Chapter 03 · The knowledge
What it has read, reconciled, and can answer from.
A self-maintaining second brain that compiled an entire working record into one curated, provenance-backed, contradiction-checked asset.
I decide against reconciled evidence, not memory.
Every claim traces to its source; disagreements are logged, not smoothed. The pattern is public; running it at real-world scale with audit discipline is the contribution.
How the first brain was built
1. Curationstage
A full workspace scan produced a manifest of knowledge-bearing files, with tooling and media excluded by rule.
2. Copystage
Resumable bulk copy. Lesson: cloud-only placeholders had to download in stages, so the copy was made re-runnable.
3. Extractionstage
Every binary document converted to text with ordinary code, parallelised, with per-file timeouts and OCR for scans. Compute, not model spend.
4. The exhaustive readstage
Eleven parallel reader agents read the whole corpus, each producing a dense, provenance-tagged digest. The real model spend sits here.
5. Reconciliationstage
The digests merged into 44 cross-linked pages, every page sourced, with 21 contradictions logged. An adversarial review by a second model family found and fixed 14 issues.
6. Hardeningstage
Two incidents became architecture: a personal-data exclusion rule, and a flat path-safe layout after a path-length failure (longest path 454 to 148 characters).
The contradictions register
The page that catches what no human reader of thousands of documents could hold in their head. Genericised examples: a readiness criterion recorded as met when the record shows it was waived with compensating controls; a termination date recorded inconsistently across documents; a dispute where two parties gave incompatible accounts in the same meeting. For an audit or a board question, this page is the difference between an archive and an asset.