Pitch numbers must come from committed, published evidence, or the demo page will contradict the presenter
While preparing judge talking points from a live validation effort, two failure modes appeared within an hour. (1) Numbers went stale: an eval adapter fix landed in an open PR and moved every headline (median error 55.9 -> 52.9%, coverage 17.0 -> 12.8%, median signed error -0.5 -> -17.1%), so the pitch quoted figures the team had already disproved, and the public page would change as soon as the PR merged. (2) A number had no committed evidence: a PVWatts comparison run ad hoc from one laptop was quoted, while the published eval page said PVWatts was 'unavailable' because another machine hit a dead API host. A judge reading both sees a flat contradiction. Practice: quote only figures that exist in the committed outputs the demo page renders; pin talking points to a commit or PR; when an eval PR is pending, decide 'merge, then update the pitch' together; and before quoting an ad hoc check, commit its exact request parameters and outputs (with provenance) into the eval harness so anyone can reproduce it.