Skip to content

T2: full truth replay on fa0fd90, 2026-10-16

Verdict: PASS (compare.txt). compare.json equals that of 2026-10-15 (../../2026-10-15/t2/README.md, which gives the counts) in every field: 173 of 173 recordings in the report; 942,445,814 of 942,451,934 values match; 0 zero-sign and 0 NaN-payload mismatches; the same 6,120 value mismatches at the old emulator recording's NaN-decode outputs; hash-only 2,251 of 2,251; hardware errors 113 of 193 reproduced; 2,466 not applicable; fastacc cross-check 0 mismatches in 512,512 values.

  • Code under test: fa0fd90 (a clean git worktree; repo_commit in replay_mmaemu.json); 63b7395, later, renamed the profiles; T2 on its code gave the same compare.json (../rtx-5070-ti-renamed/README.md). On fa0fd90 the adapter (tools/truth_adapter.py) replays the RTX 4060 recordings under the profile ada-rtx4060-torch, which it names explicitly; from 63b7395 on it names ada.
  • Command: python tools/replay_truth.py --out <out> --compare $MMAEMU_TRUTH_DIR/replay/replay_fixed.json --threads 12, on the CPU with CUDA_VISIBLE_DEVICES="", torch 2.7.1+cpu, numpy 2.5.4 and Python 3.12.2, as on 2026-10-15. <src> stands for the worktree and <out> for the output directory.
  • Wall time: 67 min (4011 s), with other CPU work on the host; 30 min (1772 s) on 7612fd8 (below).
  • Truth: manifest.json sha256 be1270ed64432a1311ab5e2c98ccb69028f72e3424ad4c6148d4cd468fba6695, unchanged after the run.

replay_mmaemu.json differs from that of 2026-10-15 only in the run metadata (start time, durations, paths, version string 0.1.0 instead of 0.1.0.dev0, commit) and in the seconds of each recording; every count and verdict is equal. The same replay ran on 7612fd8 (the release-candidate freeze) earlier the same day, with the same compare.json and every count and verdict of replay_mmaemu.json equal.