nw-adversarial-refutation

name	nw-adversarial-refutation
description	The adversarial-refutation review stance — assume the artifact is WRONG and try to PROVE it, default-to-refuted, diverse lenses, and an exhibited executable counterexample. The shared SSOT every DELIVER review (per-slice C_REVIEWER_AUDIT + per-feature F_FINAL_REVIEW) applies so the expensive final swarm is needed less.
user-invocable	false
disable-model-invocation	true

Adversarial Refutation Stance

The review METHOD (the HOW), not a checklist. The dimensions (nw-sc-review-dimensions, nw-at-completeness-check) are the WHAT to look for; this is the falsification POSTURE the adversarial swarm runs — the one that caught the C13/C14 catalogued-not-wired false-DONE, and on the consolidated 3-lens review the named-catch-all guard + fragile state-machine that two homogeneous reviewers would have missed. Apply on every per-slice (C_REVIEWER_AUDIT) + per-feature (F_FINAL_REVIEW) DELIVER review.

Pillar 1 — confirmation → falsification

A confirmation review asks "does this look right?" and accepts the artifact's own claim. The refutation stance INVERTS the burden:

Assume the artifact is WRONG. Try to PROVE it broken. You cannot prove an implementation correct — only fail to refute it. The burden is on the ARTIFACT to SURVIVE refutation, never on you to prove a bug exists.

Popperian asymmetry — the single biggest shift from a confirmation review. Incomplete without lens-diversity (Pillar 2).

Pillar 1b — the SCORE is INDEPENDENT of design-agreement (the coherence-maximizer counter-force)

The subtlest way a review passes a wrong artifact: it scores how well the artifact AGREES with the stated design/intent, not whether the artifact SURVIVES an independent attempt to break it. An artifact that matches the design is coherent — it is not thereby correct. A reviewer rewarded for confirming design-coherence maximizes COHERENCE, not TRUTH: it climbs the gradient toward "reads as intended" and awards its highest score exactly where the artifact and the design tell the same story — which is precisely where a shared blind spot hides. This gradient is the force the whole stance exists to oppose; default-to-refuted (Pillar 2) and the exhibited counterexample (Pillar 3) only bite if the SCORE itself is blind to design-agreement.

• Score only survival, never agreement. The verdict is a function of "did the constructed counterexample get caught?" — NOT "does this match what the design said?". Set the design aside as a claim to be broken, not an oracle to be confirmed against. Two artifacts that agree can both be wrong in the same way; agreement is evidence of nothing.
• An empirical design-question is resolved by a TARGETED probe, not by assent. When the design carries a question to settle at review/RED ("Q-1: verify X at RED", "confirm the wiring reaches production"), it is discharged by a probe that DISCRIMINATES the specific cause — not a coarse end-to-end red compatible with every cause, and never by the artifact asserting the question is answered. A walking-skeleton test that calls the target directly (callers in tests, ZERO in production) resolves nothing: it is never_wired wearing the name of the discipline that exists to prevent it. Get the production-caller count via des code-fact query.callers-of SYMBOL --root ROOT (the provider-neutral nw-code-analysis-port, degrade-LOUD if unavailable) — do not eyeball it.

Pillar 2 — two levers, BOTH required

1. Default-to-refuted — uncertain = REFUTED, not "probably fine". Forces the artifact to EXHIBIT why it survives instead of riding the reviewer's benefit of the doubt. ALONE it degenerates into skepticism-noise — it becomes coverage only when paired with lens-diversity.

2. Perspective-diversity — each reviewer/pass attacks through ONE DISTINCT LENS (correctness · does-it-reproduce · security · vacuity · wiring · oracle-soundness). The power is the UNION of lenses, not redundant skepticism; N identical refuters catch less than N different ones. (Empirical: an architecture lens caught a named-catch-all a test-fragility lens missed, and vice-versa — homogeneous refuters miss one.)

