Nine phases that eliminate the lies a legal document can tell.
Built on a ground-truth foundation (Phase 0), each phase isolates one class of deception — from what the contract says to how the counterparty will react — and resolves it before the next begins.
Built Once, Used Twice
How the Invesco JV work became the engine for the PSA tool.
We didn't build a second system to draft purchase agreements. We took the system that analyzes our joint-venture deals and pointed it the other way — the expensive part was built once, and now does double duty.
| In the JV system | In the PSA system | |
|---|---|---|
| Backbone | provisions → positions |
psa_provisions → psa_positions — identical shape |
| Provision types | 5 partnership-governance types | 15 asset-transfer types — Purchase Price, Deposit / Escrow, Reps & Warranties… |
| Position direction | single-sided | explicit Buyer / Seller toggle |
| Built-in provenance | every row tagged | same — stamped from day one |
The JV speaks the language of partnerships; the PSA speaks the language of buying a building — but the structure underneath is the same. Counts are a point-in-time snapshot, illustrative rather than a live metric.
-
01
We reused the backbone instead of rebuilding it
The JV system stores every provision in a contract, then the firm's position on each one. PSA uses the exact same structure —
provisions → positionsbecamepsa_provisions → psa_positions. PSA work didn't start from a blank page; it started on a backbone already proven across the entire Invesco JV provision set. -
02
It was a deliberate cut, not a lazy copy-paste
We didn't clone the JV database. Its machinery for tracking how a contract changed between draft 1 and draft 8 is essential for analyzing a counterparty's document — and meaningless when drafting one that doesn't exist yet. We kept the reusable spine and dropped the JV-only parts. A structured comparison confirmed it: bolting PSA onto the JV database scored worse than a clean schema that reuses the pattern.
-
03
One small addition did all the directional work
JV analysis asks what does their clause mean, and how do we fight it? PSA drafting asks what should our clause say to protect us? The difference, in the data, comes down to essentially one new field — a Buyer / Seller toggle on each position. The firm's library of negotiating positions, the most expensive thing it owns, gets re-pointed from reading to writing with a single switch.
-
04
The expensive part — trust — transferred for free
The costliest thing in legal AI isn't the analysis. It's the guardrails that make an attorney trust it: an executed contract outranks any AI summary, the AI only proposes while a human approves, and every decision leaves a complete audit trail. All of it was built for the JV and carried into PSA unchanged, because it governs how the system behaves, not which document it's working on.
The one adaptation: with no signed contract to check against while drafting, the PSA side uses the letter of intent plus attorney sign-off as the stand-in source of truth. -
05
The discipline that keeps JV analysis honest keeps PSA drafting safe
The JV pipeline has a hard rule: capture the facts before assigning any meaning. That shaped the PSA build order — for the first PSA deal, the provisions were loaded and verified first; the firm's positions come next, deliberately, rather than letting the model guess and fill in blanks.
— what changed, and is it market?
— which position protects us here?
The deal attorney's judgment — every correction, every preferred position she's given us — is encoded once, and now powers both. That's the whole leverage: the system gets smarter with every deal instead of starting from zero each time the work changes shape.
- The shared data model
- The governance / trust framework
- The evidence-first discipline
- The reused position library
- First PSA deal (Swiss Gables) loaded & verified
- The drafting tool itself
- Assembling clauses into a first draft
- Producing a counter-draft — the benchmark we're driving toward
A foundation that's real, and a drafting capability that's in build. The message is here's how we got the leverage — not the robot lawyer is finished.