Crossfoot

An arithmetic-reconciled corpus of public statistical tables: each messy PDF/HTML table is transcribed into a typed .cells.json whose own internal arithmetic — row totals, subtotals, percent closures, balance identities — must reconcile exactly under reconcile.py. The source's redundancy is the proof: a transcription typo almost always breaks a cross-foot, so correctness is derived, not trusted.

// a cell, and a relation the reconciler re-derives from leaf cells
{ "id": "r15c1", "row": 15, "col": 1, "value": "6274791", "role": "total" }
{ "type": "sum", "sources": ["r1c1", "...", "r14c1"], "target": "r15c1",
  "tol": "2", "why": "Table 8 prints: 'Details may not add to totals due to rounding.'" }

All arithmetic runs in exact Decimal — never floats. Values are decimal strings exactly as printed. A tolerance is only legal when the source's own rounding note is quoted in why; strict coverage (every total targeted, every leaf feeding a relation) is the default gate.

The corpus

501 units, all GREEN (2026-09-17):

family	source	units
treasury-mts	Monthly Treasury Statement, May, June, July + August FY2026 complete (Tables 1–9 incl. full Table 5 outlay detail and Table 6 Schedules A–E)	325
census-p60	Census P60-282 Income in the United States: 2023 — the complete appendix (A-1…A-7, B-1…B-5)	91
omb	OMB FY2027 Budget Appendix, Legislative Branch chapter — complete (every account with a printed schedule)	55
sec-10k	Apple FY2023 + Microsoft FY2025 10-K statement sets (complete: balance sheet, income, comprehensive income, cash flows, equity, parentheticals)	12
fec	Official 2024 presidential general election results — complete	11
bls-cpi	CPI relative importance of components, Dec 2024 — complete	7

Per-unit specs live in BACKLOG.md, the authoritative manifest.

The corpus is US sources only (US government publications and SEC public filings). Non-US families are out of scope.

Every vendored source in sources/ is sha256-ledgered with provenance in sources/SOURCES.md — transcribed verbatim, never edited.

Process

The corpus is transcribed and audited by a rotating fleet of AI agents — 29 model identities to date, credited in CONTRIBUTORS.md — under a fixed discipline:

• one session = one unit = one new file, gated by reconcile.py GREEN with zero coverage warnings (AGENTS.md has the rules; RUNBOOK-transcriber.md the session flow);
• every 10th unit gets a different-agent spot-audit against the source (RUNBOOK-auditor.md); the full audit trail, including the defects the cadence caught, is in AUDITS.md;
• PDF units are render-anchored (a vision model must look at the page, not just the text layer); HTML units run text-only.

Running it

uv venv && uv sync                                     # once
uv run python reconcile.py tables/<family>/<id>.cells.json
uv run pytest                                          # oracle self-tests
uv run python reconcile.py --all                       # full sweep (empty selection is red; any unit failure is red)

License

MIT for code and transcriptions (see LICENSE). sources/ contains vendored public documents (US government / SEC public filings) with per-file provenance in sources/SOURCES.md.