Skip to main content

Module audit_report

dsfb_gpu_debug_demo::cli

Module audit_report 

Source
Expand description

S-REAL.1 — Deterministic, no-JavaScript HTML audit report renderer.

WHY: The S-REAL.1 audit gauntlet’s central credibility deliverable is that a human reader can open reports/s_real_1/<dataset>/audit_report.html in any browser, see exactly what DSFB-GPU saw on the upstream fixture, cross-check every hash against the receipt files, and confirm that two consecutive dispatches produced byte-identical output. The renderer must therefore be:

  • Byte-stable: same inputs → same byte stream across two calls and across two machines. We enforce this by sorting all map iteration (BTreeMap), pre-sorting the episode list by (entity_id, start_window, end_window), and never emitting wall-clock timestamps, process IDs, or any other non-pinned dynamic content.
  • No JavaScript: the report renders entirely server-side; the HTML carries inline CSS for layout. A reader on an air-gapped audit workstation can open the file without any network or runtime dependency.
  • No charts (v1): charts would require either an external library (network dependency) or a hand-rolled SVG renderer whose visual conventions would themselves need provenance. Tables of admitted episodes + stage hashes are the honest v1 surface; charts/heatmaps are S-REAL.1.1 scope.

Section order (panel-locked):

  1. Input provenance — fixture identity + SHA-256 byte pin.
  2. Residual-projection lowering law — exact rule used to map cells into TraceEvents.
  3. Run configuration — contract grid, scale parameters, episode count.
  4. Admitted episodes — table of bank-admitted episodes with motif + reason + peak fields.
  5. Stage digest / hash chain — every per-stage hash + final case-file hash.
  6. Replay verification — byte-identity between two consecutive runs.
  7. Limitations and non-claims — verbatim panel-locked text.

Non-claims (rendered verbatim in section 7 of every report):

  • Does NOT claim DSFB has identified the “real” anomaly in the dataset.
  • Does NOT claim DSFB outperforms any other anomaly detector.
  • Does NOT claim DSFB has discovered causality.
  • Does NOT claim fitness-for-purpose on regulated / safety-critical use.
  • Does NOT claim the dataset is “correctly labeled” or “ground truth”.
  • Does NOT claim the corpus or registry is exhaustive.
  • Does NOT claim replay determinism across driver / CUDA versions; the replay receipt records the toolchain explicitly.

License: Apache-2.0. Background IP: Invariant Forge LLC.

Structs§

DatasetManifest
Provenance + identity for the dataset under audit.
ReplayVerification
Replay-verification metadata.
SchemaMap
Schema map describing how upstream cells project into DSFB-GPU events.

Functions§

render_audit_report_html
Render the complete audit report as deterministic, no-JavaScript HTML.