Indexed online-edge execution evidence

The complete archive retains all six workers, complete results/configs, source and tool snapshots, schema files, import receipts, logs, protocol and acquisition receipt. Its index binds 445 files by SHA256 and size. Raw images/scans and compilation caches are excluded. The archive checksum and comparison are available separately.

Implementation	Whole-worker median
NumPy with parallel preparation	62.852 s
Compiled raw-score CPU with parallel preparation	31.042 s
CUDA with overlapped initialization/preparation	14.036 s

This cohort measured 4.478x against NumPy and 2.212x against the compiled raw-score CPU control. Two observations per implementation ran in the declared counterbalanced order on an unreserved host. All 48 prepared frames and all six complete result payloads match exactly. Execution metadata is checked separately. Earlier development screens are not pooled with this cohort.

This is a working-copy source snapshot, with source_commit: null and base revision cdd95486a0540a91c3d1378e1f8e2ef6b38a2b97. All 400 package modules and seven benchmark tool/schema files are frozen by content. The record makes no committed-source or SOTA claim. The post-run Git content binding verifies the frozen package and toolchain against commit 4406195f9dd2e6e86a746eeac733b82364c8c90a, without rewriting the original lineage or measurements. The later benchmark export correction is not part of the measured toolchain. Bindings compare UTF-8 text after universal-newline normalization, matching the source snapshot and Python's source parsing. Binary content is unchanged.

python tools/audit_online_edge_cuda_archive.py \
  docs/assets/online_edge_cuda/complete-evidence.tar.gz \
  --output outputs/online-edge-archive-audit.json

The auditor uses CPU-only JSON/schema/hash checks. It does not execute archived Python or import CuPy/Calibrex. It rejects unsafe paths, duplicate files, unindexed content and external schema references. With --bind-commit REF, it also checks every package module and benchmark tool against that later Git revision; the output labels this as a post-run content binding.

To reproduce the saved binding, add --bind-commit 4406195f9dd2e6e86a746eeac733b82364c8c90a to the audit command.

Reproduction uses the benchmark driver described in the execution guide. The input is the same 48 already-spent KITTI image/cloud pairs, with the upstream archive and all 101 input-file hashes in receipt.json. Calibration is an initializer reference for this performance control, rather than independent accuracy truth.