Why Science to Startup

Every score, prediction, and enrichment on Science to Startup is backed by an Ed25519 signed receipt. The receipt locks the claim hash, the model version, and the computation timestamp into a tamper-evident envelope that anyone can verify against the public key.

Receipts are published in real time on the Proof Feed and can be verified against our JWKS endpoint or the founder PEM key.

No other research intelligence platform signs its outputs. This is not a feature -- it is the foundation the rest of the proof stack builds on.

Signal Fusion scores are not black-box model outputs. Each component signal -- citation velocity, author lineage, GitHub traction, community buzz -- links directly to its source data: arXiv metadata, OpenAlex citation graphs, GitHub repository snapshots, and social-mention threads.

The Signal Canvas surfaces every citation as an inspectable card with the original text, page number, and extraction receipt. Reviewers can audit any claim without trusting the model.

Source provenance is not optional -- it is structurally enforced. The enrichment pipeline rejects any score that cannot cite its upstream data, and the methodology page publishes the Brier calibration across all graded sub-areas so accuracy is verifiable, not asserted.

Every prediction is evaluated against a holdout ground-truth set using the Brier score -- the same metric used by forecasting tournaments like Metaculus and Good Judgment Open. Lower is better; our public gate is 0.30.

The Methodology page publishes the current Brier score, the 95% Wilson confidence interval, a calibration plot, a reliability diagram, and per-sub-area small multiples -- all SSR, no client JavaScript, refreshed by a daily recompute cron.

Ground truth is Opus-4.7-graded with founder spot-check, not founder-graded. This matters: the grading model is disclosed publicly, and anyone can pull our open-source Brier helpers from apps/web/lib/methodology/brier.ts and re-run the math.