How the manuscript audits are run and tracked
Every audit in this research follows one short routine, every source is pinned to an exact version, and two reviewers work through the collection from opposite ends. On 19 October an audit of the records themselves found that the exact completion percentage is unknown.
A run, one sitting of research, follows a routine written in the project's method file: check both indexes and the latest log; record the precise question and primary source; read the actual source before writing any statement; separate verified content from analysis from speculation; log the result; and use only disposable synthetic data for experiments. The last rule on the list says the Authority System Suite integration is pending, so nothing may claim compliance with it. The Suite is my set of rulebooks that make an AI show its work, and every entry ends “Authority Systems Suite unverified” for that reason.
Every audit cites one fixed version of the source
The source collection is mutable. On 14 October it described 719 manuscripts in 372 families (groups of related manuscripts), and every audit since cites one commit, openai/math@1581d4c…, plus the Git blob of the exact PDF. The blob is a fingerprint: if the file changes, the fingerprint stops matching. The status line then says what was and was not read, in fixed words: source audited, PDF pages pending, theorem unverified, never “fully reviewed” until the printed pages are read.
Two reviewers work from opposite ends so they never overlap
Two tracks work through the families from opposite ends so they never duplicate each other. Before starting, a run writes a claim file, a short note naming its family and where the other track last was.
One reverse-track claim says: “Forward track latest observed 024-A, next 025-A; no family overlap as of this claim.” A forward-track claim adds: “Claim is a coordination record, not a full review.”
Some early reports were blocked and only saved later
Several early entries say a report or ledger write was blocked, with a “recovery handoff retained” locally. One audit on 14 October is an example. It enumerated every Boolean function on up to three variables at four noise levels, 1,104 cases for each of two measures with zero violations, but the report could not be created.
Its full report landed on 20 October, with 736,886 assertions passed and exhaustive checks through four variables. It includes a control: two noisy copies of the same secret bit reveal 0.3319 bits where the one-release bound is 0.1887, which shows the theorem's independence assumption is needed.
The completion percentage is unknown
On 19 October a coverage audit checked the repository against the official catalogue: 372 families and 719 manuscripts now, from an initial release count of 722 with three withdrawn. It found that the coverage ledger is a narrative rather than a machine-readable census, and that some reports exist only in local archives.
It established no family as fully independently proof-verified. On completion it states: “Exact completion percentage: unknown, not 125/372.”
Its corrective list starts with an immutable snapshot of the catalogue and a status per family: not started, triaged, source reviewed, proof checked, independently verified. My open-questions file still lists the exact collection and the 6 October release chronology as not confirmed from primary sources, and the later entries have not closed that.
Sources and notes (9)
- openai/math — the pinned source collection
- openai/math: CONTENTS.md — the catalogue of manuscripts and families
- Method — the notes this was written from
- Open questions — the notes this was written from
- Coverage audit, 19 Oct — the notes this was written from
- COMP-119 checkpoint — the notes this was written from
- Claim REV-346A — the notes this was written from
- Claim FWD-039C — the notes this was written from
- Run log — the notes this was written from
Cryptography researchAll 10 parts →
