Ga naar inhoud

Pilot report template

A report for the published-dossier pilot (docs/reference/evaluation.md, "The published-dossier pilot") is a copy of this template with its fields filled in. Two sections are Mark's alone: the ranking-gate slot and the PII-gate slot. Nobody else fills those, and an empty slot is reported as "not yet run" — never as passed, never as failed.

1. The check

Copied verbatim from the pilot's check.json (see docs/reference/ingest.md, "check.json — the pilot's completed check"). Not re-described: check.json stays the source and the ingest gate (document-lifecycle) stays as it is.

field value
checked_by
checked_at
robots_txt
anubis
decision

2. The corpus

field value
number of documents
number of pages
pages without a text layer
scans

With no scans in the corpus, the OCR recovery path (unprocessable_ocr) is barely tested by this pilot.

3. The ranking-gate slot — Mark fills this, nobody else

The gate: the expected document in the top 8 for at least 4 of the 5 questions below, gating this pilot run only.

The questions and their expected documents are committed before any result is seen; the two timestamps below are the only way to read that order off the report afterwards.

  • Questions committed at:
  • Time of the first search:
# question expected document document found in top 8?
1
2
3
4
5

Outcome: not yet run.

4. The PII-gate slot — Mark fills this, nobody else

Mark's read of the anonymized text of every pilot document against the source. Any leftover direct identifier of a private person (name, address, e-mail, phone, BSN, IBAN) fails the pilot for that document, unless the document is excluded below with the reason written down.

Published material is already redacted by the municipality before wordsworth ever sees it; there is no unredacted original to measure what wordsworth would have missed. This report and this template make no claim about that — anonymization of unpublished documents remains unproven by this pilot.

document leftover identifier found excluded (reason)

Outcome: not yet run.

5. Findings so far

  • The 567-document dossier gooise-meren-woo-2022 has unknown provenance.
  • Provenance is not recorded when a document is delivered again: a duplicate keeps its first recorded provenance and fails the manifest run (see docs/reference/ingest.md, "Read-back: the proof, not a re-run"). Neither finding is fixed by this pilot or this report.