Nestbolt results

Superseded final acceptance run (seeds 1000-1004, held-out jobs included) of build 484ce8b. This build's final acceptance run stopped with an error on one development job (b-albano-strip-sheet, seed 1004, 60 s): a repacking trial that its own check rejected was treated as fatal, so G1 (every output valid, no crashes) was not met. Fixed in 75cbb32 with a regression test; Nestbolt's final run was then repeated in full with the fixed build against the same competitor records.

Material used = (sheets − 1) + used length of the last sheet ÷ its length, from an independent Clipper2 validator; lower is better. Each tool runs single-threaded at equal wall-clock budgets on the same PC, at most eight runs at once. A competitor's value per job is the better of its two settings (identical constraints, its own best settings); jobs where no competitor produced a valid layout earn no wins. G2 passes when the median reduction against the best competitor is at least 5.00%, Nestbolt wins or ties at least 90% of jobs and is never more than 1% worse (plan, section 6; the earlier raise to 10.77% was waived on 2026-10-12, decision 4).

Generated

2026-10-13 12:22 UTC from commit c12442f946

Hardware

AMD Ryzen 9 7950X3D 16-Core Processor, 16 cores / 32 threads, 63 GiB, Windows-11-10.0.26200-SP0; runs pinned one per physical core of the first CCD

Seeds

1000, 1001, 1002, 1003, 1004 for every tool

Tools

60 s budget

9.17% less material than the best competitor at the median over 55 jobs; wins or ties on 48 (87.3%); 1 more than 1% worse (worst -1.35%). This does not meet the G2 target. 19 of 75 jobs have no valid competitor result. Nestbolt layouts valid: 374/375.

Records: seed 1000, seed 1001, seed 1002, seed 1003, seed 1004; competitor medians: the audited final batches, every run listed in BENCHMARKS.md (excluded sessions: final-excluded.json, replaced harness faults: final-replaced-runs.json). Hover a value for its settings or per-seed results. 1 jobs with a failed Nestbolt seed count as failures.

300 s budget

9.61% less material than the best competitor at the median over 11 jobs; wins or ties on 11 (100.0%); 0 more than 1% worse (worst 0.82%). This meets the G2 target. 1 of 12 jobs have no valid competitor result. Nestbolt layouts valid: 60/60.

Records: seed 1000, seed 1001, seed 1002, seed 1003, seed 1004; competitor medians: the audited final batches, every run listed in BENCHMARKS.md (excluded sessions: final-excluded.json, replaced harness faults: final-replaced-runs.json). Hover a value for its settings or per-seed results.

Strip parity (G3)

Nestbolt's strip adapter against the Sparrow CLI on the 13 strip jobs at 60 s: median regression -0.041%, worst 0.137% (pass: median at most 1%, never more than 2%): pass.

Speed (G4)

Nestbolt at 30 s reaches the best competitor's 300 s median on 10 of 11 jobs (pass: at least 80%): pass.

All gates

Gate-by-gate evidence, including every failed and superseded run, is in STATUS.md and BENCHMARKS.md. All numbers on this page are in results.json; rebuild it with python scripts/results_page.py.

Back to the Nestbolt app