phase1-final-pext-pgo.exe (shipped as 1.5.0).
A 2026-06 analysis re-baselined the plan: the depth deficit vs Stockfish is eval accuracy + search selectivity, not nps and not time management. PLAN.md Phases 2-7 were rewritten accordingly (Texel pipeline, search efficiency wave, eval features, time management, feature menu, NNUE last).
Step 2.0 is complete. src/EvalParams.h created: struct EvalParams holds all ~900 tunable eval weights with defaults identical to the old inline constants; EVAL_PARAM_LIST X-macro enumerates every group. eval.cpp updated to read from g_eval_params; init_eval_tables takes const EvalParams&. Bench fingerprint on release-pext (non-PGO) on this machine: 4,283,684 nodes (the recorded 4,972,548 was the PGO build; 4,283,684 is the correct non-PGO baseline for refactor verification). 8/8 CTest passed.
Step 2.1 is complete. Under BASILISK_TUNE: load_eval_file_if_set() reads BASILISK_EVAL_FILE env-var path on startup; run_dumpeval() prints all 900 parameters in name index value format. dumpeval console command added to UciLoop. Round-trip verified byte-identical (dump → load → dump). Release build ignores both; bench fingerprint unchanged: 4,283,684 nodes, 8/8 CTest passed.
Step 2.2 is complete. CMake TEXEL=ON option added; basilisk-texel target built from src/eval.cpp (TEXEL_TRACE+BASILISK_TUNE) + tools/texel/tuner.cpp. EvalTrace struct holds int16_t mg[900] and eg[900] flat count arrays; reconstruct() computes tapered dot product. All 900 linear eval parameters instrumented with TR_MG/TR_EG/TR_BOTH macros; pawn cache bypassed under TEXEL_TRACE; king safety traced as one-hot SafetyTable lookup. Acceptance test: basilisk-texel --verify reconstructs exactly on all positions (OCB/50-move/mate-drive absorbed into rest). Release bench 4,283,684 nodes, 8/8 CTest passed.
Step 2.3 is complete. tools/datagen.ps1 and tools/texel/extract.py created. datagen.ps1 runs fastchess self-play (nodes=8000, 30k rounds = 60k games) writing tools/texel/data/selfplay.pgn. extract.py uses python-chess to skip first 16 plies + last 6 plies, drops in-check and capture/promotion moves, samples ≤12 positions per game, deduplicates by FEN (4 fields), splits by game (5% holdout), writes train.csv / holdout.csv. Round-trip pipeline verified: extracted positions pass basilisk-texel --verify.
Where the plan stands (restructured 2026-06-18, PLAN.md §0.5): Phase 2 texel infra (2.0–2.3) and the cheap, structure-independent scalar fits are done/accepted (material +29, mobility +8.8, passers +16.6, pawn-structure +30.7, hanging baked; rooks rejected). The plan was reordered so eval structure is built before the heavy eval tuning and the search SPSA. The remaining old 2.4 tuning (king safety 2.4c, PSTs 2.4d, polish 2.4e, plus the untuned pawn-threats) is deferred into the new Phase 4.
1.6.0 shipped (commit Version 1.6.0); the repo is on branch v1.7.0.
Phase 3 (eval structure build-out, Steps 3.0–3.11) is ✅ COMPLETE (2026-06-24…27). The whole enlarged HCE structure is in place, seeded inert and Texel-traceable; bench fingerprint is now 4,168,590 (re-baselined by the two behaviour-changing steps). Per-step gates were green throughout (bench identity for the inert steps, 9/9 CTest incl. test_endgames 18/18, texel --verify exact 8,598/8,598, test_eval symmetry).
- 3.0–3.4, 3.6–3.9 behaviour-identical (seeded-inert structure; bench unchanged at each step): attack maps + pins, threats, king-safety v2 danger model, per-count mobility tables, pawn/passer refinement, space + winnable coupling, small positional terms, material imbalance, HCE survey adds.
- 3.5 (endgame knowledge) behaviour-changing — exact KPK bitbase, KBNK corner mate, KNNK/no-pawn-minor/OCB draw scaling. Bench → 4,377,437. SPRT −4.17 ± 5.20 (KEPT) = the per-node NPS tax (EPD-proven correct), not a chess bug; recovered by 3.10 + 3.11 below.
- 3.10 apply_endgame guard (behaviour-identical): skip the endgame census in the opening/middlegame — removes the cost behind 3.5's −4.
- 3.11 lazy eval (behaviour-changing): skip the heavy positional block when the material/PST margin is decisive (LAZY_MARGIN=700; apply_endgame runs on both paths so KBNK/KPK survive). Bench → 4,168,590. SPRT +16.64 ± 7.03 Elo, LOS 100%, H1 (LazyOn vs LazyOff) — a clear net gain that recovers the whole Phase-3 NPS tax and then some.
The next implementation work is Phase 4 — the eval data-fit campaign (+80–160 Elo), starting at Step 4.0 (build the finite-difference --tune-kingsafety re-eval path — the biggest lever — plus the feature-support / bucketed-holdout tuner-readiness gate). None of the Phase-3 structure is tuned yet; Phase 4 is where it activates. Ships as release 1.7.0.

The program in one table (overview · model picker · Elo)

Read PLAN.md §0.5 for why this order: build all eval structure first (Phase 3, no games), fit the eval once (Phase 4), harden time management (Phase 5), then run the search SPSA once last (Phase 6). Texel fits are cheap; SPSA/SPRT games are the conserved resource.

Phase	What	Status	Model(s)	Elo
1–3	search SPSA, cheap eval scalars, robustness, eval structure	done (1.5.0/1.6.0)	(historical)	+27 / +54 / +16.6
4 Eval data-fit campaign	staged Texel refit over the full structure	done (1.7.0)	Sonnet 4.6 medium (driving)	+280.74 vs phase1-final
5 Time-management hardening	clock-at-`go` fix; 5.8 TM SPSA washed → reverted	done	Opus 4.8 medium / Sonnet 4.6 medium	~+3
6 Search-efficiency wave	6.1 TT-bound eval; 6.2–6.8 mostly washed/inert; SPSA → 7.4	✅ COMPLETE (6.10 +9.14 H1)	Sonnet 5 med; interaction-risk ports Fable 5 med-high (alt Opus 4.8)	+9.14 (bundle)
7 Eval-refresh cycles + one search tune	SF distill +6.75, ×5 self-play cycles (+21.02/+19.51/+18.29/+15.32); 7.4 SPSA skipped; 7.5 LTC gauntlet	✅ DONE — 1.8.0 released 2026-07-08	Fable 5 med (alt Opus 4.8) + Sonnet 5 med	~+93 fast / ~+40 LTC vs 1.7.0
8 Feature menu	demoted — audit found nothing to add	skip unless evidence	—	~0
9 NNUE	the ceiling-breaker	after Phase 7 flattens	Fable 5 high (alt Opus 4.8 high)	+200…+400

Per-step models + Elo are on each PLAN step header and in the Phase-7 EV table. SPRT is the only verdict.

Gauntlet validation (2026-06-19, 35k games @ tc=3+0.03): Basilisk 1.5.1 (dev, partial scalar tuning only) was 2nd of 9, +54 over 1.5.0, tied with "SF-capped-2700" — the eval lever is confirmed, with KS/PST/structure still to come. Two cautions: (1) SF UCI_Elo is calibrated for 120s+1s/CCRL-40/4, so it is not a true anchor at this fast TC — for a real CCRL number run a slower-TC gauntlet that includes Critter 1.6a (it forfeited at 3+0.03). (2) 1.5.1 lost 65 games on time (vs 1.5.0's 18) — add a hard time-safety floor to compute_time_limit now; forfeits contaminate every gauntlet.

