hw-apex-bench v0.2.0 · Evidence, not estimates

refrel3 / q4k / off

Synthetic evidence establishes correctness, not performance. Unknown machines remain separate; missing values are not zeros. Axis 5 v1 supports synthetic experiments only.

Leaderboard row

{
  "axis": 5,
  "crash": 0,
  "detected": 0,
  "evidence": "evidence/844b19c7cbb91239705694a34349abce5e5846fe51e621b45e57cc6c302d777d/evidence.json",
  "evidence_sha256": "844b19c7cbb91239705694a34349abce5e5846fe51e621b45e57cc6c302d777d",
  "format": "refrel3",
  "hang": 0,
  "harmless": 0,
  "hash_mode": "off",
  "kind": "synthetic",
  "machine": null,
  "refusal": 100,
  "scope": null,
  "silent_error": 0,
  "status": "PASS",
  "table": "axis5-844b19c7cbb91239705694a34349abce5e5846fe51e621b45e57cc6c302d777d-hash-off",
  "variant": "q4k"
}

Exact evidence JSON · SHA-256 844b19c7cbb91239705694a34349abce5e5846fe51e621b45e57cc6c302d777d

All evidence fields

Every evidence field is in the exact evidence JSON (SHA-256 844b19c7cbb91239705694a34349abce5e5846fe51e621b45e57cc6c302d777d, 148886 B), stored once and shared by every row of that evidence.

Hash-bound files

Shared input files retain their manifest associations; association is not a machine inference.

External data-root files (referenced by hash, not copied)

Reproduce and independently verify

Run in the applied benchmark clone, with this complete site at ./site. The output directory must be new. These commands verify existing evidence and regenerate HTML; they do not rerun measurements. DATA_ROOT is the release-input root holding the referenced catalog directories with exactly the listed bytes.

set -euo pipefail
python3 -m tools.release_v020 --verify site/downloads/release --manifest-sha256 8208dc3d1f7f07d0b8a345db239dd8076ff151f0ed93c8bfe0ca6b90d517d774 --data-root DATA_ROOT
python3 -m tools.build_site site/downloads/release --out rebuilt-site --manifest-sha256 8208dc3d1f7f07d0b8a345db239dd8076ff151f0ed93c8bfe0ca6b90d517d774 --data-root DATA_ROOT