spec: adversarial hardening of the Assay — measured trust, not asserted

[mdheller]

Self-review of the merged Assay (#198) — I red-teamed my own work and found four defects. This fixes all four and documents a fifth as intentional.

Findings & fixes

#	Severity	Defect	Fix
F1	🔴 Critical	The verifier's conclusion was never recorded — only its reliability. bad was reachable only via broken integrity, so a claim refuted by a calibrated verifier (clean integrity) could not project bad.	Add verifier.judgment (supported/refuted/abstained). refuted + calibrated → bad; refuted + uncalibrated → sad (an uncalibrated judge cannot force a verdict).
F2	🔴 Critical (ironic)	AssayStandard.calibrated was a self-asserted boolean — the exact assert-vs-measure anti-pattern the framework exists to kill, on the one flag that gates ok.	Add explicit calibrationThreshold; the validator recomputes F1 from confusionMatrix and rejects any record whose calibrated disagrees with (derivedF1 >= threshold).
F3	🟠 Major	Declared metrics (f1/precision/recall) were never checked against the confusion matrix.	Validator recomputes and rejects contradictions (tolerance 0.02).
F4	🟠 Major	agreement.effectiveVotes could exceed arms, defeating the decorrelation discount.	Validator rejects effectiveVotes > arms.
F5	🟡 By design	The projection doesn't consume the agreement axis for ok.	Documented as intentional (a deterministic computed+inline result needs no vote) in schema + README, rather than left silent.

Verification

tools/validate_reasoning_examples.py is green, and every new gate was confirmed to bite via negative tests:

• lying calibrated: true on an F1-0.26 standard → rejected
• bogus f1: 0.95 contradicting the matrix → rejected
• arms: 1, effectiveVotes: 99 → rejected
• refuted + calibrated → bad; refuted + uncalibrated → sad

Adds a genuinely-uncalibrated AssayStandard fixture (deployed-nli, F1 0.26, κ "slight") so the sad-path references real recorded-but-weak reliability instead of an absent lookup.

Theme

The original framework measured a claim's trustworthiness but let the judge assert its own. This closes that loop: trust is measured and cross-checked at every level, never asserted — the principle applied to the judges themselves.

Note: sourceos-spec auto-merges on green CI, so this is expected to land automatically; flagging for post-merge visibility.