So if you say "implement the next step," the model should know (2026-07-04): 1.7.0 has shipped (Phase 4 eval, +280), the dev head is 1.8.0 (bench 13,495,492), and Phases 5 and 6 are both COMPLETE (Phase 6 closed at 6.10 = +9.14 ± 4.62, H1). Phase 7 is the active phase and its on-policy self-play cycles keep paying ~+20 each: 7.1 +6.75 (b8c469b), 7.2 +21.02 (d113471), 7.3 +19.51 (9741070) — all H1, cumulative ≈ +47, curve not flattening. The next step is another self-play cycle from the 7.3 head, same pipeline each time: datagen.ps1 with -Book tools\texel\data\beast_seed.epd -BookFormat epd → extract_parallel.py --balance-phase 1.5 → --tune all + sequential --tune-kingsafety (bake tune-all → rebuild basilisk-texel → re-fit KS → bake KS) → build_test → SPRT vs the prior head. Stop when a cycle yields <~+8 (trust the SPRT — the holdout MSE stalls long before the Elo does).

The Basic Rhythm

Most work is a short ping-pong:

You   -> "Implement next step of the plan."
Model -> Reads PLAN.md, inspects current state, makes only needed edits, and
         tells you exactly what to run.
You   -> Run the command and paste the short result.
Model -> Acts on the result: keep, revert, rerun, or move to the next gate.

For SPSA and SPRT, the model cannot honestly guess the result. Your report from the long-running command is the decision input.

Next Commands

Phase 2.9.2 — done (2026-06-19). tools/gauntlet.ps1 now has a -TC/ -MoveTime/-TimeMargin parameter set mirroring sprt.ps1: clock tc=10+0.1 by default (the phase-boundary condition from §10), fixed st=… only when -MoveTime is given, and the console banner reports the real TC used. Pure tooling, no engine change.

Phase 5 Step 5.1 / 5.1b (TM foundation; old labels 6.1/6.1b) — done (2026-06-20). A first LittleBlitzer probe (521 games) of the original Step 2.9.1 patch (a reserve clamp bolted onto the old tiered-percentage formula) showed an oh1/oh10 tpm gap and a lone forfeit on the aggressive Move Overhead=1 variant, traced to the old formula being tighter to the margin than Rarog's already-rewritten SF-style one. Rather than re-tune that formula's margins, compute_time_limit's clock path is now a direct port of Rarog's Phase 2.2 Stockfish-style rewrite (D:\code\rarog\src\time_manager.rs) — logT-based optConst/maxConst, ply-aware sudden-death and explicit-movestogo branches, same constants, no new BASILISK_TUNE knobs (matching Rarog, which exposes none either). The 2.9.1 reserve is folded in using Rarog's exact mechanics: hard_ceiling = raw_time - 2*overhead (Basilisk's original patch used 3x effective margin from double overhead subtraction; now it's 2x, identical to Rarog). compute_time_limit also gained a game_ply parameter, threaded from both call sites (Searcher::search and the ponderhit branch of check_stop).

bench 13 unchanged (4,033,379 nodes, non-PGO release-pext — TM doesn't touch fixed-depth bench), 8/8 CTest passed. Manually verified wtime from 60000 down to 0 (winc=30): smooth degradation, no hangs; at the gauntlet TC (wtime=3000 winc=30) budgets land ~68-78ms, in the same range as the whole opponent field's tpm (65.6-70.0) from the prior LittleBlitzer snapshot.

LittleBlitzer validation — PASSED (2026-06-20). Full 35,000-game overnight pool at tc=3+0.03, default Move Overhead: t=0 for Basilisk 1.5.1-sftm (and every other engine in the pool — confirms the harness itself is sound, not just this engine). Rating 2696.8, 62.7% score, 2nd of 5 — essentially tied with Stockfish-2700-capped, well clear of Rarog 2.1.0 (2620.0) and SF-2600/2500. This conclusively answers the time-safety question; no remaining doubt that the Step 6.1/6.1b fix works.

Skippable, not recommended: a second-harness forfeit check. Step 6.1/6.1b is a forfeit fix, not a strength change — its true Elo delta vs phase1-final is expected to sit near 0, which is the worst case for an SPRT (elo0=0 elo1=3 converges slowest exactly when the true value sits on the H0 boundary; it can run tens of thousands of games without crossing either bound — do not use sprt.ps1 here). If you still want a second-harness sanity check per Rarog finding #3 (LB can mismeasure some engines), use gauntlet.ps1 (fixed games, no stopping rule) and read the t= count only — ignore the score:

.\tools\gauntlet.ps1 `
    -Engine tools\test_engines\basilisk-phase6-sf-tm-pext-pgo.exe `
    -Opponents tools\test_engines\basilisk-phase1-final-pext-pgo.exe `
    -Name Phase6SfTm -TC "10+0.1" -Games 1000

Given the LB result already showed t=0 for every engine in the pool (not just Basilisk, so it's not a Basilisk-specific harness blind spot), this adds little. Skip straight to the release unless you want extra peace of mind.

Then 2.9.3 (release 1.6.0). Then the eval work begins with Phase 3.0 — the attack-map substrate (PLAN.md §6): a behaviour-identical refactor that computes attacked_by[color][pt], attacked[color], attacked2[color] once per evaluate() and routes mobility, king safety, and hanging through them. No games — the gate is the bench fingerprint.

Ask the model:

"Build a PGO test binary for the time-safety fix and run a clock-TC gauntlet vs phase1-final to confirm time forfeits drop to ~0."

Then verify nothing changed:

# Rebuild and check the refactor is behaviour-preserving.
cmake --build build\release-pext --target basilisk
.\build\release-pext\basilisk.exe   # then: bench 13   (expect the recorded baseline)
# And keep --verify exact:
cmake --build build\release-pext --target basilisk-texel
.\build\release-pext\basilisk-texel.exe --verify tools\texel\data\holdout.csv

Most of Phase 3 is the same shape: implement a structural step with new sub-terms seeded inert → confirm bench 13 unchanged + --verify exact + CTest → move on. The games (SPRT) only start in Phase 4, when the staged Texel campaign activates the new terms.

The texel toolchain and datasets already exist (Phase 2): tools/datagen.ps1, tools/texel/{sample_fens,extract,import_beast}.py, the basilisk-texel target, and train.csv/holdout.csv (self-play) plus beast_sf_*.csv (Stockfish-labelled). Reuse them in Phase 4 — no new data needed unless holdout loss stalls.

This creates about 1.52M train / 80k holdout Stockfish-target positions in FEN;target format. The importer treats the raw Stockfish WDL target as side-to-move perspective and converts black-to-move rows to white perspective, which is what basilisk-texel trains against. Keep it separate from the self-play data when reporting results.

After the material tune finishes, inspect the printed material deltas before baking them into EvalParams.h; plausible values still need SPRT before they are accepted.

For the full Beast import (beast_sf_all_train.csv / beast_sf_all_holdout.csv), add --max-positions 0 to the tune command. Without it, basilisk-texel intentionally loads only 5M positions from each split to protect later, wider tune groups from accidental huge memory use.

