We categorized every failed rubric criterion across the trajectory set into recurring failure modes, then counted how often each occurs. The frequencies below are the share of all failed criteria that fall into each mode (a criterion can touch more than one). Every mode is grounded in a complete example — prompt context, the exact rubric criterion, the model’s response, the judge’s verdict, and the judge’s justification — so the failure is legible rather than asserted.

What it means: The task asks for a report interpretation; the model answers like a treating clinician — naming a surgery, drug regimen, surveillance interval, or therapy the report neither contains nor requests.

Why this is a clean over-reach: The specimen is a tissue diagnosis: “active chronic gastritis, H. pylori present.” Nothing in the report calls for therapy, and the prompt asks for an interpretation. Prescribing a named, dosed, multi-drug antibiotic regimen is purely a treating-clinician act — the model has stepped from describing the slide to managing the patient. Unlike a borderline case where a recommendation might read as reasonable clinical context, there is no defensible reading in which naming “bismuth, tetracycline, and metronidazole for 14 days” belongs in a pathology read.

It is not a single-model quirk: The same guardrail caught the strongest model on the same task: Opus 4.8 (reward 90%) lost the criterion for recommending both bismuth quadruple therapy and clarithromycin triple therapy and naming tetracycline, metronidazole, clarithromycin, and amoxicillin. When even the top run on a task over-reaches into a drug regimen, the failure is structural — a default pull toward being helpful past the edge of the document — rather than a weakness specific to one model. The same pattern recurs across domains: naming Mohs surgery off a punch-biopsy basal cell carcinoma (report_2605653), prescribing steroid therapy for an organizing-pneumonia pattern that is not even a confirmed clinical diagnosis (report_2605435), and listing endocrine agents, letrozole, anastrozole, exemestane, off a breast core biopsy (report_2605225).

What it means: A limited specimen is used to infer something it cannot establish — an overall stage, a missing biomarker, a primary tumor type, or a final subtype. This is the mirror image of over-reach: instead of adding management, the model adds findings.

This is the hardest task in the benchmark (mean 51.9%) precisely because the right answer is mostly about what not to say. The strongest run on the same task (GPT-5.5, 81%) named the same limitations explicitly — “no primary specimen, so stage and receptor status cannot be assigned here” — and stopped.

What it means: The report deliberately leaves a question open — an unresolved differential, a discordance between modalities, a missing test — and the model resolves it too decisively. The clearest case is the marrow workup that anchors finding 3.

The winning Opus run (92%) reached the same leading diagnosis but framed the core–aspirate discordance as a substrate problem rather than a contradiction, and kept the aplastic-anemia differential and the missing PNH-clone test in view. Same destination, different epistemics — and the benchmark rewards the epistemics.

What it means: multi-report tasks are not just extraction — the answer has to say how morphology, flow, cytogenetics, and immunostains agree or remain in tension, and which modality should dominate. Weaker traces list each modality’s findings without naming the synthesis. Opus is strongest here; the AML-remission case (report_34_AML_remission) is the cleanest illustration: the rubric required stating that a 0% morphologic blast count and a 0.20% flow MRD result are not discordant — flow is simply more sensitive — and the weak Gemini run framed them as a “clear discrepancy,” inverting the intended reading.

Each deep-dive contrasts a low-scoring and a high-scoring trace on the same task. For each, we give enough context to read the task cold — the prompt, the rubric anchor in play, the verdict, and the judge’s justification — plus the high-level reason the scores diverged. These are the cases that make finding 3 concrete: the difference is rarely extraction and almost always interpretation.

What it means: A four-modality marrow workup where the controlling skill is preserving an unresolvable differential rather than committing to a diagnosis.

Why it diverged: Both runs used the same number of tools and reached the same leading diagnosis. The 31-point gap is entirely about whether the answer held the differential open. Trajectory length explained nothing; epistemic discipline explained everything.

What it means: A partial-nephrectomy renal-tumor report with an internal staging contradiction: the report assigns pT1b but documents tumor in perirenal fat (which defines pT3a).

Why it diverged: The weak run had the right instinct (something is off with the stage) but stopped at a vague flag. The benchmark rewards naming the discrepancy precisely. Here the shorter trace (2 tool calls) scored far higher than the longer one — restraint plus precision, not effort.

What it means: A benign thyroid FNA (Bethesda II). The skill under test is holding a benign call rather than over-working it.

Why it diverged: The weak run had the right instinct (something is off with the stage) but stopped at a vague flag. The benchmark rewards naming the discrepancy precisely. Here the shorter trace (2 tool calls) scored far higher than the longer one — restraint plus precision, not effort.

Realm Medical-Reasoning is hard because it is concrete: the rubric grades exact figures, preserved limits, and the absence of unsupported escalation, with no credit for sounding authoritative. Three takeaways follow, each tied back to the benchmark goal.