Package 01Collection + reconciliation
Codex collection and reconciliation sources
Original-project API collector plus portable derivatives of the accepted 26-year reconciliation and independent validator.
Open replication/codex on GitHub ↗Replication record · candidate-publication layer
Collection, independent reconstruction, and exact-set corroboration.
This page documents how the broad Federal Register publication universe was collected through the public API, repeatedly checked with Codex, rebuilt in a procedurally count-blind Claude Code run, and compared only after the fresh union was sealed.
- 788,470
- Distinct publication identities
- 26/26
- Annual partitions passed
- Fourth
- Complete Codex check in the project summary Ordinal reported by the project log
- 26
- Isolated Claude Code annual sessions
- 0 ↔ 0
- Exact set differences after sealing
- PASS
- Candidate-identity layer
What is established
A verified identity universe for one documented source vintage.
Both systems produced the same exact set under the identity rule publication_date|document_number. The comparison found no differing annual count and no identity present in only one system.
This closes the candidate-publication identity layer. It does not validate sanctions classification, the routed-publication layer, the future country–product index, or the legal meaning of any publication.
Method overview · FederalRegister.gov API
Bounded enumeration with retained request evidence.
Endpoint: /api/v1/documents.json · no API key · order=oldest · per_page=1000 · no substantive filter in the all-publication inventory.
- 01
Partition by month
For each year, issue twelve non-overlapping date-only queries with inclusive calendar-month bounds.
- 02
Enumerate API pages
Request up to 1,000 records oldest-first and follow the exact same-host next_page_url returned by the accepted page.
- 03
Preserve transport evidence
Retain response bodies, headers, URLs, timestamps, byte counts, SHA-256 values, validation errors, and failed retries.
- 04
Construct exact identities
Key each publication as publication_date|document_number; validate nonblank fields, ISO dates, and year boundaries.
- 05
Reopen from disk
A separate verifier reads the retained bodies rather than accepting the collector’s in-memory result.
- 06
Freeze, then compare
Build and validate the 26-year union before consulting the historical result or comparing the two systems.
This six-step summary describes the clean live-API replication. The original Codex collector also retains broader agency and term queries for downstream routing, plus an annual inventory query; those auxiliary searches do not redefine the all-publication identity denominator.
Figure 01 · Source corpus
Annual coverage of the candidate publication universe, 2000–2025
788,470publication records
Federal Register publication records retained before sanctions-related screening. Bar lengths share a 0–34,000 scale; exact annual counts are printed.
Execution history
What ran, in what order, and what “independent” means here.
Independence is computational and procedural: a different model, isolated annual sessions, a separate implementation, and a post-freeze comparison. Both systems necessarily share the same public API source and the declared identity definition.
- 01Reported run history
Repeated Codex validation
The project summary identifies the accepted result as the fourth complete 26-year check.
That ordinal comes from the project run history. It is not inferred from the final attempt directory alone.
- 02PASS
Accepted Codex reconciliation
Attempt 005 passed after four earlier aggregate implementations were rejected.
The rejected implementations failed runtime, discovery, or schema-normalization design gates. They are preserved as method-development history—not counted as four successful replications.
- 03PASS
Count-blind Claude Code run
Twenty-six isolated annual sessions rebuilt the universe from the live API.
No target or prior count was consulted until the fresh union was sealed and independently validated. The sessions shared one frozen method and package.
- 04Exact set match
Post-seal corroboration
A neutral parser compared the identity sets, not merely their totals.
All annual counts matched. The symmetric difference was empty in both directions, and a common identity-only serialization produced one shared digest.
Verification ledger
Different operations, one comparison boundary.
