Scattered public records.
One traceable answer.

One buyer, one output, and a test that could kill it.

The test I set

The plan was to ask one lender for its last 50 funded Florida restaurant files, run the receipt on each, and sort the results into three piles.

1. Pile 1Silent wrong bind. The lender funded company A, but the food or alcohol license at that address belongs to company B.
2. Pile 2Refuse. A famous name on the wrong block.
3. Pile 3Split. Two plausible shells, so do not pick.

If pile 1 is empty, walk. That is a matcher, not a must-buy.

From a government row to one cited answer.

Every ingested row becomes an observation tied to its source record and URL. Identifiers decide identity first; names and addresses decide only behind guards. When the evidence is close, the resolver says so instead of guessing.

Measured on rows nobody had seen.

Blinded customer‑path test, July 15 to 17

Fresh restaurants were drawn from the state’s live food-license extracts at run time. Each upload row carried only what a lender collects: legal name, trade name, street, city and ZIP. License numbers were withheld and kept as ground truth. Rows went through the real product path, from sign-in and CSV upload to the full run, and the analysis used a read-only connection.3

auto-matches correct

215 of 215

rows sent to review or left unresolved

35 of 250 (14%)

wrong in three re-runs after fixes: 83 of 83, 67 of 67, 72 of 72

0

Where the hard rows went. None auto-attached. Every wrong candidate sat in review: an inn’s food-service row matched to the property’s bed-and-breakfast lodging license, a venue stand matched to its neighbor with a near-identical license number, and two food trucks matched to a different truck at the same commissary. All three unresolved rows in run 2 were non-seating licenses, catering or kitchen-only, with a thin public footprint.

What this does not prove. No named person reviewed the books; that gate stayed at 0 of 150 rows, and the report says an automated workflow cannot honestly close it. Ground truth was the withheld license number. On the same three books an older full-book matcher had scored 115 of 116; its one false positive is case 8 above.

Deterministic baseline against an agent, July 20

50 real restaurant submissions, each with three real candidates and an upstream pick chosen to mislead: 20 wrong picks, 5 stale predecessors, 15 correct picks among close candidates, and 10 rows whose true entity was withheld to test abstention. The baseline keeps the upstream pick. It is not the production resolver, so its 30% is not the lookup’s score; it shows what accepting an upstream candidate costs.4

The agent bought precision with time and caution: 237 times slower, and it abstained on 14 of the 40 rows that had a true answer. The benchmark states its own limits: the slates were composed, not replayed from matcher logs, and it is not a blinded human baseline. The cut that followed kept models out of identity: “Do not let an LLM write canonical identity.”

Gates on every pull request

• 48

  Resolver precision cases, 29 that must match and 19 that must not: co-tenants at one address, a shared FEIN, a shared state document, one phone across different names, weak-name near ties, generic names in one ZIP, and trade name against legal name. The test asserts zero false positives.

• 23

  Labeled merge pairs, 10 merge and 13 must not, most negatives taken from real production dry runs. In the fixture’s words: “Zero false positives is a hard CI gate.”

• 130

  Rows in a regression set run against production on demand rather than in CI: all 80 license-graded rows correct with 0 false positives, and all 8 abstention traps abstained. It mostly tests exact-license lookup, so the blinded run above is the real test.

Two failures, and what they changed.

Both were mine to own. Each write-up covers what happened, why, how it was found, the fix and how the fix was verified.

I stopped when the test said walk.

The engineering held up; the market case didn’t. I pitched lenders and insurance carriers, and none signed on, so I ran my own backtest to find out whether they were missing something.

A statewide loader for secured-loan (UCC) filings bound each debtor to a food- or alcohol-licensed business through the same fail-closed resolver. A read-only analysis raised flags from the past year of filings for 8 lenders, and a verifier checked them.

lenders in my own backtest

8

spot-checked flags that held up

10 of 10

flags a cheap per-borrower search would not have found

1 of 36

The flags held, but only 1 of 36 was something a lender couldn’t get from a cheap per-borrower UCC search. Meanwhile, keeping about 60 sources fresh as one person was a full-time job on its own; Every check green, no data shows what that upkeep looks like.

No buyer signed on, the signal was rarely unique, and that upkeep sat behind it. Under the kill test I had written in August, that meant walk. I closed the line on September 27, 2026.7

What else I cut

• MCP. I built a 13-tool local MCP server and a hosted endpoint limited to three tools, which stayed dark behind a rollout flag. The MCP surfaces saw 17 production calls, so I cut them. The route now returns 410 Gone.
• Sediment. I unmounted the book product, the autonomous diligence agent, the risk file and MCP, and each route now returns 410. On August 22 one pull request deleted the local MCP package and the other dead surfaces. The README calls what was left “sediment. Do not extend them.”
• A rule for the agents. If a task would add an MCP, book-review, autonomous, diligence, risk-file or case-file surface and is not a deletion, stop and point at the product page.

I directed the agents that wrote it.

I founded EarlyCo. Claude Code and Codex wrote the code under my direction, with some Cursor early on, and a contract engineer helped set up the initial database in March. I scoped the work, wrote the product lock and the agent briefings, designed the evaluations and their ground truth, reviewed and merged through pull requests, approved production writes as bounded runs, and decided what shipped and what was cut.

commits on main, March 11 to September 2, 20268

2,110

merged as numbered pull requests

1,931

test files

767

SQL migrations

144

commits on main co-authored by Claude

675

dated Codex entries in the handoff ledger

132

How the agents were kept honest

• One product authority. Every agent briefing pointed at the product page as the only product authority, and the cut list carried its own rule for agents.
• Gates before merge. Main required three status checks and rejected direct pushes. Every pull request ran the static checks and the unit and integration suites, precision gates included.
• Dry run by default. Production write scripts for repair, rollback, re-linking and merging do nothing without an explicit apply flag. Production writes ran as bounded runs, verified by separate read-only runs.
• Schedules on an allowlist. A workflow-hygiene test rejects any autonomous schedule that is not explicitly allowed.
• Agent errors caught by design. The language-model cluster verifier counted a shared address as corroboration, so it “could re-merge exactly the pairs the guard protects.” Its prompt now states license hand-offs outright, so it cannot undo the code guard.

What I would do differently

1. Run the sales test before building surfaces. I wrote “Do this before adding surfaces” in August, after most of them existed. Before that lock the product had pointed at vendor prospecting, insurance underwriting data, insurance books and lender credit.
2. Keep one governed matcher. The worst incident came from a second matcher that hadn’t been through the same checks. Name containment should never decide identity on its own.
3. Treat built and running as different states. A job counts once it runs on a schedule and something checks that it delivered.
4. Keep raw bytes from the first ingest. State feeds are rolling windows. A gap that outlives the window is permanent.
5. Scope every accuracy claim to its evidence. A July copy audit flagged “zero false-cleans” as true only of one dated 15-restaurant batch. Every number on this page carries its source and date for that reason.