After the failed SF-supervised material SPRT, use Beast FENs as start positions only and let Basilisk self-play generate the labels:

python tools\texel\sample_fens.py A:\Chess\Beast\data\txt\positions.txt `
  --out tools\texel\data\beast_seed.epd --count 100000

.\tools\datagen.ps1 -Suffix phase2-base `
  -Book tools\texel\data\beast_seed.epd -BookFormat epd `
  -OutputPgn tools\texel\data\selfplay_beast_seed.pgn `
  -Rounds 75000 -Nodes 8000

python tools\texel\extract.py tools\texel\data\selfplay_beast_seed.pgn `
  --train train_beast_seed.csv --holdout holdout_beast_seed.csv `
  --skip-start 0 --max-per-game 12

The Phase 2 scalar tuning is complete (§0.5): material, mobility, passers, pawn-structure, and hanging are baked into EvalParams.h; rooks was rejected. Do not continue the old 2.4 subgroups (king safety, PSTs, threats, minors) — they are deferred into the Phase 4 campaign that runs after the Phase 3 structure exists.

For a fresh clone, the historical Phase 1 commands are:

# One-time setup on a fresh clone
.\tools\setup_tools.ps1

# Build the original default tune baseline, only needed for pruning from scratch
.\tools\build_test.ps1 -Suffix phase1-defaults

# Configure pruning SPSA
.\tools\setup_spsa.ps1 -ConfigGroup pruning -Iterations 5000

# Run the tuner
cd tools\weather-factory
python main.py

Stop SPSA with Ctrl-C when values look stable or the planned run is complete. State is saved under tools\weather-factory\tuner\state.json, so rerunning python main.py resumes. Running setup_spsa.ps1 again starts a fresh run and archives old tuner state unless you pass -Resume.

After pruning is accepted, build the current accepted head as the LMR parent and run LMR SPSA from that engine:

cd ..\..
.\tools\build_test.ps1 -Suffix phase1-lmr-baseline
.\tools\setup_spsa.ps1 -ConfigGroup lmr -EngineSuffix phase1-lmr-baseline -Iterations 5000
cd tools\weather-factory
python main.py

The narrowed combined polish already ran and failed under the old split harness. Do not run it as a Phase 1 continuation. Its one justified retry folds into the relocated Step 7.4 histshape SPSA, after the eval refit, using the unified clock harness:

cd ..\..
.\tools\setup_spsa.ps1 -ConfigGroup combined -EngineSuffix phase1-lmr -Iterations 2000
cd tools\weather-factory
python main.py

Treat that command as a historical reference; the current search-tune config is config_histshape.json (Step 7.4) — see tools/spsa_configs/README.md.

What To Report Back

SPSA Result

Minimal:

"Pruning SPSA stopped at 5,000 iterations. Final values: RfpCoeff=128, RazorCoeff=276, NullEvalDiv=180, AspirationDelta=21, ..."

Helpful extras:

"HistPruneCoeff sat at the max for the last 1,000 iterations."

The model will decide whether to bake the values, rerun with adjusted ranges, or discard the noisy group.

SPRT Result

Minimal:

"SPRT: H1 accepted after 1,840 games."

or:

"SPRT: H0 accepted after 2,210 games."

Helpful extras:

"Score 53.1%, LLR crossed +2.94."

H1 usually means keep the candidate. H0 usually means revert or retry once if the SPSA run was obviously flawed.

Bench Result

For pure refactors:

"bench 13 = 4,972,548 nodes."

For tuned candidates:

"bench 13 = 4,812,903 nodes."

A changed bench fingerprint is expected after tuning. It is a behavior fingerprint, not an Elo score.

Texel / Tuner Result (Phase 4)

Minimal:

"Threats fit: train loss 0.0981 → 0.0972, holdout 0.0986 → 0.0979."

Helpful extras (these are what the model acts on — loss alone is never the verdict, SPRT is):

"Bucket losses: all held or improved except king-attack (0.094 → 0.097); feature-support flagged threat_by_minor[Q] with only 31 observations; the new safe-check unit weights fit with plausible signs."

A regressing bucket, a sparse-feature warning, or an implausible sign/shape is a stop-and-investigate signal before spending SPRT games, not after.

Errors

Paste the important error line:

"fastchess exited with: engine option RfpCoeff not found."

or:

"bench 13 returned 0 nodes; engine crashed on startup."

The model can diagnose from that.

Decision Points

Situation	Usual decision
SPSA values are stable and plausible	Bake candidate values, build, test, SPRT
SPSA values are noisy	Run longer or reduce the group/ranges
One plausible value hits a boundary	Widen that range once and rerun
Many values hit boundaries	Treat the run as suspect; narrow the problem
SPRT accepts H1	Keep, record, move to next group
SPRT accepts H0	Revert; retry once only if the run was flawed
End of Phase 1	Run fixed-game validation before release work
Phase 2 asks for Texel data	Build dataset and holdout before tuning

Do not keep running repeated SPRTs against tiny changes until one passes. That is just statistical fishing wearing a little hat.

Phase Progress Tracker

Update this when work completes.

Phase 0 - Harness

tools/setup_tools.ps1 exists for fastchess/weather-factory setup.
tools/sprt.ps1 exists and uses repo-local book/default paths.
tools/gauntlet.ps1 exists for fixed-game phase-boundary validation.
tools/build_test.ps1 builds pext PGO tune binaries into tools\test_engines.
tools/setup_spsa.ps1 writes pruning/LMR weather-factory configs.
Baseline bench 13 fingerprint recorded: 4,972,548 nodes.
Calibration self-vs-self SPRT skipped intentionally.

Phase 1 - Search Constants

Phase 1 is complete. Keep phase1-final as the accepted head.

Phase 2 - Eval Tuning (Texel pipeline; PLAN.md Section 4)

Full per-candidate record (commands, intermediate holdout losses, build fingerprints, PGN paths) lives in PLAN.md §4 — this is the outcome summary.

2.0–2.3 infra: EvalParams (bench-identical, 4,283,684 non-PGO baseline); tune-only loader + dumpeval (900 params, round-trip exact, hidden in release); basilisk-texel (TEXEL=ON, 900-param trace, --verify exact incl. frozen-term edges); datagen.ps1 + extract.py; Beast/SF import verified (1,520,109 train / 79,891 holdout).
2.4a material — SF-WDL labels rejected (-16.56 ± 9.56, 2,982 games), so pivoted to Basilisk-labeled Beast-seed self-play; that candidate accepted +29.05 ± 11.36 (1,930 games). Baked mg_val={85,323,363,514,1085}, eg_val={96,310,319,557,996}.

2.4b — tuner gained scalar subgroups with sign/shape clamps + best-epoch restore (the broad all-scalar pass was not baked: implausible signs/shapes). Per-subgroup SPRT outcomes:

Subgroup	SPRT verdict (games)	Baked values
mobility	accepted `+8.77 ± 5.68` (7,288)	`mob_mg={0,0,5,5,1,2,0}`, `mob_eg={0,0,5,7,7,12,0}`
passers	accepted `+16.57 ± 8.28` (3,462)	`passed_mg[1..6]={8,8,8,25,90,90}`, `passed_eg[1..6]={10,10,53,84,106,124}`, support/cand/free-EG ≈0, `pass_safe_eg=16`, `prox_base=11`
pawnstruct	accepted `+30.74 ± 11.76` (1,836)	`doubled=(-2,-8)`, `isolated=(-9,-17)`, `connected=(18,1)`, `backward=(0,-12)`
hanging	baked, SPRT pending	`hang_pen={0,0,22,39,33,36,0}`
rooks	rejected `+3.13 ± 4.74` (10,666; inconclusive, awkward shape) — reverted	—

2.4b cheap structure-independent scalars done (mobility/passers/pawnstruct/hanging accepted; rooks rejected). Stop 2.4 here (§0.5).
2.4b threats/minors pawn-threat scalars — DEFERRED to Phase 4.2/4.5 (Phase 3 rewrites the threats term).
2.4c King safety block — DEFERRED to Phase 4.3 (Phase 3.2 rewrites king safety first; do not tune the toy model).
2.4d PSTs (+material refit) — DEFERRED to Phase 4.7.
2.4e Global polish — DEFERRED to Phase 4.8.
2.5 2000-game validation vs phase1-final — moved to the Phase 4 boundary.

