PMFKit

Trust pills

Verified, Likely, Unknown. Every claim in a PMFKit report carries one.

Most LLM output reads like a confident essay. PMFKit is not an essay. Every individual claim in every report has a label, and the label tells you how much weight to give it.

The three labels

  • Verified. Direct from a deterministic public source. The page itself, an API response, a structured Schema.org block. You can click through and confirm it yourself.
  • Likely. Inferred from corroborating sources where no single deterministic source confirms it. Above the AI fallback floor; not at the level of a direct quote. Examples: domain age via WHOIS, traffic order of magnitude via a public estimator, mention counts across social.
  • Unknown. AI inference where the deterministic sources failed or returned nothing. The report says so explicitly: "We couldn't verify this directly." Unknown is not the same as wrong; it means the model produced a best guess and we're telling you so.

Per claim, not per report

Confidence is computed per claim, not per report. A single Discoveries section can carry Verified pricing data, Likely traction estimates, and Unknown competitor positioning all side by side. The label appears inline next to the claim; the source breadcrumb is one click away.

How the label is computed

For each claim, we identify the most reliable source that supports it and inherit that source's tier. We downgrade by one tier if the model had to synthesize across multiple sources to produce the claim, and we never upgrade. A claim's tier ceiling is its highest tier source.

The verdict (Continue, Pause, Kill) carries an aggregate confidence label derived from the claims that feed it. A verdict supported entirely by Verified claims is high confidence. A verdict mixing Verified and Likely is moderate. A verdict resting on a majority of Unknown is low confidence, and PMFKit will explicitly recommend re running after you provide additional input or after time has elapsed and more public signal is available. See the verdict page.

Why every claim carries a label

Founders are technical readers. The senior founder reading a public report card on social will dismiss the entire output the moment one specific claim is wrong, unless they can see which claims rest on hard evidence and which rest on inference. The trust pills are how PMFKit stays useful in front of an adversarial reader.