Five months, five versions, and 1,017 commits — from an empty repo and a list of ambitions to a system that knows what it can prove and what it can't. The shape of that journey, in five chapters and two failures.
An empty repo, a thesis, and a list of things the system was supposed to one day do. Twenty agents were specified; six were ever implemented. A 14-category JV taxonomy was sketched. Three deals — Chaparral, Vibe, Invesco — were drawn into early analysis to see what the prose actually wanted to be.
The lesson of v0 was that you cannot specify your way out of a problem you do not yet understand. Most of what was written here was discarded; what survived was the instinct that structure had to come before tooling.
The schema was the work, and the work was the schema — twelve-plus iterations over thirty days, each one refining what a "provision," a "position," a "feedback record" actually had to mean for an attorney to trust it.
By the end of v1 the system had a vocabulary — and could read its own output back to itself without flinching. That was the point at which it stopped being a sketch.
The Constitutional Compliance Audit landed on December 23, 2025 and the system declared itself NOT READY for AI-operated status. Six constitutional risks were named. Four security gates were stood up. One hundred-plus files were renamed in a single day to enforce the new naming convention.
v2 was the month the project learned to refuse its own output. The audit became a tool the system ran on itself, not a ceremony performed at the end.
Of 693 clauses across two versions of the Invesco JV, 93 actually changed; 84 received a direction call (added / removed / modified / strengthened / weakened); 25 were attorney-verified by Andrea against her own prior judgment.
This was the month the difference between "the LLM said something" and "the system can show you why" became visible in the data, not just in the architecture diagrams.
The legal-logic schema — five provision types, eight enums, fourteen tables — was validated against five hand-extracted provisions and approved by Andrea on a structural review (six of seven items agreed, one to discuss). The extraction pipeline (A1–A3: context assembler, validator, orchestrator) was wired and is what powers Phase 4 today.
v4 is short, on purpose. What the system can do today is the smallest honest claim the architecture supports.
Six constitutional risks. Four gaps found. ~one month of remediation work queued. The audit was supposed to bless the launch. Instead it stopped it. The lesson: build the audit before you trust what it grades.
Ten direction calls were fully inverted; twenty-two were partially mis-attributed. ~two weeks of remediation, an entity map rebuild, and a full pipeline re-run. The lesson: "the LLM read it" is not a citation, and never was.
After twenty deals through Phase 6, the system will be able to do for a new attorney what it now does only for Andrea on Invesco: hand them the firm's accumulated judgment, by topic, with citations that hold up. The first twenty are the project. Everything after that is the product.