The Codex path tests retained-capture custody and consistency; the Claude Code path performs a fresh source enumeration.
| Control | Codex | Claude Code |
|---|---|---|
| Primary operation | Reconcile 26 retained annual identity sets against three active file families | Enumerate 26 years afresh from the FederalRegister.gov public API |
| Execution architecture | Repeated project checks; accepted aggregate reconciliation attempt 005 | 26 isolated annual sessions using one frozen method |
| Checksum evidence | 5,349 manifest checks; zero failures | 2,323 accepted annual checks; zero failures |
| File and identity controls | 78/78 active files unchanged; zero identity defects | Zero blank, duplicate, contradictory, or wrong-year identities |
| Final independent gate | Independent final validator: PASS | Independent aggregate validator: PASS |
| Pre-comparison boundary | Aggregate frozen before the historical comparator was supplied | Target or prior count not consulted until the union was sealed and validated |
Set-level result
Equal totals, equal annual counts, equal identities.
Comparison began only after the Claude Code union was sealed and its independent aggregate validator returned PASS.
| Test | Codex | Claude Code | Difference |
|---|---|---|---|
| Distinct identities | 788,470 | 788,470 | 0 |
| Annual counts | 26 values | 26 values | 0 differing |
| Duplicate row excess | 0 | 0 | 0 |
| Codex-only identities | — | — | 0 |
| Claude-only identities | — | — | 0 |
Representation-independent digest
72a011164d60c4d0ddc78a1de11bc9dd6598a78148f1ec617349821df3659943UTF-8; sorted exact identities; one identity per line; LF after every line. The two original union-file hashes differ because Codex and Claude Code used different JSON structures; that file-level difference is not a set difference.
Repository packages
Source is stored in Git, not hidden behind a download form.
Repository access follows its current GitHub visibility. Public access begins only when maintainers deliberately open the repository; that source-publication gate is separate from release of the dataset.
Package 02Live-API replication
Claude Code count-blind live-API replication
Portable live-API collector, disk verifier, aggregator, independent aggregate validator, 26 task prompts, and execution-source manifest.
Open replication/claude-code on GitHub ↗Claude annual checksum arithmetic ↗Set-comparison code ↗Machine-readable comparison receipt ↗Machine-readable page record →
Comparison receipt SHA-256: 5a7553991d54ea55333c56a5f003f359ec57ed2e2fa83313bbbb8393f8c684b4
Disclosed repair
The Claude collector was patched before the accepted runs.
A one-line repair supplied saved response headers to retry bookkeeping after a KeyError: 'headers'. Pre-fix SHA-256: 444421c34061b69771c9c884f91f3ad35a15d36c7f46947cd35e28f34c8f4ba9. Accepted SHA-256: c5ba487a903658b23cac43d88650eeb8e7cc5371e38d668516ee4bdc911c20b4. Query construction, pagination, identity formation, and verification were unchanged. Pre-fix attempts remain preserved in the private research archive and are deliberately excluded from the compact repository package.
Correction to the run narrative
wc -l was not a valid CSV row counter.
An interim wc -l reading was wrong because quoted CSV fields contain embedded newlines. Parsed as CSV, 2013 contains 30,771 rows and 30,771 distinct identities; duplicate excess is zero. The correction was made before this publication record was written.
Limitations and rights
What readers should not infer from this PASS.
- Identity membership, not metadata equivalence
- Codex recorded metadata-drift diagnostics in eight years separately. Those differences did not change membership in the exact publication identity set.
- Not permanently live-complete
- The Codex audit closes the frozen retained capture, and the Claude Code run corroborates it against the API at that run time. A later live API run is a new source vintage and may differ.
- Not legal authority
- FederalRegister.gov is an informational service. Legal research should verify the official edition on GovInfo and the operative source text.
- Not independent human coding
- This is cross-system computational corroboration, not a conventional double-coded human reliability study.
- Not a data release
- The 42+ GiB working archive and large source unions remain excluded. The future finished research dataset and API are separate publication gates.
- Not open source yet
- All project-authored materials remain all rights reserved unless a file states otherwise. Source visibility does not grant an open-source license.