AK-Benchbenchmark report

docs/EXPORT.md

Export

akbench export writes one experiment out as a self-contained artifact: the

descriptor, the pinned descriptor snapshots, the variants, the runs, the

metrics, the manifests and the artifacts, plus the same document as JSON and

CSV.

Formats: html (default — the experiment page from the report, standalone),

json and csv.

bash
akbench export --experiment exp-0001 [--format json|csv|html] [--output DIR] [--json]

Filtering secrets

Before anything is written, every string in the export is screened for

material that must not leave the machine:

  • environment variable names that the project classifies as secrets;
  • JWTs and PEM blocks;
  • OpenAI-style and AWS-style keys;
  • Authorization headers and Bearer tokens;
  • long hex strings that read as keys.

A match is not printed: the value is replaced by [REDACTED:kind] and the

kind is listed in the redactions section of the result, so a reader can see

that something was withheld without learning what it was. The screening

applies to JSON and CSV; the HTML export is the report, which already omits

secrets because the history stores them only as presence, never as value.

What an export contains

  • the experiment descriptor and its identity;
  • the pinned descriptors with their digests;
  • every variant with its configuration hash and baseline flag;
  • every run with its verdict, score, failure class and error message;
  • the recorded aggregates of the experiment;
  • the artifacts with their checksums, and the manifest digest of each run;
  • the redactions made, by kind and count.

The export is deterministic: the same history produces the same bytes, apart

from the generated_at timestamp, which is a field, not part of the payload

that is redacted or hashed.