Phase 2.9 - Robustness Quick Win (DONE — shipped in 1.6.0; PLAN.md §4.9)

2.9.2 Add -TC/-MoveTime/-TimeMargin to tools/gauntlet.ps1 (clock tc=10+0.1 default, mirrors sprt.ps1; banner fixed). Pure tooling, no engine change, parse-verified.
2.9.1 (superseded by Phase 5 Step 5.1/5.1b, old labels 6.1/6.1b, 2026-06-20 — see below) — original standalone reserve patch on the old formula; replaced rather than re-tuned after a LittleBlitzer probe showed it binding more than Rarog's equivalent.
2.9.3 Release 1.6.0 prep done (2026-06-20) — version bumped in src/Constants.h + CMakeLists.txt, CHANGELOG.md entry added, clean non-tune release build verified (options hidden, bench 13 matches, 8/8 CTest), dist binaries built, committed. Pending (user-only): create tag v1.6.0 and push — the model never tags or pushes.
(calibration, anytime) Slower-TC gauntlet with a CCRL-rated anchor (Critter 1.6a / Fruit 2.1), pin with ordo -A "<name>" -a <ccrl> (PLAN.md §10).

Phase 3 - Eval Structure Build-Out (PLAN.md §6; EXECUTE FIRST — bench-identical, no games)

Phase 4 - Eval Data-Fit Completion (PLAN.md §6.5; the campaign, SPRT per stage)

Phase 5 - Time Management hardening + tuning (PLAN.md §7) — promoted to Phase 5, executed BEFORE the Phase 6 search wave (2026-06-29)

The 5.1/5.1b/5.2 foundation (Rarog-port formula + reserve + the 2026-06-20 LB validation) is done and stays. Reopened (steps 5.3–5.9) because LB still shows time-losses post-Phase-3/4 (heavier eval per node) and the TM constants were never tuned for Basilisk. Root cause (2026-06-29): (1) the clock starts inside the worker, missing go-receipt→search dispatch latency the GUI charges (likely the main LB forfeit); (2) fixed 2048-node poll overshoots the tiny bullet budget now that nodes are heavier; (3) thin static 2*overhead reserve

10 ms default Move Overhead; (4) zero TM-constant tuning. Aim: generally strong across bullet→slow, not a TC specialist.

PHASE 5 STATUS (2026-06-29). Diagnosis (5.3) overturned the premise: the engine's TM is sound — no overshoot (≤1 ms / 12k bullet moves), correct allocation; the LB forfeits are GUI-pipe latency, not an engine bug. So the robustness "repairs" were unnecessary: 5.5 dropped (no overshoot), 5.6 dropped (no real forfeit), 5.7 deferred (existing adaptive logic already extends on instability; a mid-iteration version needs an SPRT). Code-complete and committed: 5.3 (TM_Debug instrument), 5.4 (clock-at-go, bench-identical, SPRT-validated +2.95 ± 6.74, kept), 5.8 (9 TM knobs exposed for SPSA, behaviour-identical at defaults). One maintainer validation remains to close Phase 5:
~~5.4 non-regression SPRT~~ — DONE (+2.95 ± 6.74 at 4,468 games, stopped early as a confirmed non-regression).
5.8 TM-constant SPSA (the Elo lever, +8–25):
./tools/setup_spsa.ps1 -ConfigGroup tm -EngineSuffix phase5tm -Iterations 5000
cd tools\weather-factory ; python main.py
at tc=10+0.1; then re-validate the tuned set at 1+0.01 and 60+0.6, bake into SearchParams.h defaults, and SPRT the baked set vs phase5tm.
After 5.8 bakes + 5.9 cross-TC check → Phase 5 ships and Phase 6 (search wave) opens. Until then the Phase-5 head is behaviour-identical to 1.7.0 except 5.4.

Phase 6 - Search Efficiency Wave (PLAN.md §5; renumbered to Phase 6 on 2026-06-29 — executed AFTER Phase 5 time management. SPSA last — driving Sonnet 5 medium, dense/interaction-risk ports Fable 5 medium-high with an Opus 4.8 alternative — Fable 5 is available for a limited time)

6.1 TT-bound eval refinement — ACCEPTED +7.18 ± 4.05 Elo (H1, LOS 99.97%, 13,162 games, 2026-07-01). In negamax, a separate eval for pruning decisions only (RFP/razoring/NMP/futility) prefers the TT score when its bound proves it tighter than the corrected static eval (exact / fail-high above / fail-low below); ss->eval/static_eval stay raw so improving + correction-history are unaffected. Mirrored in qsearch stand-pat. Gain SPRT elo1=3 at 3+0.03 (phase61-ttbound vs phase5final) crossed H1 (LLR 2.96; Ptnml [318,1562,2639,1654,408]). Merged to development (commit 00370b8, the validated b600571). New head fingerprint 13,503,085 (was 13,466,664; the pruning cuts the hardest bench position hard — top-pos share 19.6% → 7.2%). Early SPRT trend was negative → recovered to +7.18: a reminder not to react to sub-3k-game noise. Follow-up: a mate-score guard on the refinement (branch commit 75650d6, house-style abs(tt_score) < MATE_SCORE - MAX_PLY, matching the existing ProbCut guard) is NOT in the validated head — bundle it into 6.2's SPRT as a non-regression rider, or apply as a rare-trigger correctness bugfix. — Sonnet 4.6 medium
[REJECTED] 6.2 History: 6-ply continuation table — REGRESSED −7.70 ± 4.77 Elo (H0, LOS 0.08%, 9,204 games, 2026-07-01). Added cont_hist6_ (ply-6 continuation, /3 weight) mirroring cont_hist4_, on branch phase6-6.2-history (commit 0f4d08d), bundling the 6.1 mate guard. Bench dropped 13,503,085 → 11,266,326 (~−16% nodes) — but the node cut came largely from the extra history magnitude feeding LMR/history-pruning (tuned without it) → over-pruning → clean regression. Dropped; NOT merged. Development stays at the validated 6.1 head. Key lesson (PLAN §0.5): cont_hist6 is a Rarog feature that gains for Rarog but hurts Basilisk — a direct confirmation that we must benchmark against SF and let SPRT decide, not copy the sibling. The bundled mate guard can't explain −7.7 (rare trigger); it survives on phase6-6.1-ttbound (75650d6) for a future non-regression rider. If ply-6 is ever revisited, do it via the wave2 SPSA (jointly with the divisors it disturbs), not as a hand-add. — Sonnet 4.6 medium

Remaining steps REWORKED 2026-07-01 from a line-by-line audit of SF master (snapshot 2026-06-30) + Ethereal 14.40 + Weiss 2025 (the HCE anti-bias check — PLAN §5 has the full audit + formulas). Basilisk's search base is audit-confirmed modern (singular w/ multicut+double+negative ext, IIR, ProbCut, LMP, history pruning, corr-hist, etc.); the steps below are the proven remaining refinements, ordered by evidence strength. Rules unchanged: one at a time, SPRT elo1=3 @ 3+0.03, knobs exposed at introduction, SPSA only in 6.9. Models: Fable 5 is limited-time — Opus 4.8 at the same thinking level is the standing alternative.

Pattern established by 6.3 + 6.4 (both landed as exposure-only, 2026-07-01): literature constants (SF's, Weiss's) chaotically break the KBNK/KQK mate-resolution CTests because Basilisk's other search constants weren't tuned for that scale. For 6.5–6.8: implement the mechanism, expose it, verify CTest passes at a provably-inert default (mathematically argued, not just empirically observed to pass), and defer the real value to 6.9 — do not spend cycles hand-picking a canary-passing constant.

