Internal classifier export

The classifier keeps no state. Every verdict, passes included, is stored on the request's log.gateway_content_filter_response (internal classifier entries have model: "internal-classifier") and is cleared by log retention after 30 days. Fingerprints are present only while the classifier runs with its hash-all option. See the gateway skill for how the gateway calls it.

Admin dashboard export

Content Filter page → "Internal classifier checks", backed by GET /admin/content-filter/internal-checks/export (apps/api/src/routes/admin-internal-classifier-checks.ts): up to 24 hours per download, real request and organization ids. Its CSV formula guard prefixes ' to minhash_b64 values starting with +, so do not send dashboard downloads to the vendor; use the scripts below.

Vendor sample

Load the database skill first: this reads log in production. The scripts read only the content filter columns, never prompts or completions, inside a read-only transaction, through log_content_filter_created_at_idx.

1. Dump. Needs kubectl access to the production cluster. Run per window of about 20 minutes (~9k rows); longer kubectl exec streams drop and are retried from scratch:

   KUBE_CONTEXT=<production context> sh .agents/skills/internal-classifier/scripts/fetch.sh \
     "2026-10-27 15:00:00" "2026-10-27 15:20:00" <dir>/raw-1500.ndjson.gz
   

   Times are UTC, [from, to). Each file ends with {"__total": N}. Retried attempts are skipped: the final attempt carries the same verdicts and the model that served the request.

2. Convert with the admin export's own code. It imports workspace packages from their dist/, so run pnpm build:core first on a fresh checkout:

   pnpm exec tsx --tsconfig apps/api/tsconfig.json \
     .agents/skills/internal-classifier/scripts/convert.mts <out> <dir>/raw-*.ndjson.gz
   

   • <out>/send/checks.csv: one row per verdict in the vendor audit CSV's column order (ts … band_keys, account_id, decision), then request_key and chunk_index. Request ids are replaced by request_key, organization ids by acct-NNN (acct-001 = most checks).
   • <out>/send/rule-summary.csv, <out>/send/hourly.csv.
   • <out>/internal/export-all.csv and <out>/internal/account-pseudonyms.csv hold real ids. Never send them.

   It prints totals for the README.

3. README. Write <out>/send/README.md: window, classifier image digest and cyber profile (from the running deployment), whether fingerprints were on, the file list and column notes below, the totals, and the questions for the vendor. Then cd <out> && zip -qr -9 <name>.zip send.

Keep <out> and the dumps outside the repository or under .context/.

Column notes

• ts is when the gateway logged the request (microseconds), not the classifier's clock.
• rule is empty on passes; reasons can still list signals that fired.
• fp_status=too_short rows have empty hash columns and no minhash_b64.
• chunk_index is 0 unless the content filter reads the whole conversation and splits it into several classify calls.