docs/anonymization_report.md

# Anonymization and sensitive-content report

This release is **de-anonymized on purpose**: the paper is accepted, so author
names, affiliations, and a contact email appear in `README.md`, `NOTICE`,
`CITATION.cff`, and `registry/release_manifest.json`. That is intended, not an
anonymization failure.

What this report covers is everything that should *not* be here.

## Scan performed

A recursive case-insensitive grep across every staged file, for:

- absolute local filesystem path prefixes (user home directories)
- absolute cluster filesystem path prefixes (the project's scratch and
  campaign-storage roots)
- API-key prefixes, bearer-token headers, and credential-style key names
- private-key PEM headers
- served inference endpoint hostnames and URLs
- scheduler job scripts, job identifiers, and queue names
- the author's username as a bare string

This document deliberately does not reproduce those patterns literally. A report
that quotes the strings it forbids will match its own scan, which makes the scan
result unreadable and tempts a maintainer to add a file exclusion. The scan
itself lives in the validation script, where the patterns belong.

## Result

**Clean.** No local or cluster filesystem paths, no credentials, no endpoint
URLs, no scheduler artifacts, and no internal run logs appear in any staged file.

The author's username appears only as part of the intended contact email address,
in exactly four files: `README.md`, `NOTICE`, `CITATION.cff`, and
`registry/release_manifest.json`. Every occurrence was checked individually and
every one is that email address.

The scan was verified to be capable of failing: a canary string in the shape of
each forbidden pattern was planted in a temporary copy of the tree and the
scanner flagged every one. A scan that has only ever run on clean data detects
nothing known.

## Specific things that were stripped or never staged

- **Endpoint URLs.** The source `run_metadata.json` files record the served
  inference endpoint. Those files were not staged; the confidence-probe table was
  rebuilt column by column and the endpoint field was dropped. The model
  identifier (`openai/gpt-oss-120b`) is retained because it is scientific
  metadata, not infrastructure.
- **Absolute paths.** Source manifests (`run_manifest.json`, `diagnostics.json`,
  `confidence_probe_staging_index.csv`) embed cluster and local paths. None of
  those files were staged. Where a number from them was needed, only the number
  was carried across.
- **Scheduler artifacts.** Job scripts, job identifiers, and run logs exist in
  the source tree. None staged.
- **Reviewer and post-review discussion.** The source tree contains reviewer
  response plans, strategy notes, and OpenReview drafts. None staged.
- **Raw model responses.** The confidence-probe source table has a
  `raw_response` column of model-generated text. Dropped. Only the parsed integer
  probability, the parse status, and token counts are released.
- **Model rationale text.** The probe's free-text `reason` column was dropped;
  it is model-generated and not needed for scoring.
- **Serialized models and feature builders.** Not staged.

## Residual identifiability

Problem identifiers are stable and derived from upstream benchmarks, so they are
not a privacy concern. The benchmarks contain no personal data: competition math
problems, entrance-exam questions, textbook problems, and biology protocol
questions.

No human-subject data, no user data, and no personally identifying information
about third parties appears in any released table.