6.3 History formula — DONE 2026-07-01 as behaviour-identical EXPOSURE (commit 18c5db5; constants deferred to 6.9). The planned Weiss-constant transplant (and SF's, and blends) chaotically destabilised the mate-resolution CTests (KBNK playout + KQK mate-in-5 flipped non-monotonically under consumer knobs) — any linear form is ~30× larger at shallow depth than our d² and the consumers (hist_prune_coeff, lmr_hist_div) were tuned for the small scale. Shipped instead: bonus = min(quad·d²/64 + lin·d, max) + mirrored malus knobs + hist_ttmove_bonus (SF-only, default 0) — 7 knobs, defaults reproduce the legacy formula EXACTLY (bench identical, 9/9 CTest, no SPRT owed). The asymmetric-linear shape both references prove is reachable by the 6.9 SPSA, tuned jointly with the consumers (HistPruneCoeff range widened [1000, 28000]). 6.1 mate guard merged alongside (correctness fix): dev-head bench fingerprint now 12,736,941 (was 13,503,085 — the guard's footprint concentrates in the suite's mate positions); guard validation rides in the 6.10 cumulative SPRT. — Fable 5 medium
6.4 Post-LMR deeper re-search + conthist + double-ext cap — DONE 2026-07-01 as EXPOSURE ONLY (commit 945ec8c; real values deferred to 6.9). Hit the SAME canary fragility as 6.3, twice: (1) the do_deeper/do_shallower depth adjustment — swept flat margins 40/80/150/300/600/1200 against the KBNK CTest, non-monotonic (only 300 passed) with no natural "off" point in a sane range → dropped entirely; (2) the post-LMR conthist nudge (Weiss form) at full weight broke the KQK mate-in-5 CTest → shipped behind a new PostLmrHistScale knob (percent, default 0, provably inert — hist_update with bonus=0 is a no-op); (3) the double-ext cap — Weiss's own seed (≤5) also independently broke KBNK → shipped as real infra (ss->double_exts, propagated via SearchStack) defaulting to 200, mathematically unreachable (double_exts ≤ MAX_PLY=128). Bench 12,736,941 unchanged (provable), 9/9 CTest (incl. both canaries). Extracted history_bonus_value/malus_value/update_cont_for_move as shared helpers (pure refactor). No SPRT owed. Real PostLmrHistScale/DoubleExtMax values are 6.9 SPSA material, jointly with LmrHistDiv/HistPruneCoeff. — Sonnet 5 medium
[REJECTED] 6.5 SEE-quiet + capture futility — DONE 2026-07-01, both EXPOSED-BUT-INERT (c3f1adc → 3d14867). Capture futility (eval + 200 + 200·lmr_depth + PieceValue[captured] + capHist/32 ≤ alpha) was canary-clean so it shipped active (bench 12,736,941 → 11,806,504) and got a real SPRT vs the 6.4 head: −2.78 ± 7.50 Elo, LOS 23%, LLR → H0 over 3.6k games — wash-to-tiny-loss → REVERTED to inert (CapFutDepth=0). SEE-quiet never went active: naive base-table-lmr_depth broke KBNK completely (no mate in 250 plies, every coeff 25–120 failed; SF's lmr_depth here includes the history term that protects precise endgame quiets) → inert (QuietSeeDepth=0). Net: no Elo, but both wired + exposed for 6.9 SPSA (cap_fut_depth~7 + margins; quiet_see_depth~8 after the history-aware lmr_depth fix). Dev stays at validated 6.4 head (12,736,941, 9/9 CTest). Lesson: canary-clean ≠ Elo gain — the games are the arbiter. — Opus 4.8 high
[SKIPPED] 6.6 Fail-low prior countermove bonus — skipped 2026-07-02 (user decision). Weakest evidence in the wave (SF-only) and, after 6.2–6.5's run of rejections/washes/inert-landings, decided to go straight to 6.8 rather than spend another implement/CTest/SPRT cycle on it. Revisit only if 6.9 shows spare headroom. — Sonnet 5 medium
6.7 Fractional LMR (1024ths) + lmr_tt_capture — DONE 2026-07-02, behaviour-identical infra (commit 5daae1c). lmr_table_ → 1024ths (int(1024·(base+ln·ln/div)), floor identity keeps base reduction identical); the 4 lmr_*_adj knobs → 1024ths (defaults ×1024, ranges 0..3072); reduction accumulated in 1024ths, >>10 at the end; new lmr_tt_capture seeded 0 (inert). History kept integer-quantised so it's behaviour-identical (no SPRT) rather than the spec's fractional-history rounding step (which the 6.2–6.5 run showed would just wash). Bench 12,736,941 unchanged, 9/9 CTest; verified live in a TUNE build (LmrNonPvAdj=1536 gives a node count distinct from 1024 and 2048 — sub-ply control). Gives 6.9 five fractional LMR dims + lmr_tt_capture. — Opus 4.8 medium
6.8 Qsearch quiet checks — DONE 2026-07-02, exposed-but-inert (commit 62648ca). At qply==0, after captures fail to cut off: try quiet checking moves (gen_quiet_checks) filtered by SEE≥0, capped at QsearchCheckCap. The seeded default (6, "4–6" per spec) broke the KBNK mate CTest — the sixth mechanism in a row (6.2–6.5) to trip this canary at its literature seed. Gated behind QsearchCheckCap defaulting to 0, provably inert (loop skipped when cap ≤ 0). Bench 12,736,941 unchanged, 9/9 CTest; confirmed live in a TUNE build. No SPRT owed. Real value → 6.9. Phase 6's feature list is now complete (6.1–6.8 resolved: one gain, one rejection, six exposure-only). — Sonnet 5 medium
[MOVED → 7.4] 6.9 wave2 joint SPSA — infrastructure built (commit 74bf5f0), the RUN relocated by the 2026-07-02 EV review (triggered by a 30h negative-gain Rarog SPSA). Why: (1) family SPSA base rate at maturity = 0-for-2 (5.8 TM wash; Rarog negative); (2) config_wave2.json's start point activates known-negative mechanisms (CapFutDepth=7 = the exact −2.78 config; QuietSeeDepth=6 = the KBNK-breaking mechanism) — a local optimizer starting 5–15 Elo down a hole in 24 noisy dims; (3) §0.5's own rule: search constants are cp-denominated → tune once, at the FINAL eval scale, and Phase 7 is about to refit the eval. Replacement = config_histshape.json (11 always-live dims from current defaults) run as Step 7.4 after the eval cycles. wave2 config + phase69-wave2base binary remain valid for a possible evidence-based wave3. — Sonnet 5 medium
6.10 Phase-6 boundary validation — ✅ DONE 2026-07-04, +9.14 ± 4.62 Elo (H1). Cumulative SPRT phase69-wave2base vs phase5final (elo1=3, 3+0.03): +9.14 ± 4.62, LOS 99.99%, 9,810 games, Ptnml [223,1149,1953,1307,273]. Validated the whole Phase-6 bundle (6.1 +7.18 + the un-SPRT'd mate guard + the inert 6.3–6.8 exposures). +9.14 ≥ 6.1's +7.18 → the mate guard is vindicated (its only behaviour change beyond 6.1; neutral-to-slightly-positive, no bisect needed). PHASE 6 COMPLETE. Cross-TC riders skipped (TC over-fit implausible; re-check only if 7.5's gauntlet flags it). Release deferred to 7.5 (1.8.0 = Phases 5+6+7). — Sonnet 5 low
[SUPERSEDED — old pre-rework acceptance stubs, no open work] These are the original Phase-6 acceptance-criteria stubs from before the 2026-07-02 EV-review renumbering; every one is covered by the reworked 6.3–6.10 entries above, so they are resolved / not separately necessary:
- old "6.5 Post-LMR deeper retry" + old "6.7 Double-extension cap" → folded into reworked 6.4 (exposure-only: do_deeper/do_shallower dropped as canary-non-monotonic; DoubleExtMax=200 shipped provably inert).
- old "6.6 Qsearch quiet checks" → reworked 6.8 (exposed-but-inert, QsearchCheckCap=0).
- old "6.8 Razoring restriction experiment" → not pursued — it wasn't on the SF/Ethereal/Weiss audit shortlist and its EV was ≈0; revisit only on new evidence.
- old "6.9 Second-wave SPSA" → relocated to Step 7.4 (the config_histshape.json run) by the EV review; see reworked 6.9 [MOVED → 7.4] above.

Phase 7 - Non-NNUE ceiling: eval-refresh cycles + the one search tune (PLAN.md §7.5; THE ACTIVE PHASE after 6.10 — reworked 2026-07-02 EV review, absorbs the relocated 6.9 SPSA as 7.4)

The EV table (decision aid — every step individually skippable; gains overlap):

Step Expected Elo (range / median) Cost Wash risk Model

7.1 SF-distill refit ✅ DELIVERED +6.75 ± 4.07 (est. +10…+35/~+20) fit = minutes; 1 SPRT 2h48 scale drift did NOT materialize Fable 5 med (alt Opus 4.8 med) pipeline; Sonnet 5 med fit

7.2 self-play refresh ✅ DELIVERED +21.02 ± 7.73 (est. +5…+25/~+12) datagen 45 min + fit + SPRT 49 min overperformed ~2× — 7.3 now high-EV Sonnet 5 med

7.3 cycle 2 ✅ DELIVERED +19.51 ± 7.04 (est. 0…+10/~+4 — under-called ~5×) same (~2h/cycle) curve not flat; iterate on (MSE-stall ≠ Elo-stall) Sonnet 5 med

7.4 histshape SPSA 0…+12 / ~+4 ~12–18h, abort @600 iters ~50% (5.8 wash + Rarog −30h precedent) Sonnet 5 med

7.5 boundary + 1.8.0 0 (banks the above) ~4–8h — Sonnet 5 low

(Phase 8 menu) ~0 — demoted, audit found nothing — — skip

(Phase 9 NNUE) +200…+400 months low Fable 5 high (alt Opus 4.8 high)

Step	Expected Elo (range / median)	Cost	Wash risk	Model
7.1 SF-distill refit ✅	DELIVERED +6.75 ± 4.07 (est. +10…+35/~+20)	fit = minutes; 1 SPRT 2h48	scale drift did NOT materialize	Fable 5 med (alt Opus 4.8 med) pipeline; Sonnet 5 med fit
7.2 self-play refresh ✅	DELIVERED +21.02 ± 7.73 (est. +5…+25/~+12)	datagen 45 min + fit + SPRT 49 min	overperformed ~2× — 7.3 now high-EV	Sonnet 5 med
7.3 cycle 2 ✅	DELIVERED +19.51 ± 7.04 (est. 0…+10/~+4 — under-called ~5×)	same (~2h/cycle)	curve not flat; iterate on (MSE-stall ≠ Elo-stall)	Sonnet 5 med
7.4 histshape SPSA	0…+12 / ~+4	~12–18h, abort @600 iters	~50% (5.8 wash + Rarog −30h precedent)	Sonnet 5 med
7.5 boundary + 1.8.0	0 (banks the above)	~4–8h	—	Sonnet 5 low
(Phase 8 menu)	~0 — demoted, audit found nothing	—	—	skip
(Phase 9 NNUE)	+200…+400	months	low	Fable 5 high (alt Opus 4.8 high)

