About

A diagnosis first. Then a verdict you can sign.

Zoe exists because the numbers pass and the company still fails. Financial diligence is excellent at what it measures. But the thing that kills the deal is how the company actually operates, and that was never measurable in the window a deal gives you.

The origin

The pattern kept repeating. A fund would spend six figures and eight weeks on diligence, sign, and then watch the thesis come apart over the following year for reasons that were visible in the company’s own systems the whole time. The CEO was already burned out. The top three customers were already most of the revenue. Engineering and product had already stopped talking.

None of it was hidden. It just wasn’t measured, because measuring it meant interviewing forty people about their own behavior. Which is exactly the method that produces the least reliable answer.

So Zoe reads the behavior instead, and holds the deck to it. The pitch deck becomes a ledger of claims. The company’s own systems, connected read-only by the target, say which claims hold. That’s the whole idea: evidence over opinion, said versus true, on a 24-hour clock.

How it got built

Six decisions, in order.

No dates on this. The sequence is the point, not the calendar. Each step closed off an easier version of the product that would have been worse.

01
Measure behavior, not answers
The first prototype was a questionnaire. It worked exactly as badly as every other questionnaire: people described the company they wanted to work at. The metadata described the one they did. The one place answers count is the Founder Call, and even there the questions come from gaps in the Claim Ledger and every answer is scored against the record: answered, dodged, or contradicted.
Ruled outA survey product. Faster to build, and structurally incapable of being evidence.
02
Patterns, not paragraphs
Scoring message content would read richer signal. It would also make every authorizer say no, and it would make Zoe a surveillance tool. So in the connected systems the analysis stays on the shape of the activity. Message bodies are never stored, and what is extracted is discarded by default. The boundary is a product decision before it is a privacy one.
Ruled outContent analysis of the connected systems: sentiment scoring, storing or scoring what anyone wrote in Slack, email or tickets. The Founder Call transcript is the one exception, and the founder is on the call for it.
03
Nine dimensions, and never a tenth
The set closed early and has not moved. If something matters, it rolls up into one of the nine. Otherwise every unusual quarter grows a new axis and the score stops being comparable.
Ruled outA configurable dimension set. Sells well in a demo, destroys comparability.
04
A finding is a causal string
“Improve communication” was in an early readout. It is not a finding. “Engineering and Product stopped talking on 15 December” is, because it names the metric, the movement, and the date.
Ruled outAdjectival scoring, letter grades, traffic lights without words.
05
Every number opens to its pull
The evidence trail was going to be a premium feature. Charging for the ability to check the number would have meant selling a claim rather than a measurement.
Ruled outA tiered evidence layer.
06
A verdict only when the evidence earns it
Zoe was asked repeatedly for a buy or walk answer on thin data. She refuses that. When confidence clears the bar she signs an opinion, and the IC memo says what the data supports, what it does not, and what is still open. The decision still belongs to the people who carry the consequence.
Ruled outAn unsigned opinion. A verdict with no confidence level and no evidence appendix behind it.

What we’re building next →

A diagnosis first. Then a verdict you can sign.

Six decisions, in order.

See a Readout Package opened live.