source-ingest

Source Ingest

Use this skill to process a raw source into a normalized record before analysis.

Do not skip this step when the user provides an external document that will be used as evidence. Running analysis directly on an unprocessed source risks attributing the wrong source type, missing date/author context, and carrying in unsupported claims.

When to use

• User provides a PDF, DOCX, XLSX, or file path.
• User pastes a URL or article text.
• User shares a transcript (speech, meeting, interview, briefing).
• You are about to run agenda-intelligence or a policy-risk memo and the evidence base is user-provided or a specific retrieved document.

When not to use

• General web browsing without a specific document.
• Reasoning-only analysis where no source is being ingested.
• You are writing or editing a memo that is already sourced.

Live retrieval limitations

Many authoritative sources (OFAC/Treasury, FATF, Reuters, major think-tanks) block automated retrieval in CLI and agent environments. If a URL fetch fails:

1. Do not fabricate the document content.
2. Mark the source as [primary][verify] or [secondary][verify] based on what it would be if accessed.
3. Note the retrieval failure explicitly in the source record under "Source limitations."
4. Downgrade the downstream memo's evidence mode to mixed or reasoning_only as appropriate.
5. Include a [verify] flag on every claim derived from that source.

The memo's limitation note must reflect which sources were not accessed.

Output format

Produce a source record in this shape:

## Source record

**Title:** [Full title or best approximation if absent]
**Author / publisher:** [Name, institution, or "Unknown"]
**Date:** [Publication or document date, YYYY-MM-DD or YYYY-MM if only month known]
**Document type:** [official-doc | regulatory-filing | court-record | company-filing | research-report | media-article | think-tank-report | transcript | dataset | other]
**URL or file:** [URL or filename]
**Retrieval / received date:** [YYYY-MM-DD — when you accessed or received it]
**Language:** [Language of the source]

### Axis A classification (SOURCE_POLICY)
**Source type:** [primary | secondary | user-provided]
- `primary` — official document, legal text, regulator statement, sanctions list, court record, company filing, parliamentary record, treaty text, directly accessed.
- `secondary` — think-tank report, media article, analyst commentary, research note.
- `user-provided` — provided by the user in this session, not independently verified.

### Axis B flags
[Add if applicable]
- `[verify]` — if a key claim should be confirmed against the original before operational use.
- `[stale-risk: YYYY-MM]` — if the document is dated and the covered facts may have changed.

### Key claims (extracted)

For each material claim in the document, produce one line:

| Claim | Page / section | Provenance | Action flag |
|---|---|---|---|
| [Quoted or closely paraphrased claim] | [p. N or §N] | [primary / secondary / user-provided / inference] | [verify / stale-risk: YYYY-MM / none] |

Limit to 5–10 claims most relevant to the user's task. Do not list everything.

This table is the accountable claim inventory for the downstream memo. The analysis that consumes this record must account for every row: each claim ends as **used** (woven in, with its provenance tag), **flagged-but-not-used** ("I cannot verify [X]; it is not used"), **conflict-surfaced** (both positions named with provenance), or **out-of-scope** (explicitly excluded, with a one-line reason). Silently dropping a listed claim between this record and the memo is an honesty failure, not a length optimization.

### Key excerpts (quoted)

> "[Direct quote from source, including enough context to be intelligible.]"
> — [Author if named], [Title], [Date], [Page or section if known]

Include 2–5 excerpts that are most likely to be cited in the downstream analysis.

### Source limitations

State what this source does and does not establish:

- **Establishes:** [What factual claims are directly supported by this document.]
- **Does not establish:** [What the document cannot support — scope limits, missing data, what it does not screen or verify.]
- **Potential bias or framing:** [Publisher orientation, advocacy position, or conflict of interest if relevant.]

### Routing

Based on the document's topic, suggest which source-requirements pack and skill references to load next:

- **Source-requirements (Agenda Intelligence MD):** [e.g.,
  `sanctions.json`](../../source-requirements/sanctions.json) or the relevant
  sibling file under `source-requirements/`.
- **Skill reference (Agenda Intelligence MD):** [e.g., the sanctions
  reference](../agenda-intelligence/references/sector/sanctions.md) or the
  relevant sibling file under `../agenda-intelligence/references/`.
- **Vertical source guide:** If the document relates to a specific region, load the relevant vertical skill's source guide for source tiering, freshness horizons, and authoritative URL pointers:
  - Central Asia / Caspian: `docs/source-guide.md` in [central-asia-caspian-hybrid-intelligence-skill](https://github.com/vassiliylakhonin/central-asia-caspian-hybrid-intelligence-skill)
  - Gulf / Middle East: `docs/source-guide.md` in [gulf-middle-east-hybrid-intelligence-skill](https://github.com/vassiliylakhonin/gulf-middle-east-hybrid-intelligence-skill)
- **Evidence-pack JSON mode:** one of `live_source_backed`, `user_provided`,
  `reasoning_only`, or `mixed`, as defined by
  [`SOURCE_POLICY.md`](../../SOURCE_POLICY.md).
- **Agenda-memo JSON mode:** one of `reasoning_only`, `user_provided`, or
  `mixed`, as defined by
  [`agenda-memo.schema.json`](../../schemas/v1/agenda-memo.schema.json). The
  product shell does not accept `live_source_backed`; sources retrieved upstream
  are supplied to it by the caller and therefore map to `user_provided` or
  `mixed`.