nocturn9x’s tests

Switch to remiel-v2 (bench 4810274) Use MattBench run 49, superbatch 1100, with a 768-neuron FT and 32/32 hidden layers. Match the synthetic fixture and SIMD checks to the production dimensions. On all 7,153,653 lichess-big3 positions, the absolute mean is 928.76 for remiel and 949.83 for remiel-v2 at scale 282. getScale yields 275.744417, rounded to EVAL_SCALE=276. Co-authored-by: GPT 6.1 Sol <codex@openai.com>

Scale soft time bound according to half move clock (bench 5611448)

Scale soft time bound according to half move clock (bench 5611448)

Double check zobrist collisions using our position stack (bench 5611448) Co-authored-by: GPT 6 <codex@openai.com> Not actually coauthored by ze slopper btw, I just had it fix missing benches and it upid. Keeping this up for accountability purposes

Merge master and select A8607157 (bench 5611448) Bring tyrfing-12ksn up to date with master and select the new A8607157 network with the unchanged multilayer TI architecture. Keep the weights local in networks/files/tyrfing-12ksn-v2.bin; network publication remains deferred until the net is merged. Calibrate on all 7,153,653 positions in lichess-big3-resolved.book. At scale 305, getStats reports absolute means of 927.90 for master's 510EEBF8 network and 1004.55 for A8607157. getScale recommends 281.727639, rounded to 282. At scale 282, the new absolute mean is 928.76. Update the documented default scale. Validation: make dev and the runtime-check build pass. Scalar and AVX2 NNUE tests each pass 17,731 incremental/full-refresh comparisons with identical checksums. Python regressions pass (99 tests, 5 skipped) with Git signing disabled for temporary test commits. The repository commit hook also passes make test. The optimized depth-13 bench is 5611448 nodes. Network SHA-256: a8607157fbb8ac900ac7fe25d45d772509db9ecf3d9b5cb5000e5a581e1d0890 Co-authored-by: GPT 6 <codex@openai.com>

Guard search heuristics and skip static evaluation in check (bench 5401484) Checked nodes carry noEval through the search stack and TT. Main-search futility pruning and the fail-low extension after a reduced parent search previously consumed static eval in check; require present evaluations before using them. Guard improving, reverse futility, null move and correction-history arithmetic likewise, and compute TT-adjusted eval only where needed. Disable qsearch stand pat when static eval is absent. Preserve capture-only move generation and SEE pruning here; full check evasions and qsearch mate recognition are isolated in a follow-up commit. Until then, checked qsearch nodes with no searched capture retain -SCORE_INF. At the hard ply boundary, return a neutral score in check instead of evaluating. Enforce the rawEval precondition with doAssert, including in danger builds. Subprocess tests verify this invariant and absent-eval unwrapping survive there. Validation: focused AVX2 checked/danger and scalar checked tests; exhaustive optional-eval/TT round trips and unchanged layouts; Python suite (99 run, 5 skipped). Optimized depth-13; depth 14 visited 8078278 nodes. Twelve alternating depth-14 pairs on CPU 2 versus the representation-only commit show -1.25% throughput change, 95% bootstrap interval [-7.32%, +3.49%]. Search trees differ, so this measures throughput rather than identical work or playing strength. Co-authored-by: GPT 6 <codex@openai.com>

Make qsearch keep track of PVs, avoid qsearching with invalid windows and don't cutoff QS in PV nodes (bench 5515662)

Switch to tyrfing-12ksn (bench 4979204) Select tyrfing-12ksn network 77F0596E and set the multilayer scale to 288. Update the documented default scale. At scale 305, the 7,153,653-position Lichess comparison gives absolute means 927.90 for master and 981.15 for the new network; getScale yields 288.446721, rounded to 288. Validation: make bench with 77F0596E at scale 288 gives 4979204 nodes at depth 13. Confirmed the network ID and compiled scale; the 10,000-position absolute mean is 911.04. Runtime-check build, full-book evaluation and Python regressions passed (5 skipped). Network weights remain local pending separate publication. Co-authored-by: GPT 6.1 Sol <codex@openai.com>

Switch to tyrfing-12ksn (bench 5228946) Select the completed SB 1100 tyrfing-12ksn network and set the multilayer evaluation scale to 274. Update the documented scale. The 7,153,653-position Lichess comparison gives absolute means 927.90 for master and 1032.37 for the new network at scale 305; getScale yields 274.135727, rounded to 274. Validation: make bench with the selected network at scale 274 gives 5228946 nodes at depth 13. Confirmed the built-in network ID and scale. Runtime-check build and Python regressions passed (5 skipped). The network weights remain local pending separate publication. Co-authored-by: GPT 6.1 Sol <codex@openai.com>

Document direct executable downloads and Unix permissions (bench 5551733) Co-authored-by: GPT-6 Astra <codex@openai.com>

Document direct executable downloads and Unix permissions (bench 5551733) Co-authored-by: GPT-6 Astra <codex@openai.com>