DeepfakePolicy review worksheets — 7 October 2026

These are editorial worksheets for your own review process. They are not an API
request, an imported company template, a detector report or a claim decision.
Downloading a worksheet performs no analysis and creates no paid job.

detector-pilot-blank.csv
  Header only. Add your independently labelled files and their recorded outcomes.
  Empty fields mean unavailable or not yet assigned; they must not become zero.

detector-pilot-nasa-observation.csv
  One editorial transcription of the live check made on 2 October 2026.
  Exact input: AS17-148-22727_lrg.jpg, 743020 bytes.
  SHA-256: 83a4a11f58dc1ef12162b009c7601704f01eb96bd235a491b420265f3684998e
  AI-generated image signal: 6.9%; conclusion: No strong AI finding returned.
  This is one attributed genuine-source web JPEG reproduction. It is not a
  benchmark, an error-rate measurement or a test of generated images or video.
  Source attribution was established separately from the detector check.
  No transformed variants or new live checks were run to produce this worksheet.
  Inspect the original files and report at:
  https://deepfakepolicy.com/our-tests/nasa-blue-marble-ai-image-check

Detector worksheet fields
  record_id: recorded job/report identifier or your stable local file identifier.
  record_kind: distinguish a plan, a live observation and a fictional example.
  source_family: group derivatives of the same source; do not count them as
    independent source examples when reporting a sample size.
  variant: original, resize, recompression, screenshot or another documented copy.
  filename / input_sha256: identify the actual bytes submitted.
  reference_origin / label_basis: independently established origin and its basis;
    another detector score is not ground truth.
  analysis_mode / executed_at: recorded processing scope and timestamp.
  reported_category / ai_signal_percent / coverage: preserve the returned finding
    and limitations; do not invert a signal into an authenticity probability.
  review_needed: your frozen operational rule, if assigned.
  reviewer_label: explicit independently supported reference label, if assigned;
    a reviewer disposition alone is not automatically ground truth.
  notes: uncertainty, processing failures, unavailable checks and follow-up.

Counting rules
  Freeze thresholds on development data before evaluating held-out labelled data.
  Record attempted inputs, completed reports, failures and inconclusive results.
  State which inputs are eligible for each metric and publish denominators.
  Keep unresolved or unsupported inputs visible; do not quietly exclude them.
  Sensitivity = true positives / all eligible labelled generated inputs.
  False-positive rate = false positives / all eligible labelled genuine inputs.
  A small pilot cannot establish a population error rate.
  Re-exporting preserves a recorded result. Reanalysis is a new job and may cost.

claim-evidence-blank.csv
  Header only. Record your question, description, source/page, observed fact,
  missing evidence, next request, actual report citation and reviewer decision.

claim-evidence-fictional-example.csv / .json
  Entirely fictional editorial worked example CLAIM-DEMO-001.
  No live detector or comparison was run. The imagined photo and invoice are not
  supplied files. Dates and repair items illustrate follow-up questions only.
  They do not establish fraud, claim validity, coverage, repair value or payment.
  The JSON is a worksheet, not a POST /v1/comparisons payload or report schema.
  actual_report_citation and reviewer_decision remain empty; do not invent them.

Claim worksheet fields
  case_id / record_kind: identify the local case and whether data are fictional.
  review_question / description_fact: the specific question and supplied claim.
  source_name / source_page: your file/page reference, when present.
  illustrative_observation: fictional in the example; use actual cited facts
    when filling your own worksheet, and rename this column if appropriate.
  open_point / next_request: missing evidence and the precise follow-up request.
  actual_report_citation: fill only from a real recorded comparison, if obtained.
  reviewer_decision: your authorised review decision, separate from model findings.

Keep authorised originals in your own evidence system. These worksheets do not
retain uploaded originals or replace the chosen workflow's retention rules.