research-document
Protocol
Protocol
Scope and control
The independent variables are (1) explicit uncertainty representation and (2) traceable evidence lineage. Framework K has neither requirement; Framework M adds uncertainty; Framework R adds lineage to uncertainty. All three retain identical operating guidance and identical required substantive sections. No case materials, domain guidance, or general reasoning instruction may differ by framework.
Blinding and isolation
The executor keeps identity-key.private.md sealed. Evaluators receive only framework identifiers, case files, raw outputs, and scoring materials. They must not inspect the key, prior scores, or other evaluations. Each run uses a fresh model context and one framework file plus one case file. Raw outputs are immutable; normalization may remove transport metadata but may not repair content.
Minimum matrix
| Case | K | M | R |
|---|---|---|---|
| 01 simple | 3 | 3 | 3 |
| 02 incomplete | 3 | 3 | 3 |
| 03 conflicting | 3 | 3 | 3 |
Minimum total: 27. Preferred: repeat the full matrix with a second provider family (54 runs). Record provider/model/version, temperature or deterministic setting, framework, case, trial, prompt hash, raw-output hash, timestamp, and word count for every run.
Standard diagnostic task
Every run receives exactly this task after its framework and case material: “Analyze this case using the provided diagnostic framework. Produce the framework's required response.” The case and framework files contain the substantive requirements; do not append variant-specific coaching.
Stop conditions
Classify the experiment as compromised and record the failure if identities are revealed before evaluation; prompts drift beyond the variables; a variant gets additional material; earlier outputs leak; an evaluator is told which framework represents FE; traceability to framework/case/provider/trial fails; or an output is manually repaired before evaluation. Do not silently restart.
Completion boundary
Completion requires the execution matrix, blinded evaluations, adversarial review, frozen analysis, unblinding record, final analysis, and a principle decision. This proposed package has none of those results yet.