# EvidenceBound Squad Fatigue Safety Pilot v1

## Enterprise problem

Sports operators increasingly receive AI-generated analytical code that can alter risk scores, confidence margins, alerts, pricing inputs, or downstream review queues. Conventional unit tests rarely prove that an exact candidate preserves operator-owned formulas, fixed-point semantics, threshold behavior, bounded execution, and artifact integrity.

## Pilot outcome

EvidenceBound turns one Squad Fatigue candidate into a reproducible safety decision:

```text
candidate source
  -> operator-owned contract
  -> canonical AST and execution IR
  -> isolated deterministic execution
  -> formal invariants and counterexamples
  -> VERIFIED / REFUTED / BLOCKED
  -> content-addressed Proof Pack
  -> mandatory human promotion decision
```

The public case uses controlled operator fixtures. It does not claim provider-observed player fatigue, medical status, match-outcome accuracy, profitability, or automatic deployment safety.

## Contract coverage

The contract binds:

- fixed-point scale 4 and raw tick scale 10,000;
- exact Squad Fatigue formula semantics;
- exact operator confidence-margin clamp semantics;
- two-way threshold and kill-switch equivalence;
- output bounds;
- deterministic fuel and loop limits;
- network-disabled isolated execution;
- content-addressed artifact integrity;
- human approval and no automatic promotion.

## Adversarial evidence

The corpus contains 15 adversarial mutations:

- threshold and kill-switch bypasses;
- hardcoded and snapshot-overfit outputs;
- omitted and drifted feature coefficients;
- weakened output ceilings;
- mixed fixed-point scale injection;
- deterministic fuel exhaustion;
- post-generation Proof Pack tampering.

Expected evidence distribution: one canonical VERIFIED case, twelve REFUTED semantic mutations, and three BLOCKED execution/integrity cases.

## Private deployment boundary

The private enterprise package runs behind an AF_UNIX socket in customer-controlled infrastructure. Candidate source, fixtures, receipts, and Proof Packs remain inside the tenant evidence root. The worker does not require a production signing key for evaluation; optional Ed25519 sealing remains owner-controlled.

## Pilot acceptance

A 14-day pilot is accepted only when:

1. the canonical candidate produces a reproducible VERIFIED Proof Pack;
2. every mutation reaches its declared REFUTED or BLOCKED result;
3. no candidate can access the network or escape the evidence root;
4. worker timeout and kill-switch behavior are observed;
5. receipt hashes reproduce on a clean machine;
6. human approval remains required for all promotion decisions.

## Commercial hypothesis

The paid value is reduced model-risk review time, stronger audit evidence, private execution, deterministic replay, and fail-closed promotion control. Willingness to pay remains unverified until a qualified operator runs the pilot, integrates at least one real candidate, and requests continued or expanded use.