# Responsible America — Source Policy

Status: internal operating procedure for anyone finding, citing, or
maintaining source records on this platform.

Last reviewed: 2026-07-30.

## 1. The source hierarchy

Sources are ranked in the following fixed order, most to least
authoritative. A source's `source_type` field (`source.schema.json`)
must be one of these 15 values:

1. **Constitution and official constitutional materials** (`constitution`)
2. **Statutes and regulations** (`statute`, `regulation`)
3. **Court opinions** (`court_opinion`)
4. **Official government datasets** (`official_government_dataset`)
5. **GAO** — Government Accountability Office reports (`gao`)
6. **CBO** — Congressional Budget Office analysis (`cbo`)
7. **CRS** — Congressional Research Service reports (`crs`)
8. **Inspectors General** (`inspector_general`)
9. **Federal and state agencies** (`federal_or_state_agency`)
10. **Legislative records** (`legislative_record`)
11. **Peer-reviewed research** (`peer_reviewed_research`)
12. **Established nonpartisan research institutions** (`nonpartisan_research_institution`)
13. **Reputable journalism** (`journalism`)
14. **Advocacy organizations representing competing positions** (`advocacy_organization`)
15. **Public commentary and social media**, used only as evidence of
    public reaction, never as evidence of fact (`public_commentary_or_social_media`)

Higher-tier sources take precedence when they conflict with a
lower-tier source on a factual question. A tier-13 news article
describing what a statute says is not used in place of the statute
itself once tier-2 text is available.

## 2. Original sources over secondary description

When both are available, the original document is cited and used —
the actual bill text, the actual court opinion, the actual dataset —
rather than a news article or think-tank summary describing it.
Journalism and secondary analysis are used to establish context,
significance, and framing, and to represent perspectives an original
document does not itself state (e.g., how a ruling was received). They
are not used as a substitute for reading the primary document when the
primary document is publicly available.

## 3. Controversial claims and competing interpretations

For a controversial claim — one where credible, informed people
genuinely disagree on interpretation, not just on politics — the
sourcing includes credible sources representing the major competing
interpretations, drawn preferentially from tiers 9–14 rather than tier
15. This is what allows a claim to be marked `disputed` honestly
(`docs/fact-checking-policy.md`) rather than resolved by citing only
the side that was easiest to find.

## 4. What every source record must capture

Per `data/schemas/source.schema.json`, every source record includes:

- `source_id` (pattern `src-[a-z0-9-]+`)
- `title`, `publisher`, `author` (nullable)
- `publication_date` and, where relevant, a separate `event_date` —
  the date of the underlying event or data point, which is not always
  the same as when it was published
- `accessed_date` — when Responsible America actually retrieved it
- `url`, and `archive_url` where an archived copy has been captured
- `source_type` — one of the 15 tiers above
- `credibility_notes` — anything a reviewer should know about the
  source's reliability, funding, or limitations
- `relevant_excerpts` — the actual text being relied on, not just a
  link
- `related_reform_ids` — which reform_id(s) cite this source (named
  `applicable_reforms` in this list until a Phase 6 review found the
  field was never actually named that in `source.schema.json` — the same
  kind of stale-name correction as section 6 below, not a schema change)
- `is_current` — false once superseded
- `superseded_by` — the source_id of the record that supersedes this
  one, when `is_current` is false

**A source record is never created for a document that has not
actually been retrieved and reviewed.** Listing a source that a
researcher merely believes exists, has not opened, or is inferring
from a citation in another document is not permitted — this applies
equally to human researchers and to AI-assisted research (see
`docs/ai-usage-policy.md`).

## 5. Link checking and staleness

- **Broken links.** Source URLs are checked periodically. A broken
  link is flagged for review, and an `archive_url` is used or added
  where available so the underlying document remains verifiable even
  if the live link fails.
- **Statistics past their review date.** Claim records carry
  `as_of_date` and `last_verified`; when a claim's underlying figure is
  past a reasonable review window for its `claim_type` (e.g., annual
  budget figures reviewed at least yearly), it is flagged for
  re-verification rather than left to quietly age.
- **Never replace verified data solely because a newer page exists.**
  A newer webpage on the same topic is not, by itself, grounds to
  replace a verified figure. The replacement happens only when the new
  document is actually authoritative for that fact and genuinely
  supersedes the old one (e.g., a new official dataset release
  replacing last year's release). When that happens, the old source
  record's `is_current` is set to `false` and `superseded_by` points to
  the new source's `source_id` — the old record stays in place and
  queryable, not deleted.

## 6. Correction (2026-07-31, Phase 4): field names updated

The `is_current` boolean and singular `superseded_by` field described
in section 5 above were replaced in Phase 3 by a richer
`current_status` enum (`current` / `superseded` / `historical_reference`)
and array fields `supersedes_source_ids` / `superseded_by_source_ids` —
see `data/schemas/source.schema.json`. Section 5's *behavior* is
unchanged (the old record stays in place, not deleted); only the field
names were current as of Phase 1/2 when originally written. Left the
original wording above rather than silently editing it, per this
project's no-silent-history-erasure practice — see `docs/change-log.md`.

## 7. Deduplication (Phase 4)

A duplicate is two source records describing the *same underlying
document* — not two records that merely share a title, a topic, or (for
regulations) a common enabling-statute citation. Different editions,
revised reports, amended statutes, separate docket documents, majority
opinions, concurrences, dissents, appendices, and archived versions are
each their own record.

