empirical-collapse

Empirical Collapse

Collapse converges the superposition onto one survivor using evidence only. The rule is deliberately boring: verification gates, findings rank, conviction never enters. The engine (pqa.collapse.select_survivor) is canonical; the judge mirrors it and surfaces the reasoning — when they disagree, the engine wins and the disagreement goes in the report.

The collapse rule

1. Filter out branches with a critical unresolved adversary finding (deadly hits).
2. If any remaining branch has tests:
   • keep only those that passed verification;
   • none passed → no survivor: all branches failed; surface every death;
   • otherwise rank by (findings_resolved desc, coverage desc, non-incremental first).
3. If no branch had tests at all: collapse on adversary findings only and flag the result UNVERIFIED — recommend tests.

The final tie-break toward the less incremental branch is the quantum jump: when the evidence is equal, take the bigger defensible swing.

Protocol

1. Feed the judge structure only. pqa-collapse-judge (model fable) receives the N digests, the findings JSON, and the verifier results — never raw code. P-relativize is judgment over evidence, and the evidence here is structured; "let me read the code to get a feel" is the eloquence reflex in a lab coat.
2. Hold survivors as both-possibly-correct until the rule selects. Two verified branches that both resolved their findings are BOTH live candidates; apply the ranking tuple, not a preference.
3. Corroborate with the engine. Build a BranchResult(name, verified, has_tests, coverage, findings_resolved, findings_total, conviction, incremental) per branch and run select_survivor. The invariant gate (scripts/check_invariant.py) statically forbids conviction from entering _rank_key — that is the one rule that cannot break, enforced in CI.
4. Qualify confidence on every outcome: verified (94% coverage) / passes-but-thinly-tested (40% coverage) / verified (coverage unmeasured) / UNVERIFIED — no test suite. Coverage is a confidence qualifier, never a victory lap.
5. All-fail is a valid result. Report each branch's specific death; never promote a least-bad branch silently. The dead branches feed the failure taxonomy — that is this run paying for the next one.

Worked example

from pqa.collapse import BranchResult, select_survivor
results = [
    BranchResult("b0", verified=True,  has_tests=True, coverage=91.0,
                 findings_resolved=1, findings_total=2, conviction=None,
                 incremental=True),
    BranchResult("b1", verified=True,  has_tests=True, coverage=84.0,
                 findings_resolved=3, findings_total=3, conviction="high",
                 incremental=False),
    BranchResult("b2", verified=False, has_tests=True, coverage=None,
                 findings_resolved=2, findings_total=4, conviction=None,
                 incremental=False),
]
outcome = select_survivor(results)
# b2 is out (failed verification). b1 beats b0 on findings_resolved (3 > 1) —
# coverage is never consulted, and conviction was never consulted at all.
# outcome.survivor.name == "b1"; outcome.confidence == "verified (84% coverage)"

Had b0 also resolved 3 findings, the tie would fall to coverage and b0's 91% would beat b1's 84% — the non-incremental bit breaks only full ties, last. Walk the tuple in order; never improvise the ranking.

Anti-patterns

• Eloquence tipping the scale. The better-written rationale is not evidence.
• Conviction tipping the scale. Conviction shapes exploration; at collapse it is telemetry. A high-conviction loser still loses — and that row is the valuable part.
• Premature collapse. Picking "the obvious winner" before verifier results land is the collapse reflex the loop exists to suppress.
• Least-bad merging. An all-fail reported as a soft win poisons the registry and the operator's trust in the same move.
• Judge overriding engine. The judge explains; the engine decides. A silent override voids the auditability that justifies having a rule at all.
• Re-ranking after the fact. The survivor is the survivor; "but I prefer b0" belongs in a human review, not in the collapse record.