Documentation: accurate README + full /docs tree

The README described intent; the repo is now real. Rewrote it to reflect
what actually exists — 17,506 validated files across four entity types,
the deterministic pipeline, the commands — and added the /docs tree:

  government-as-code.md  the philosophy: git primitives -> government
  data-model.md          entity types, IDs, file format, the tree, determinism
  sources.md             where every fact comes from, and how current
  roadmap.md             done / next / later, honest about what isn't built
  contributing.md        the disciplines, and how to add a source

Plus data/sources/sources.yaml — the machine-readable source registry.

Every count, path, and source claim is grounded in the current tree, not
the aspirational brief. Internal links verified; validation still clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/sources.md

@@ -0,0 +1,41 @@
+# Sources
+
+Republic OS is a mirror. Every fact in it comes from a public source, and every record carries field-level provenance and a confidence level. This page lists the sources feeding the mirror today; the machine-readable registry is [`data/sources/sources.yaml`](../data/sources/sources.yaml).
+
+The pipeline is **repo-canonical**: raw source exports land in `data/*.jsonl` untouched, and [`scripts/build.py`](../scripts/build.py) generates the entity tree from them in one deterministic pass. Nothing is hand-edited into the tree.
+
+## Active sources
+
+| Source | Feeds | Raw file | Authority |
+|---|---|---|---|
+| **Atlas / influence.tools** | Officeholders — federal, all state legislatures, FL county + municipal | `officeholders-v2.jsonl` (11,285) | Aggregator |
+| **congress-legislators** ([unitedstates project](https://github.com/unitedstates/congress-legislators)) | Bodies, leadership roles, committee memberships | `bodies.jsonl` (233), `leadership.jsonl` (28), `committee_memberships.jsonl` (3,879) | Community, authoritative |
+| **FEC** (Federal Election Commission) | 2026 federal candidate filings | `fec_candidates.jsonl` (2,494) | Official |
+| **Census ACS 2023** | County & congressional-district demographics | `acs_county.jsonl` (3,231), `acs_cd.jsonl` (363) | Official |
+
+Atlas itself draws from official upstreams — House Clerk XML, Senate.gov XML, OpenStates, Census TIGER/Line geometry, and the Florida Cities partner API — and records the specific upstream in each record's `sources` block. So a Person file's provenance names not just "Atlas" but the document behind the field (e.g. `tenure: House Clerk XML 2026-06-10`).
+
+## Provenance in every record
+
+Each entity's frontmatter carries a `sources` list and a `confidence` level:
+
+```yaml
+sources:
+  - field: tenure
+    source: "House Clerk XML 2026-06-10"
+  - field: roles
+    source: "congress-legislators (unitedstates project)"
+confidence: official
+```
+
+`confidence` is one of `official`, `reported`, `inferred`, or `unverified`. Everything in the mirror today is `official`.
+
+## Reconciliation and its findings
+
+Sources use different keys — Atlas uses internal UUIDs, congress-legislators uses bioguide IDs, FEC uses committee IDs, ACS uses FIPS codes. Where no shared key exists, records are reconciled by normalized name. Federal officeholders reconcile at ~99% (531 of 532 members matched to their leadership and committee data).
+
+Disagreements between sources are not errors to hide but **findings to surface**. Where Atlas's officeholder roster names a different person for a seat than the authoritative current roster does, the mismatch stands as a visible discrepancy — the gap between the model and reality is itself information.
+
+## Known sources not yet ingested
+
+Named in the [technical brief](../Technical%20Brief%20—%20Government%20as%20Software.md) and [roadmap](roadmap.md), not yet in the mirror: campaign finance flows and contracts (the "money" layer), local agenda portals (Legistar / Granicus / CivicPlus), the Federal Register, lobbying registrations, and property records. These are where the mirror grows next.