Evidence quality
What they said is not what they did
Foundry grades startup evidence quality by customer behaviour, not by how encouraging it sounds. What a customer said ranks below what they did, which ranks below what they committed, which ranks below what they paid. It also weighs whether each source is independent or founder-controlled, whether the evidence is current, and whether you tried to prove yourself wrong.
The evidence ladder
PAIDWhat they paid
Money changed hands, or a contract commits them to pay.
- Payment · can reach Strong
COMMITTEDWhat they committed
They put real time, people or budget behind it.
- Resource commitment · can reach Strong
DIDWhat they did
They acted: built a workaround, shared real data or access, changed how they work.
- Existing workaround · caps at Medium
- Shared data or access · can reach Strong
- Workflow change · can reach Strong
SAIDWhat they said
Opinions, stated intent, and a problem described in their own words.
- Opinion · caps at Weak
- Stated intent · caps at Weak
- Confirmed problem · caps at Medium
The ladder is an explanation, not the whole policy. Two things on the same rung can carry different ceilings: a workaround and a changed workflow are both things a customer did, but only the second can count as strong. The outcome always comes from the full policy.
Who the evidence comes from
- Independent
- Not connected to you or to your other sources. Only independent evidence counts towards the requirement for more than one channel.
- Related
- Connected to you or to another source: a friend, a former colleague, someone a previous interviewee introduced.
- Founder-controlled
- A source you control, such as your own team or a company you run. It is recorded, and it never satisfies an independence requirement.
High-risk decisions, and every Kill, need independent evidence from at least two different channels. Five interviews from one channel are one channel.
Whether it is still true
You mark each item current or stale. Foundry applies no decay score and no half-life. When two evidence bases are otherwise equal, current evidence outranks stale evidence.
Disconfirmation, the falsifier, and repair
- Disconfirmation
- Did you actively look for the customer or situation most likely to prove you wrong? Without that, the policy will not return Build.
- The falsifier
- What would prove this assumption wrong, and when you wrote it down. A falsifier defined after you started collecting evidence cannot be used to stop an assumption, because it could have been fitted to the result.
- Repair attempts
- Before an assumption is stopped, you must have changed something real: the pricing, the target customer, the message, or the channel. Saying you tried is not counted; the changed condition is.
Decision risk
You choose how costly or hard to reverse it would be to be wrong. The bar rises with the risk:
| Risk | What Build requires |
|---|---|
| Low | At least one piece of evidence, and a disconfirmation attempt. |
| Medium | Evidence that reaches Medium strength, at least one item of observed behaviour, and a disconfirmation attempt. |
| High | Strong observed behaviour, independent sources in at least two channels, a disconfirmation attempt, and your explicit confirmation. |