How it works
The same policy, every time
Foundry performs a deterministic startup evidence evaluation. You enter one assumption and classify each piece of evidence in a structured form. A fixed, versioned policy then checks that structured input against explicit rules and returns Build, Iterate, Kill or Abstain. The same input always produces the same result, and no AI model decides the outcome.
A review, in four steps
Pay once
€199 through Stripe Checkout. Foundry confirms the payment with Stripe directly before the intake opens.
Describe the assumption and the evidence
State the assumption, choose how costly it would be to be wrong, and add each piece of evidence with its behaviour type, source relationship, channel, recency and date. Record what would falsify the assumption and when you decided that.
The policy evaluates it
The policy runs on Foundry’s server, on the structured input only. Free text you write is kept for your record and is never read by it.
Read the result, and come back to it
The result is stored with the exact input and policy version that produced it. You can reopen it from your reviews page.
What the policy checks
- Is the question answerable? If the assumption is not stated, or the question already assumes its answer, the result is Abstain.
- How strong can each item be? Each behaviour type has a ceiling. An opinion can never count as strong evidence, however many you have.
- Does the evidence meet the bar for the risk? Low risk needs some evidence. Medium risk needs observed behaviour. High risk needs strong behaviour from independent sources in at least two channels, and your explicit confirmation.
- Did you try to prove yourself wrong? Build requires a disconfirmation attempt.
- Is stopping justified? Kill requires a falsifier written before the evidence, a real repair attempt, contradiction repeated at least seven days apart, independent evidence, and your confirmation. Missing any one, the policy will not return Kill.
- Is there a better-supported alternative? If your assumption has weakened and an alternative has comparable or stronger evidence, the result is Iterate, pointing at the alternative.
Evidence bases are compared by their best item, never by adding items up. More of the same weak evidence does not become stronger evidence.
Where AI is, and is not
The outcome is produced by ordinary, deterministic code. No language model reads your evidence, classifies it, or chooses a result, and your evidence is not sent to one. If Foundry later uses a model to help explain or format a result, it will not be able to change the outcome the policy produced.
Every result can be replayed
Foundry stores the structured input next to the result and the policy version, currently [email protected]. Running that version on that input again gives the same result. A result is not edited after the fact.