[JacobStephens2]

Summary

paired Trial Scores + fixed Bar
  → paired Dirichlet intervals for accuracy and wrong findings
  → pass / fail / inconclusive, with the paired dropout guard
  → posterior-predictive task estimates when inconclusive
  → JSON with the Bar, sampler provenance, counts and pass³

Adds the standalone Verdict module. The default Bar uses a signed ten-point tolerance (-0.10), a 95% credible level, one uniform pseudo Eval task, and 20,000 seeded draws.

Closes #999

Test seams

• The Verdict module on its own: verdict::judge accepts paired Eval task Scores and a Bar, and returns the serializable Verdict. This is parent Spec: Experiments, which measure a change before it reaches production #977's pre-agreed seam 2.

Evidence

• Before: the focused regressions failed for the unimplemented interval, missing decision rules, retained void tasks, unreversed wrong findings, absent pass³, missing forecasts and missing JSON provenance.
  After: cargo test --bin thirdshift verdict::tests passes all 10 focused tests, including a sure loss inside the margin and the first forecast count reaching 0.8 assurance. Independent Python simulations corroborated the one-Trial interval and the small-gap forecast range.
• cargo fmt --all --check and cargo clippy --all-targets -- -D warnings pass.
• /usr/bin/time -p scripts/affected-tests issue-977 selected cargo nextest run: 2,386 passed, 56 skipped. Wall time 2,513.50s, including compilation (3m 20s); nextest execution 2,309.858s on Linux.

Merge Danger

Door: two-way

Blast Radius: Experiments

The statistical module has no session or storage dependencies. Forecasts assume new non-void Eval tasks and record their finite simulation budget; exhaustion returns an explicit unfinished estimate.

Review coverage

Reviewed against issue-977 (d05aa971). Standards: one naming finding, addressed by using “Eval task” in the three test names. Spec: no confirmed findings. Both reviewers inspected all seven changed files, including Cargo.lock.

Unread changed files: none on either axis. Generated-output exclusions: none. Final reports are saved locally as .thirdshift-review-bzWxTy/standards.md and .thirdshift-review-bzWxTy/spec.md.

Unaddressed findings

Standards

None.

Spec

None.

Built with codex · gpt-6.1-sol · xhigh