Skip to content

README.md

Latest commit

 

History

History
65 lines (54 loc) · 4.64 KB

README.md

File metadata and controls

65 lines (54 loc) · 4.64 KB
Raw

Post-pin confirmation gate — October 6, 2026 (reviewed by Claude, October 7)

Focused same-machine A/B confirming that the final dependency pins (GRiD 8dccbfa + GLASS 9e57178, consumed by HJCD 121982d) did not change quality or latency relative to the already-audited pre-pin build (2302d22, measured in audit_timing_2026-10-05/). It is not a comparison against the paper's hardware or competitor solvers, and it does not re-measure the October 5 refinement gains.

Endpoints (frozen, provenance re-checked read-only on October 7 with run.sh --check)

Label Meaning Commit Header sha (build_info) Binary sha
main pre-pin audit build (docs/open-tasks/audit_gate_2026-10-04/branch, own venv) 2302d22 efc2c758… a1f5a163…
branch post-pin primary repo (docs/open-tasks/upstream_2026-10-05/.venv, editable) 121982d 602db063… 2c91ca0a…

Both Release, CUDA 13.2.86, Python 3.12.3, NumPy 2.5.3, panda_hand_joint frame, identical foam spheres, hard collision mode, automatic refinement precision (S=1 fp64, S=4 fp32). Workload: four alternating rounds (endpoint order rotates per round), S ∈ {1, 4}, open-world B ∈ {1, 2000, 8000, 32000} on the shared Halton targets (targets.json sha 89d2d46e…), box_panda B ∈ {100, 2000} (tests/mb_problems.json sha 69a3eae8…), 100 targets per cell. Launcher/driver/helper hashes in SHA256SUMS and manifest.json match the tree at 121982d. Installed binaries re-hashed October 7: unchanged.

Independently verified (October 7)

  • 96 unique CSV rows (48 per endpoint), 32 worker logs with 100 targets each, 9,600 measured attempts; no worker errors; analysis.json recomputed from results.csv and equal to the stored file to 1e-9.
  • Every matched per-round cell has identical solved / empty / returned counts. Open-world S=1: 95/100 at B=1 and 100/100 at B≥2000 (both endpoints, every round); S=4: 96/100 at B=1, else 100/100, 400 configurations returned; box_panda: 100/100 and 100/400 everywhere; zero empty outputs. "Solved" means at least one returned candidate satisfies pos < 5 mm AND ori < 0.05 rad on the same candidate under the driver's independent FK, with reported errors cross-checked; in collision mode every returned candidate was also verified free under the hjcd sphere oracle (the driver raises otherwise).
  • Launcher log: GPU at 0 % utilization, 29 °C, 6.5 W, idle clocks at start; the gate aborts on any foreign compute PID before and during every worker (2 s polling) and none appeared. Run wall time 64 s (04:12:39–04:13:42 UTC). Exit was clean (analysis.json is only written after analyze() succeeds). The coordinator-side account of CPU/graphics quiet was not available to the reviewer.

Paired latency ratios (branch / main, per-round medians; 100 targets per cell)

Workload S B main median ms branch median ms paired ratio round range
open-world 1 1 2.0067 2.0063 1.0001 0.9989–1.0005
open-world 1 2000 1.2610 1.2602 0.9946 0.9856–1.0382
open-world 1 8000 1.6822 1.6800 0.9988 0.9915–1.0066
open-world 1 32000 2.7604 2.7688 1.0046 0.9941–1.0078
open-world 4 1 1.4154 1.4153 0.9997 0.9992–1.0006
open-world 4 2000 1.3883 1.3879 1.0009 0.9977–1.0021
open-world 4 8000 2.4880 2.4891 1.0013 0.9970–1.0034
open-world 4 32000 6.8365 6.8428 1.0013 0.9993–1.0033
box_panda 1 100 2.0903 2.0932 1.0016 0.9997–1.0031
box_panda 1 2000 2.0894 2.0928 1.0016 0.9979–1.0048
box_panda 4 100 1.8667 1.8653 0.9990 0.9966–1.0040
box_panda 4 2000 1.9408 1.9423 1.0005 0.9973–1.0040

Verdict: unchanged. All twelve paired medians lie within 0.9946–1.0046; no cell is consistently slower or faster across rounds. The one round outside ±1 % (open-world S=1 B=2000, one round at +3.8 % while the cell's other rounds are at −1.4 … +0.2 %) is single-round noise, not a trend; it does not meet the "median > 1.03 or consistently slower" review trigger. Ratios near one are "no change", not a win. This confirms the pin update is performance-neutral on this corpus; it establishes nothing about other hardware, the competitor baselines or the paper tables.

Files

results.csv, manifest.json, analysis.json, launcher.log, worker_logs/ (verbatim), gate.json, run.sh, SHA256SUMS (the frozen launcher inputs). Original outputs remain untouched under docs/open-tasks/postpin_timing_2026-10-05/results-20261006T041238798290236Z/.