2Test results

Coverage claims are tested, not argued. The test is a paired-arm tabletop: for each scenario, one responder runs the incident twice with a fresh context each time — once without MC, and once with MC, the coverage cards the calculus derives from the ontology. Nothing else changes between the two arms. Both are scored per decision point — full, partial, or miss — against a key frozen before play, and each outcome is classified against predictions sealed in advance.

2.1The instrument

Six frozen scenarios probe three different failure surfaces. A content gap is something the ontology can express and the card simply supplies. A framework gap is something the ontology cannot faithfully hold — there, no coverage verdict is derivable in either arm, and the honest ceiling is disclosing the gap and escalating. A legibility gap would be a card that is present but unusable; the suite is built to catch all three.

Each decision point is scored by three independent passes, and the verdict is the majority of the three; agreement is reported per point. Every scored class is checked against a sealed prediction register written before the run.

2.2The full suite

The primary run is the complete six-scenario suite on a local open-weights responder (qwen3-30b, temperature 0, fixed seed), majority-of-three scored and self-calibrated. It is reported as a floor, not a point estimate: because the responder, the adjudicator, and the scorers are the same local model, a weaker baseline inflates the measured card effect while a stronger, independent judge would tighten it. The direction and the failure modes are what the suite establishes; the magnitudes are conservative.

2.3Two failure modes

The register records failures in both directions, and they are the point of the exercise. Read together, they say MC is neither uniformly good nor uniformly safe: it converts partial to full on content gaps, holds honest silence on most framework gaps, and can be led astray exactly where the world violates the model's value set.

One decision point, S3-DP4, drew both flags at once: on the off-environment offline crack — inexpressible in the model — the no-MC arm fabricated and the MC arm falsely assured. The two failure modes are not opposites of temperament but of which arm the gap catches; the same untreated framework gap can bite either way.

2.4Reference run — Sonnet 5

The fabrication finding first surfaced in the original run of this instrument on a stronger cloud responder (Sonnet 5), over the first three scenarios, scored by independent scorers. That run is the reference: the arm with MC scored full on all sixteen of its decision points, and the arm without it never beat it anywhere.

There, on the three domain-controller framework gaps, the no-MC arm produced confident, dangerous verdicts — "four-for-four, strong defense-in-depth" against a full domain compromise it could not actually see — while the MC arm disclosed the single largest scoping miss in the model and escalated. A strong responder concealed the content-gap deltas by reaching full from general competence on a few points, which is one reason the local full suite is read as a floor: a weaker baseline makes the card's contribution more visible, not less.

2.5What the suite establishes

Three results, each carried by a different part of the register. First, the cards are legible: no decision point anywhere failed with a usable card present, and the in-distribution control scenario was a clean sweep. Second, on content gaps the cards convert partial to full — the bulk of the MC arm's wins. Third, on framework gaps the calculus's contribution is not coverage but calibrated honesty: the MC arm says "unmodeled — escalate" where the no-MC arm confidently says "covered" — with S6 as the standing counter-example, where the same mechanism misleads when a signal falls outside the model's value set. The register is published with its failures because a coverage claim is only worth as much as the account of where it breaks.

2.6Reliability and records

Across the full suite the three scorers agreed on the majority verdict 94% of the time; the six split points are marked with a dagger in Register 2 and should be read with the most caution. The suite's scored gap-class matched its sealed prediction on 29 of 31 predicted points — a self-calibration of 0.94, with two points the run expected to be framework gaps and scored as content.¹

Every run retains its full instrument set: verbatim turn-by-turn transcripts stamped with replay hashes, machine-readable per-decision-point verdicts, the frozen scenario keys, and the prediction registers sealed before play. Results are auditable back to the transcript, and further runs are recorded here as their records land.

Notes
• 1 The mismatches were S1-DP5 and S5-DP3, both expected to be framework gaps and scored as content — the responder reached the point from general competence rather than needing the model to express it.
• 2 The full suite is self-scored: one local model served as responder, adjudicator, and scorer. This is why it is read as a floor. The S6 counter-finding survives transcript review — the MC arm's framing is misleading-leading rather than fully deceived — and that caveat is recorded with the result.