7.0 Non-NNUE ceiling analysis read & stop-point understood. SF11 (~3440 CCRL) is the clean classical proof a complete HCE reaches the high-3000s; the 2026-07-01 SF/HCE audit (vs Ethereal 14.40 + Weiss 2025 + SF-classical) confirmed Basilisk's eval is feature-complete pre-NNUE — so Phase 7 is a label/data-quality grind, not new terms. Don't size off NNUE-era "HCE" ratings (Berserk/Rubi/Stash 3300+ are their NNUE builds, classical ~3000–3150). King-bucketed PSTs = NNUE input shape → do Phase 9 instead, not more HCE. — read-first
7.1 SF-distillation bootstrap — ✅ DONE 2026-07-04, ACCEPTED +6.75 ± 4.07 Elo (H1, LOS 99.94%, 16,066 games, Ptnml [623,1865,2834,1999,712]; SPRT 7.1-sfdistill vs 6.10-head, elo1=3 @ 3+0.03). Merged to development (b8c469b), bench 12,736,941 → 13,544,942. Re-centered the HCE against Stockfish @60k nodes (Hydra's annotate_sf.py on the quiet-filtered beast set), converting FEN;cp→FEN;wdl via 1/(1+10^(-cp/400)). Fit basilisk-texel --tune all on 2.0M/100k, L2=1e-6, K=1.372: holdout MSE 0.02226→0.02134, data phase-balanced 47/30/22 (vs v17's 59% endgame). The eval-scale risk did NOT materialize — the L2-to-prior anchor pinned material (Pawn 83/96, Knight 322/310, Rook 514/561, Queen 1085/998) and kept startpos at +50cp@d12; 55 eval lines nudged toward SF (tempo 15→20, rook_open 25→27, PSTs, imb/mob/threats ±1-2). 9/9 CTest — KBNK/KQK canary clean; one shallow-depth sanity bound widened 100→150cp (distilled eval reads +122@d5). KS funnel re-fit (--tune-kingsafety) DEFERRED as a follow-up rider (finite-diff too slow to finish inline; isolating the distillation gave a cleaner verdict) — rolls into 7.2's --tune-kingsafety on fresh data or runs standalone first. The 10+0.1 boundary gauntlet rolls into 7.5. — Fable 5 medium (alt: Opus 4.8 medium) pipeline; Sonnet 5 medium fit
7.2 On-policy self-play refresh cycle — ✅ DONE 2026-07-06, ACCEPTED +21.02 ± 7.73 Elo (H1, LOS 100%, 4,618 games, Ptnml [165,495,795,604,250]; SPRT 7.2-selfplay vs 7.1-head, elo1=3 @ 3+0.03). Merged to development (d113471), bench 13,544,942 → 14,095,183. Biggest single-step gain since the Phase-4 eval campaign — ~2× the ~+12 EV midpoint; proves the HCE eval was not near its ceiling and the self-play curve is still climbing (→ 7.3 is high-EV). Joint refit on 968,771 unique on-policy positions (200k games from the 7.1 head seeded by beast_seed.epd; --balance-phase 1.5 → 25/37/37 from natural 9/34/55). --tune all --l2 1e-6 (K=2.5) + the KS funnel re-fit deferred from 7.1 (sequential joint bake on the tune-all baseline). Eval moved substantively (queen_infiltration 5→8, minor_king_ring 1→5, safety_table tail 296→342/351, ks_central_king 0→3; material pinned). 9/9 CTest — KQK/KBNK canaries clean. Historical plan/pipeline notes retained below. Must follow 7.1: self-play WDL optimizes Basilisk winning from positions its own search reaches, correcting distillation's off-policy + representability bias. Regen self-play with the strongest head (datagen.ps1 → extract_parallel.py), one joint low-lr refit (--tune all + --tune-kingsafety), SPRT + gauntlet — NOT a re-stage. Fold in: phase-rebalance (v17 was 59% endgame), quiet-filter (extract.py already skips in-check/capture plies), optional blended labels λ·WDL+(1−λ)·σ(Basilisk-own score); ride-along re-fits: deferred nonlinear KS re-fit (now on correct data), sigmoid-K, tempo, lazy-margin, shelter/storm→danger fold. Expected ~+10–40 Elo. ⚠️ DATAGEN NEEDS THE DIVERSE BOOK. A first attempt (2026-07-05) with datagen.ps1's default SuperGM_4mvs.pgn book collapsed 200k games → 31,880 unique positions (fixed-node self-play from a tiny book replays identical games). The proven fix (PLAN.md §637–664, and datagen.ps1's own book-size warning) is to seed from tools/texel/data/beast_seed.epd (100k diverse openings from the Beast corpus) — Phase-2 got 1.73M unique from 200k games that way. Aside (2026-07-05): reusing Hydra's ready beast_train.csv (WDL on the same 2M positions 7.1 already distilled) was tried as a shortcut and abandoned — off-policy, the eval barely moved, and it slightly degraded KQK (mate 5→6); the null result confirmed on-policy self-play is the only productive 7.2. Concrete pipeline (this cycle): (A) datagen.ps1 -Suffix phase71-sfdistill -Book tools\texel\data\beast_seed.epd -BookFormat epd -Rounds 100000 -Nodes 8000 -OutputPgn ...\selfplay_72_beastseed.pgn; (B) extract_parallel.py ... --balance-phase 1.5 (float R = cap each bucket to R× the smallest; the parallel extractor now matches extract.py's flag exactly, commit 4b399b4); (C) --tune all --l2 1e-6; (D) --tune-kingsafety --max-positions 100000 (~8 min — the finite-diff path times out at 600k, calibrated 2026-07-05); (E) bake both (--allow-pst for the eval dump) → build_test.ps1 -Suffix phase72-selfplay; (F) SPRT elo1=3 vs basilisk-phase71-sfdistill. — Sonnet 5 medium
[~] 7.3 Iterate (further self-play cycles) + stop — ▶ ITERATING, curve NOT flat. Cycle 1 ACCEPTED +19.51 ± 7.04 Elo (H1, LOS 100%, 4,440 games, Ptnml [91,505,847,618,159]; SPRT 7.3-selfplay vs 7.2-head), merged 9741070, bench 14,095,183 → 13,495,492 — ~as much as 7.2's +21, so cycle-4 datagen is running. Repeat 7.2 with the now-stronger head; stop when a cycle yields <~+8 Elo. ⚠️ The stop signal is the SPRT, NOT the holdout-MSE stall — cycle 1's linear eval barely moved (holdout 0.0797→0.0793, all movers ±1/±2) yet gained +19.5; the KS funnel keeps re-shaping each cycle (safety_table tail 351→373). Each accepted cycle is releasable. Sustained flattening = pure-HCE tuning spent → Phase 9 (NNUE). — Sonnet 5 medium
7.4 histshape SPSA — the ONE search tune, at the final eval scale (relocated 6.9). Only after 7.1–7.3 settle. config_histshape.json = 11 always-live dims (6.3 history-shape 7 + LmrHistDiv + HistPruneCoeff + LmrCutNodeAdj + LmrTtCapture) starting from current defaults (iteration 1 = the validated head — unlike the superseded wave2, whose start activated known-negative mechanisms). Pre-registered rules: 1000–1500 iters (~12–18h) not 5000; abort at ~600 if the trend ≤ 0 and keep defaults; converged candidate → bake → PGO build → SPRT vs pre-tune head AND full ctest; a wash → keep defaults, done tuning. Gated 6.4/6.5/6.8 mechanisms stay inert unless an evidence-based wave3 case appears. ./tools/build_test.ps1 -Suffix phase74-histbase → ./tools/setup_spsa.ps1 -ConfigGroup histshape -EngineSuffix phase74-histbase -Iterations 1500 → cd tools/weather-factory && python main.py. — Sonnet 5 medium
7.5 Boundary validation + release 1.8.0 — ✅ DONE 2026-07-08. Validation was an LTC field gauntlet in colosseum (10+0.1, 12-engine field) + an independent LittleBlitzer cross-check, not the originally-planned cumulative SPRT + cross-TC riders (superseded by the direct gauntlet). 1.8.0-rc beats 1.7.0 head-to-head ~+40 Elo at LTC (the +93 fast-TC gain compresses to ~+40 at long TC — normal for eval refits; both tools agreed directionally). The absolute colosseum field-estimate (2939, Δ−111) is not a regression — it's anchored on strong fixed opponent ratings; the within-gauntlet 1.7.0 head-to-head is the trustworthy number. 7.4 histshape SPSA skipped (low EV). 1.8.0 = Phases 5+6+7 bundled, released from the cycle-5 head (bench 12,661,251). — Sonnet 5 low

Later

Phase 8 feature menu — demoted 2026-07-02 (EV ≈ 0; the SF/Ethereal/Weiss audit found no missing features pre-NNUE). Skip to Phase 9 (NNUE, +200–400, Fable 5 high / alt Opus 4.8 high) unless a specific gauntlet weakness points at a listed item.
Phase 9 NNUE remains the terminal option; eval boundary stays clean.

External gauntlet (opponents) + time budget

After each phase — especially after Phase 4 (eval), which over-fits self-play most — run tools/gauntlet.ps1 at tc=10+0.1. Recommended opponents (pick by measured score): Basilisk 1.5.0 (phase1-final) + Rarog; Critter 1.6a (~3150–3200, a target to clear); Stockfish capped via UCI_LimitStrength, start 2900 → 3100 → 3300 (Basilisk is already ~3100, so start high) at the level where Basilisk scores ~30–70%; one independent mid/high HCE (Lambergar / Peacekeeper / Igel). Time budget on the 5950X: datagen/SPSA at -Concurrency 24; Texel fits are CPU-minutes — run them freely.

Release points & version numbers

Last released: 1.6.0 (Phase 2 eval scalars + time-forfeit fix). 1.7.0 prepared 2026-06-29 (Phase 4 eval data-fit campaign; version/CHANGELOG/README done, awaiting the maintainer's tag). When to tag the next ones (full rationale + gates in PLAN.md §10):

Tag after	What ships (cumulative since last tag)	Version
Phase 2.9	Phase 2 eval scalars (+54 vs 1.5.0) + time-forfeit fix	`1.6.0`
Phase 3	bench-identical only (except 3.5 endgame knowledge — behaviour-changing, SPRT-gated) — nothing to release	—
Phase 4	full eval data-fit campaign (+280.74 vs phase1-final)	`1.7.0` (shipped)
Phases 5+6+7	TM (+3) + search wave (+7.18) + eval-refresh & one search tune (+30), bundled	`1.8.0` (unreleased; ships at Step 7.5)
Phase 8	demoted (EV ≈ 0) — only if a gauntlet weakness surfaces	patch per batch
Phase 9	NNUE (+200–400)	`2.0.0`

Rules: minor bump (1.Y.0) for each phase that banks SPRT+gauntlet-validated strength; patch (1.y.Z) for fix-only releases; major (2.0.0) for the NNUE architecture swap. The version tracks everything new since the last tag, not the last phase — that's why the post-2.9 release is 1.6.0 (it ships all of Phase 2's eval gains), even though Step 2.9 itself is just a fix.

Released as 1.6.0, not 1.5.1. 1.5.1 (a patch) would have undersold the biggest eval gain so far. The release was gated on the time-safety fix (Phase 5 Step 5.1/5.1b, old labels 6.1/6.1b) clearing t=≈0 and the +54 holding vs phase1-final — the forfeit fix is what made the eval gains releasable.

