Verify the Algorithm, Not the Outcome
A high-stakes analytical system can produce a bad real-world outcome even when its code followed the approved contract — and a good outcome even when code drifted outside it. EvidenceBound separates those questions so operators can prove what an algorithm did without pretending to prove the future.
Outcome quality and algorithm conformance are different claims
When an analytical recommendation becomes consequential, the first reproducible technical question is narrow: did the exact candidate execute the approved rules on the approved evidence? That is different from asking whether the real-world outcome was correct.
Why the sports case is useful
Sports analytics provides constrained examples with fixed-point arithmetic, thresholds, bounded outputs, event replay, resource limits and artifact tampering. Small semantic changes can cross a decision boundary while remaining visually plausible in review.
- Threshold boundary drift
- Fixed-point scale mismatch
- Omitted clamp or output bound
- Reversed comparison
- Resource exhaustion
- Changed evidence or result identity
The operator owns the contract
The verifier must not infer success from the candidate itself. The operator declares expected formula, scale, thresholds, bounds, resource policy, evidence identity and approval behavior first; the candidate is then checked against that declaration.
Mutation evidence is stronger than a polished demo
A happy path shows that one candidate can pass. An adversarial mutation corpus asks whether meaningful drift is caught and whether negative states retain reproducible evidence.
Human promotion remains first-class
Verification and promotion remain separate. A VERIFIED result can permit review, but it does not deploy the candidate. Interruption and recovery should preserve lineage so accountability remains visible.
From sports to a general control pattern
The generalizable element is the control pattern, not a claim that one sports implementation validates every domain. Finance, healthcare, infrastructure or other high-stakes applications require their own operator-owned contracts, domain validation and acceptance evidence.