   Where this pillar fully lands (do not over-claim). One reviewer can realize assume-wrong, default-to-refuted and an exhibited counterexample, but only approximates perspective diversity by changing named lenses. Genuine diversity requires distinct observers. Use it proportionately at the whole-delivery join; do not multiply per-slice reviews or call one reviewer's passes an independent swarm.

Pillar 3 — evidence discipline (for REFUTE and for CLEAR)

The executable counterexample is the witness. No prose-only verdicts.

• To REFUTE — EXHIBIT the failing case: the breaking input, the deleted call-site that leaves the test green, a real build error (trybuild/compile), an actual failing run. A refutation without a constructed, runnable counterexample is a suspicion, not a finding.
• To CLEAR — the artifact must EXHIBIT why it survives: run the counterexample, show it is caught. A clear without exhibited survival is a confirmation-bias pass.
• The witness runs against the REAL artifact, not a synthetic snippet — a doc-test or hand-written sample can diverge from the real impl (the H-1 fix required a real trybuild E0004, not a prose note: the synthetic test missed the named-catch-all in the real impls).

Failure-mode taxonomy (the swarm's hunting checklist)

Drive each refutation pass against these refutation-specific modes — the ways an artifact looks green while proving nothing that go BEYOND the test-quality catalog. For the vacuous / tautological / mock-dominated / zero-assertion family, apply the Testing-Theater Detection catalog in nw-sc-review-dimensions under this stance — it is the SSOT for those and is not re-listed here.

Failure mode	The attack / litmus
Catalogued-not-wired (asserts the declaration, not the behavior)	Delete the call-site that wires the new code — does the test stay green? If yes → wrong-level / never reaches production dispatch. (Testing-Theater's "delete the production code" litmus misses this: a test calling the leaf directly stays meaningful while production dispatch never wires it.)
Oracle silent today (the "true-positive oracle" does not actually fire)	Run the positive case — does the detector/AT actually FIRE? A confirmation review accepts "the AT asserts it fires" without verifying it fires.
Guard that does not guard (the enforcement mechanism is itself bypassable)	Attack the guard: a fragile wildcard, a named catch-all other =>, a too-narrow regex — the guard-test passes but can false-green.
Byte-identity claimed-not-proven ("behavior-preserving" without an exact-set oracle)	Demand the exact-set oracle (the full tuple), not a count-only check that misses a reordering/swap.
Scope-creep / silent-narrowing (the delivery ships more or less than declared)	Diff the delivered surface against the validated contract targets and obligations — anything extra or missing is unattested.
Cross-language projection leakage (POLYGLOT / language-agnostic projects only)	A language-specific sink/predicate fires on another language's file via callee name-collision. Run the projection against a SIBLING-language fixture carrying a colliding symbol (e.g. Python exec("c"+x) vs a TS/JS exec detector) — does it stay silent? If it fires on the wrong language, the projection leaks. Invisible to a single-language hermetic AT — the genericity-dual of catalogued-not-wired.

The first five modes are UNIVERSAL (any review, any stack). Cross-language projection leakage is CONDITIONAL — it applies only when the project is polyglot / language-agnostic (e.g. nWave's own multi-language detectors); skip it for single-language work.

Verdict shape

Each pass ends as either: a REFUTED finding carrying its exhibited executable counterexample, or a SURVIVED verdict carrying the demonstration that the counterexample is caught. "Looks fine" is not a verdict.

---

Empirical source: an independent nWave adversarial verifier swarm — the 15-verifier swarm that caught the C13/C14 catalogued-not-wired false-DONE, and the consolidated multi-lens DELIVER review. This SSOT productizes that stance into the standard per-slice + per-feature reviews so the full swarm is the exception, not the rule. The cross-language projection-leakage mode was found by that swarm APPLYING this taxonomy's catalogued-not-wired litmus on a real polyglot slice (F1 slice-03, TS/JS detector D2) — the method generating a new mode is the stance working on itself.