What problem does it solve? Contributors to Sonder Runtime need to know what counts as real proof of a change: which tests to run, how the hermetic pytest harness behaves, how to add tests the repo will accept, and which eval suites gate model promotion. This Skill encodes that verification doctrine so claims of "tested" or "verified" meet the repo's actual standards. ## Core Features & Use Cases - Evidence doctrine: Explains checkbox rules, the evidence ledger (requirements.jsonl with baseline/verified SHAs), and the house discipline of RED-before-fix, control cases, and mutation-proving every new guard. - Test harness contract: Documents the two conftest layers, offline sentinels, marker gating (--run-network, --run-model), the single live smoke test, and the known-red baseline at the verified commit. - Regression selection and ratchets: Covers select_regression_tests.py exit-code semantics and the three shrink-only ratchets (error signals, history privacy, architecture) that act as acceptance thresholds. - Execution-grounded evals: Inventories the verifier registry (VerifierUnavailable vs failed verdicts) and the promotion eval suites that gate model changes. - Use Case: A contributor fixes a bug and asks "did the tests really pass?" — the Skill directs them to reproduce the failure, show RED at the final item count, run the regression selector, and report exact commands with an explicit full-suite statement. ## Quick Start Ask the assistant what counts as verified evidence for your change and which tests you should run before opening a pull request.
