Evidence-grade Testing

When adding or updating test/verification logic, treat the harness setup as part of the test and ensure the evidence is causally attributable and non-vacuous.

Standard

1. Isolate the control variable (no hidden confounds): If your A/B or control relies on shared artifacts like node_modules, document and enforce when that’s a clean control. If dependency inputs change (e.g., package.json/lockfiles), either run a dependency-aware control or explicitly record the confound.
2. Prove the wiring really points to the intended code: In monorepos, internal workspace dependencies may symlink into the wrong (e.g., PR/head) tree. Add a check that resolves the actual runtime targets (e.g., realpath of resolved modules) and gate pass/fail of the verification on that.
3. Require non-vacuity for “prove the test matters” runs: If you validate a new/changed test by reverting/mutating inputs, the run must (a) reach the setup, and (b) fail the intended assertion with an expected-versus-actual mismatch message. If the revert breaks import/compile/setup before the assertion, record it as inconclusive, not proof.
4. Regression-test the harness invariants (the harness is production code): Add unit tests that pin subtle logic in your automation (e.g., stage-marker matching uses startswith, and you filter by bot author). Also ensure ids are re-resolved at patch time when threads/comments may have mixed order.

Example pattern (monorepo control isolation)

# Run from the base worktree.
# 1) Resolve what the code under test actually loads.
REALPATH=$(node -p "require('fs').realpathSync(require.resolve('@qwen-code/qwen-code-core'))")

# 2) Assert it points into the base tree (pseudo-check).
case "$REALPATH" in
  *tmp/base-tree/*) echo "OK wiring: $REALPATH" ;;
  *) echo "FAIL: base control resolved to non-base code: $REALPATH" ; exit 1 ;;
esac

Example pattern (non-vacuity rule)

• If reverting a key source hunk for a new test does not reach the intended assertion and produce the expected mismatch failure message, mark the vacuity check inconclusive and use an interface-preserving mutation instead.

Example pattern (harness regression tests)

• Add structural/unit tests that verify your comment/stage matching uses startswith (not contains) and filters by bot login, so a future edit can’t silently reintroduce clobbering behavior.