When a genuine duplicate is found:

1. Compare metadata (citation, docket, publisher, retrieval method) to
   confirm it really is the same document, not a related-but-distinct one.
2. Pick (or create) one canonical record — prefer the one with the
   higher `retrieval_level` (see below) when they differ.
3. On the canonical record, list the other's `source_id` in
   `merged_source_ids`.
4. On the merged-away record, set `duplicate_of_source_id` and
   `canonical_source_id` to the canonical `source_id`. **The merged-away
   record is never deleted** — it stays in place for provenance, exactly
   like a superseded record under section 5.
5. Redirect every claim/reform/court-case cross-reference that pointed
   to the merged-away id so it points to the canonical id instead, and
   union `related_reform_ids`/`related_claim_ids` onto the canonical
   record so no relationship is lost.
6. Document the merge in `docs/change-log.md`.

`ra-research-validate.mjs` reports any citation shared by more than one
source record that has *not* been resolved this way as a review item.

## 8. Retrieval confidence (Phase 4)

Every source record's `retrieval_level` field records how its content
was actually obtained, from `metadata_confirmed` (weakest) through
`search_result_confirmed`, `authoritative_document_located`,
`authoritative_document_retrieved`, `authoritative_document_reviewed`,
to `cross_checked` (strongest) — see the field's full description in
`data/schemas/source.schema.json`. Locating a search result is never
treated as equivalent to reading the underlying document. No claim may
be marked `verification_status: verified` on the strength of a source
at `search_result_confirmed` or below — see
`docs/fact-checking-policy.md`. Sources still below
`authoritative_document_reviewed` are tracked in
`research/queues/primary-document-review.json` for future direct
retrieval.

## 9. Source snapshot and archival policy (Phase 6)

Some official sources this platform cites are genuinely unstable —
agency press releases get reorganized, a rule's status page gets
overwritten in place rather than versioned (see reg-013's own
`historical_versions` field for a rule that has already changed legal
status once), a state attorney general's enforcement-action page
disappears after a settlement. This section governs when and how the
project may preserve a copy of something it cites, distinct from simply
recording a URL.

### 9.1 What may be preserved

Only these categories, and only when the specific document is actually
cited by a claim or source record:

- an official downloaded PDF (a statute, regulation, court opinion,
  agency report, or other public-domain or officially-released
  government document);
- an official XML or plain-text version of the same;
- an agency data file (e.g. a CSV/XLSX dataset release);
- a court opinion (public record, not copyrighted);
- any other public-domain U.S. government document (17 U.S.C. § 105 —
  works of the U.S. federal government are not subject to copyright
  domestically);
- failing an archivable original, a checksum, an archive-service URL
  (e.g. a Wayback Machine snapshot), and full bibliographic metadata
  only, so the citation remains verifiable even if the original page
  disappears.

### 9.2 What must never be preserved

- copyrighted secondary-source publications (news articles, paywalled
  reports, think-tank PDFs not released as public domain) — copy the
  citation and a short `relevant_excerpts` quotation under fair-use
  norms, never the full document, regardless of whether the URL might
  later disappear;
- leaked, hacked, or otherwise unlawfully obtained material of any kind,
  even if it would be convenient evidence — see docs/security-and-privacy.md
  and section 17 of the Phase 6 instruction;
- classified or export-controlled material;
- material where source authenticity cannot be established — mark
  `retrieval_level` no higher than `search_result_confirmed` and keep it
  out of any archived-copy workflow entirely, not just out of
  public-data;
- personally sensitive material — private individuals' identifying
  information, medical records, home addresses, or similar, regardless
  of the source category it would otherwise fall into.

### 9.3 Required metadata for any preserved file

A source record for a preserved (not merely linked) document must
additionally capture, either in existing schema fields or a future
`archived_copy` extension (not yet built — no file has been archived
under this policy as of Phase 6, since WebFetch has been blocked for
every domain tested across every phase, making direct retrieval and
therefore archival moot in practice so far):

```
source_id
original_url
retrieved_at
content_type
file_size
sha256
issuing_body
public_domain_status
license_status
repository_path
redaction_status
```

`redaction_status` matters specifically for reform 018 (defense audit)
and reform 019 (secrecy/UAP) sources, several of which are themselves
officially redacted documents — the redaction is part of the source's
authentic content, not something this project adds or removes, and must
be recorded so a reader knows the preserved copy is exactly as
officially released, not a project edit.

### 9.4 Current state

No file has actually been archived under this policy. WebFetch has
returned HTTP 403 for every external domain tested across every research
phase of this project, including fresh, direct re-tests during Phase 6
(four representative high-priority documents: a Supreme Court opinion,
a Federal Register final rule, a GAO report, and a justia.com case-law
mirror) and Phase 7 (three further attempts: a second justia.com
Supreme Court opinion, an eCFR regulation, and the U.S. House's own
uscode.house.gov statute mirror) — see
`research/queues/primary-document-review.json`'s `last_attempted:
"2026-07-31"` entries and `docs/citation-format-policy.md`. This policy
exists so that if and when direct retrieval becomes technically
available (a different environment, a future session, or an explicitly
authorized manual process), preservation
happens under clear, pre-agreed rules rather than being improvised in the
moment.