To prepare a release, say: "Release 1.6.0." The model runs the checklist in PLAN.md §10 — bumps the version in both src/Constants.h and CMakeLists.txt, updates CHANGELOG.md (README no longer carries per-version content), verifies the non-tune release build hides the tuning options + passes CTest, builds the dist binaries (the full platform matrix comes from .github/workflows/release.yml on tag push), and commits. It will not tag and will not push — those are yours. Create the tag and push whenever you're actually ready to cut the release:

git tag -a v1.6.0 -m "Version 1.6.0"
git push origin <branch>
git push origin v1.6.0

It will refuse to even commit if the version's pre-release gate (SPRT/gauntlet) hasn't passed — so settle that first.

Common Commands

# Build a named pext-PGO+TUNE binary
.\tools\build_test.ps1 -Suffix <name>

# SPRT a gain candidate (default tc=3+0.03 clock)
.\tools\sprt.ps1 `
    -EngineA tools\test_engines\basilisk-<candidate>-pext-pgo.exe `
    -EngineB tools\test_engines\basilisk-<baseline>-pext-pgo.exe `
    -NameA "Candidate" -NameB "Baseline"

# SPRT a small/tighter candidate
.\tools\sprt.ps1 `
    -EngineA tools\test_engines\basilisk-<candidate>-pext-pgo.exe `
    -EngineB tools\test_engines\basilisk-<baseline>-pext-pgo.exe `
    -NameA "Candidate" -NameB "Baseline" -Elo1 3

# LTC confirmation / TC-suspect candidate
.\tools\sprt.ps1 `
    -EngineA tools\test_engines\basilisk-<candidate>-pext-pgo.exe `
    -EngineB tools\test_engines\basilisk-<baseline>-pext-pgo.exe `
    -NameA "Candidate" -NameB "Baseline" -TC "10+0.1" -Elo1 3

# Optional old fixed-movetime sanity gauntlet
.\tools\sprt.ps1 `
    -EngineA tools\test_engines\basilisk-<candidate>-pext-pgo.exe `
    -EngineB tools\test_engines\basilisk-<baseline>-pext-pgo.exe `
    -NameA "Candidate" -NameB "Baseline" -MoveTime 0.1

# Refactor/default-equivalence SPRT, only when needed
.\tools\sprt.ps1 `
    -EngineA tools\test_engines\basilisk-refactor-pext-pgo.exe `
    -EngineB tools\test_engines\basilisk-baseline-pext-pgo.exe `
    -NameA "Refactor" -NameB "Baseline" -Elo0 -3 -Elo1 3

# Phase-boundary fixed-game validation
.\tools\gauntlet.ps1 `
    -Engine tools\test_engines\basilisk-phase1-final-pext-pgo.exe `
    -Opponents tools\test_engines\basilisk-phase1-defaults-pext-pgo.exe `
    -Name Phase1Final `
    -Games 2000

# Configure SPSA
.\tools\setup_spsa.ps1 -ConfigGroup pruning -Iterations 5000
.\tools\setup_spsa.ps1 -ConfigGroup lmr -Iterations 5000
.\tools\setup_spsa.ps1 -ConfigGroup combined -EngineSuffix phase1-lmr -Iterations 2000

# Run/resume SPSA
cd tools\weather-factory
python main.py

Bench is best run interactively:

.\build\release-pext\basilisk.exe
bench 13
quit

Why Tuning Options Exist

weather-factory can only perturb the engine through UCI. For example, it sends:

setoption name RfpCoeff value 128

That is why search constants are UCI options in tune builds. They are hidden in normal release builds through the BASILISK_TUNE compile definition.

Before any public release, verify a non-tune build does not expose the tuning option list.

Ground Rules

Do not accept a tuned value set without SPRT.
Do not interpret lower or higher node count as strength.
Do not bundle feature work with tuning defaults.
Do not skip fixed-game validation at phase boundaries.
Do not start Phase 3 features while Phase 1 or Phase 2 still has obvious tuning work left.
Keep Evaluator::evaluate(const Board&) as the boundary between search and evaluation.

The process is the strength engine here: tune, test, keep only what survives.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Basilisk Development Workflow Guide

Current Checkpoint

The program in one table (overview · model picker · Elo)

The Basic Rhythm

Next Commands

What To Report Back

SPSA Result

SPRT Result

Bench Result

Texel / Tuner Result (Phase 4)

Errors

Decision Points

Phase Progress Tracker

Phase 0 - Harness

Phase 1 - Search Constants

Phase 2 - Eval Tuning (Texel pipeline; PLAN.md Section 4)

Phase 2.9 - Robustness Quick Win (DONE — shipped in 1.6.0; PLAN.md §4.9)

Phase 3 - Eval Structure Build-Out (PLAN.md §6; EXECUTE FIRST — bench-identical, no games)

Phase 4 - Eval Data-Fit Completion (PLAN.md §6.5; the campaign, SPRT per stage)

Phase 5 - Time Management hardening + tuning (PLAN.md §7) — promoted to Phase 5, executed BEFORE the Phase 6 search wave (2026-06-29)

Phase 6 - Search Efficiency Wave (PLAN.md §5; renumbered to Phase 6 on 2026-06-29 — executed AFTER Phase 5 time management. SPSA last — driving Sonnet 5 medium, dense/interaction-risk ports Fable 5 medium-high with an Opus 4.8 alternative — Fable 5 is available for a limited time)

Phase 7 - Non-NNUE ceiling: eval-refresh cycles + the one search tune (PLAN.md §7.5; THE ACTIVE PHASE after 6.10 — reworked 2026-07-02 EV review, absorbs the relocated 6.9 SPSA as 7.4)

Later

External gauntlet (opponents) + time budget

Release points & version numbers

Common Commands

Why Tuning Options Exist

Ground Rules

FilesExpand file tree

user_dev_guide.md

Latest commit

History

user_dev_guide.md

File metadata and controls

Basilisk Development Workflow Guide

Current Checkpoint

The program in one table (overview · model picker · Elo)

The Basic Rhythm

Next Commands

What To Report Back

SPSA Result

SPRT Result

Bench Result

Texel / Tuner Result (Phase 4)

Errors

Decision Points

Phase Progress Tracker

Phase 0 - Harness

Phase 1 - Search Constants

Phase 2 - Eval Tuning (Texel pipeline; PLAN.md Section 4)

Phase 2.9 - Robustness Quick Win (DONE — shipped in 1.6.0; PLAN.md §4.9)

Phase 3 - Eval Structure Build-Out (PLAN.md §6; EXECUTE FIRST — bench-identical, no games)

Phase 4 - Eval Data-Fit Completion (PLAN.md §6.5; the campaign, SPRT per stage)

Phase 5 - Time Management hardening + tuning (PLAN.md §7) — promoted to Phase 5, executed BEFORE the Phase 6 search wave (2026-06-29)

Phase 6 - Search Efficiency Wave (PLAN.md §5; renumbered to Phase 6 on 2026-06-29 — executed AFTER Phase 5 time management. SPSA last — driving Sonnet 5 medium, dense/interaction-risk ports Fable 5 medium-high with an Opus 4.8 alternative — Fable 5 is available for a limited time)

Phase 7 - Non-NNUE ceiling: eval-refresh cycles + the one search tune (PLAN.md §7.5; THE ACTIVE PHASE after 6.10 — reworked 2026-07-02 EV review, absorbs the relocated 6.9 SPSA as 7.4)

Later

External gauntlet (opponents) + time budget

Release points & version numbers

Common Commands

Why Tuning Options Exist

Ground Rules