Claude of Record

Construction already names the parties who answer for the design: the Architect of Record, the Engineer of Record. This is the third seat. The party who answers for what the project knows.

Point an agent at every document an owner receives and it builds a cross-referenced wiki it can run real analysis from: change-order triage, allowance hunts, independent budget rebuilds. It is the method behind the project brain I built for an owner's team, written up so it travels to another project.

The method in five lines

1. Sources are immutable and scanned to text. Every figure traces to a source line, or it does not ship.
2. One record per project entity, densely cross-linked. A link to a document that has not arrived yet completes itself when it does.
3. The contract is the constitution. Evidence is tier-ranked, and a party's assertion never becomes a fact by repetition.
4. Deliverables freeze the moment they ship. Corrections are new artifacts, never edits.
5. Validation is structural and semantic, and the agent must state what it verified, what it trusted, and what it did not check.

Why it holds up

A record you cannot trust is worthless, so the repo ships the enforcement layer and not just the idea: structural and semantic audits, an index rebuilder, a reference resolver, a content-hash gate, and a write lock.

Run the validation suite against the included fictional project and it passes. Change a dollar figure to one that does not appear in the source it cites, and it fails.

What's in the repo

• Thirteen short docs on the pattern, the architecture, the constitution, provenance, validation, the honesty rules, ingestion, extraction, analysis, deployment, results, and failure modes.
• Templates for the operating rules and the constitution, plus the config that keeps project facts out of code.
• Schemas for 30 record types, and the wiki procedures and analysis patterns as skills.
• A runnable example. A small fictional corpus you can validate in one command.

If you are skeptical, start with the failure modes: everything that broke, and what each failure taught.

Take it

The repo is public: github.com/prounds-4/claude-of-record. Clone it, run the validation suite against the example, then copy the templates into a workspace and fill them in.

None of the real project's documents are in it. Every name, record ID, and figure in the docs and the example is invented.

The core pattern descends from Andrej Karpathy's LLM Wiki, where the agent reads its sources once and maintains a living, linked knowledge base instead of re-reading the raw documents on every question. The construction layers on top are the contribution.