Sample report

Illustrative Evidence Report

A fictional record outline for an evaluation question, controls, possible proof designs, and the information needed for a next decision. It contains no model measurements or observed result.

Status

Illustrative only — not evaluated

ILLUSTRATIVE RECORDILLUSTRATIVE ONLY

Preserve the path from question to evidence.

This example shows the fields that keep an observation, interpretation, intervention decision, proof mode, result, limitations, and next decision connected. It does not report a completed test.

Record identity and status

Illustrative structure only. No real model, customer, checkpoint, evaluation run, or completed result is represented.

Evidence Profile

Declare the task, intended use, evidence available at decision time, source and version details, support boundaries, missingness, subgroup and temporal coverage, confounds, and unresolved questions. These fields are not populated with customer evidence here.

Evaluation question

For an illustrative longitudinal documentation task, does the output preserve the specified behavior when the same available evidence is presented with only whitespace or section-order changes?

Exact-formatting control

Compare renderings of the same synthetic input while preserving wording, values, units, timestamps, uncertainty, and information availability. Change only whitespace or section presentation. Prespecify the behavior expected to remain stable. No response or score is supplied.

Changed-information counterexample

A material observation documented after prediction time is unavailable to the model at that earlier time. This changes the observable evidence, not just formatting; a changed answer or greater uncertainty may be appropriate.

state the prediction-time cutoff
separate clinical state from observable evidence
record how evidence is realized for the model
assess expected detectability separately from the model response

Failure adjudication

Not evaluated. No model response, failure, clinical significance, or detectability result is claimed in this example.

Intervention decision

None selected. Input or adapter repair, preprocessing, calibration, thresholding, prompt or retrieval changes, routing, abstention, targeted data, or weight adaptation may be considered only if permitted and relevant to an adjudicated gap.

Proof mode

Not selected. Sealed candidate-blind synthetic proof, patient- or episode-grouped customer cross-fitting, and stronger untouched or source-distinct real-world evidence support different claims.

Proposed controls and regression

If reviewed and approved, preserve the exact-formatting pair as a control and keep any changed-information case separate. A fixed regression case requires a declared expected behavior and scoring rule.

record the scenario and realization versions
preserve the information cutoff and source availability
record adjudication and uncertainty
retain unsupported claims and the next decision

Candidate and intervention

Not applicable in this illustration. No intervention, candidate model, or updated checkpoint exists here.

Evaluation protocol

Freeze the claim, materiality criteria, comparator, proof mode, no-regression checks, and evaluator authority before evaluation. Cross-fitting must exclude each evaluated patient or episode from the whole fold-specific adaptive pipeline.

record model, evidence, rubric, and protocol versions
do not tune against an evaluation fold
do not treat a full-data candidate as held out from its own data
state which stronger claims remain unsupported

Result and regression status

Not evaluated. No response, comparison, intervention effect, regression run, or release decision is recorded.

Next decision

Confirm the task contract, information available at prediction time, evaluation authority, and permitted changes. Then decide whether a controlled assessment is justified and which proof mode matches the intended claim.

Unsupported claims

This illustration establishes no measured failure, clinical consequence, intervention effect, clinical validation, safety result, deployment authorization, or real-world improvement. A graph relationship may be source-grounded context; it is not automatically a causal mechanism.

Result boundary

Every outcome is marked not evaluated. No model-improvement result or measured weakness is claimed.

Explore the evidence and proof boundaries.

Start with high-level, non-PHI context. Do not send patient records, model weights, credentials, source code, or held-out evaluation rows through public email.