Whether the data used to train, fine-tune, or retrieve is licensed, clean, and free of the PII and contamination that create legal and quality risk.
Free & open · no signup · read-only — it ends by asking · nothing leaves your machine
You are working inside this repo. Mission: trace the data behind the product's models — training sets, fine-tuning data, and retrieval corpora — and judge whether it is licensed for this use, free of unconsented personal data, and uncontaminated. If the product only calls a hosted model, focus on retrieval and prompt data.
Read-only pass. Read the data-loading, fine-tuning, and indexing code and any dataset docs; change nothing but the report file.
Phase 2 sweeps the codebase through every one of these, citing file and line for each finding.
where the data came from and whether its license permits this use
personal data in training sets or corpora, and the basis for using it
data scraped or reused against its terms; user data used to train without disclosure
test or eval data leaking into training; benchmark leakage inflating results
copyrighted content whose reproduction creates exposure
can you trace what went into a model or index, and remove a source if required
can specific data be deleted from datasets and downstream indexes on request
One structured report at the repo root — or in reports/, if you keep one — the same shape every time, ready for a teammate — or the optional Studio — to act on.
Every Goal Prompt follows the same four steps, so results are consistent and repeatable — no matter which one you run.
List every dataset the product trains on, fine-tunes on, or retrieves from.
Source & licensing — where the data came from and whether its license permits this use.
Rank by legal exposure and quality impact.
Create PROVENANCE.md at repo root.
Copy it and paste it into your agent inside the repo you want checked.
Install the goal plugin once — two commands — then just type /goal:training-data-provenance.
/plugin marketplace add GhostlyGawd/goal-prompts/plugin install goal@goal-promptsOr install only this Goal Prompt as /goal-training-data-provenance:
curl -fsSL https://goal-prompts.vercel.app/install | BRIEF=120 shLet an agent fetch it mid-conversation, or pull the raw Goal Prompt by URL.
https://goal-prompts.vercel.app/raw/120.mdNothing hidden — this is the whole Goal Prompt, verbatim. Read it in a minute, edit it, or copy it as-is.
# Goal: Training-Data Provenance You are working inside this repo. Mission: trace the data behind the product's models — training sets, fine-tuning data, and retrieval corpora — and judge whether it is licensed for this use, free of unconsented personal data, and uncontaminated. If the product only calls a hosted model, focus on retrieval and prompt data. Read-only pass. Read the data-loading, fine-tuning, and indexing code and any dataset docs; change nothing but the report file. ## Phase 1 — Inventory the data - List every dataset the product trains on, fine-tunes on, or retrieves from. - For each, find where it came from and under what terms. - Note whether personal or user data is among it. ## Phase 2 — Audit through 7 lenses 1. **Source & licensing** — where the data came from and whether its license permits this use 2. **PII in data** — personal data in training sets or corpora, and the basis for using it 3. **Consent & terms** — data scraped or reused against its terms; user data used to train without disclosure 4. **Contamination** — test or eval data leaking into training; benchmark leakage inflating results 5. **Copyright & attribution** — copyrighted content whose reproduction creates exposure 6. **Lineage** — can you trace what went into a model or index, and remove a source if required 7. **Retention & deletion** — can specific data be deleted from datasets and downstream indexes on request ## Phase 3 — Curate - Rank by legal exposure and quality impact: unlicensed data or PII with no basis outranks a minor attribution gap. - For each, name the control — a license check, a PII scrub, a lineage record, a deletion path. - Separate a legal risk from a quality risk; both matter, differently. ## Phase 4 — Report Create `PROVENANCE.md` at repo root: 1. **Data inventory** — dataset · source · license · contains PII? 2. **Findings** — each: risk (legal or quality) · dataset · the issue · the control 3. **Lineage & deletion** — the tracking and removal capability to build 4. **Priority** — the exposures to remediate first Start the report with today's date. If `PROVENANCE.md` already exists from a previous run, read it first and lead with what changed since. ## Rules - If you cannot say where the data came from, you cannot defend using it - Contamination invalidates your evals before it ever reaches production - No training or fine-tuning data in this repo? Say so in a one-paragraph null report and stop — a null result is a valid finding. - If a `reports/` directory exists at the repo root, write the report there instead of the root. - Before asking, present the top findings as a ranked list in plain words - Report only — end by asking which data risks to address first
reports/ directory exists at the repo root, write the report there instead of the root.Find the code that ranks, scores, or filters people, list the exact fields it consumes, and check whether any eval ever sliced results by group — suspicions labeled.
Whether the product's AI outputs are grounded in real sources or confidently invented — and whether it cites, hedges, and refuses when it should.
Attack the product's own AI like an adversary — make it ignore instructions, leak secrets, or misuse tools — and find where untrusted input can hijack it.
Diff what the UI says about the AI against what the code does — disclosure strings quoted, dropped confidence traced, retention claims checked against real payloads.
Runs inside these playbooks — curated sequences you can launch with one paste: