A gate set's split needs no case that expects a blocking Finding (#691)

Merged
DO
8d ago
+60-11

Closes #691, "A gate set's split needs no case that expects a blocking Finding", one of the gate-v1 preconditions.

The change

split_ok in harness/eval/__init__.py held every set, gate sets included, to BLOCKING_LEAST cases that expect a blocking Finding. A gate set that grades only what no deterministic check holds may hold none. Under the old rule such a set was eval_split_short on every run, so its gate could never pass.

  • A gate set: split_ok asks only kind_least_of cases of each kind.
  • blocking_pass: still computed and reported, pooled as before, and it closes nothing (decision 0150).
  • A development set: unchanged.
  • No other consumer assumes a gate set holds a blocking case: aggregate_verdict, pooled_verdict, pooled, closed_of, the ledger and the preflight read the word only as null or false.
  • What a gate report reveals: a little less than before. split_ok: true no longer implies that a blocking case exists.

Tests and mutants

  • test_a_gate_set_needs_no_case_that_expects_a_blocking_finding runs a gate set with no blocking case under BLOCKING_LEAST 1. It holds that:
    • the set is split, and aggregate_verdict and pooled_verdict pass its sealed report;
    • its own least of a kind still refuses it when raised;
    • the same cases as a development set are short, and are split only with BLOCKING_LEAST at 0.
      It fails on the old code.
  • New mutant "a gate set is held to a blocking case again", killed. Three mutants refreshed to the new text, each killed.
  • harness/eval/AGENTS.md gains the rule's row.

Record

The record amends decision 0150's blocking word for gate sets alone. It is unnumbered (NNNN) until the Factory Builder gives the number at enqueue, so the records check is red until then. make checks passed with it numbered on a temporary commit.

Checks

CheckResult
make checksOK, with the record numbered on a temporary commit, since reset
tests.test_evalsOK, 80 tests
Mutants, filtered4 of 4 killed (1 new, 3 refreshed)
gitleaksNo leaks
Full make verifyNot run locally, by the Maintainer's standing change of 2026-11-09; CI is the answer
Argon (fcd4ca5)Clean
Muse (fcd4ca5)Blocking: the record is unnumbered, by design until enqueue. Minor, fixed: the BLOCKING_LEAST comment and blocking_pass's docstring stated the old rule for every set
Claude review and security (fcd4ca5)Nothing blocks. Lows, fixed: the same two comments and a test's comment; the record's Status no longer hints at a coming gate set's make-up. Nits: harness/eval/AGENTS.md row, added; a test at a non-zero least of a kind, not added, since decision 0169's test holds the least itself. Security: nothing
AstraOnly where Muse gives no read

Built against the schema and synthetic fixtures only; no gate-set case, count or composition was read. Labelled factory:hands-off.

šŸ¤– Generated with Claude Code