Credibility System

Turn review questions into default artifacts.

Credibility is files, thresholds and isolation rules — not adjectives. This page expands the seven-step loop, COU essentials, default acceptance thresholds, audit fields and report requirements.

§ 1

Seven-step loop

Same workflow as the homepage, expanded for reviewers.

01 · COU

Context of Use

Scope, targets, out-of-scope, assumptions, risks and validation requirements in cou.yaml.

Structure examplecou.yaml

Structure example

cou.yaml

scope: research demo / teaching / early mechanism study
targets: [T_clear, AUC, TGI@dayX]
out_of_scope: [individual clinical prediction, dosing advice]
validation: { L1: required, L2: encouraged }

§ 2

COU required blocks

COU is the yardstick for validation strength. Coverage numbers are meaningless without it.

Use & readers
Prediction targets (endpoints)
Out-of-scope (incl. non-medical / non-individual)
Key assumptions & mechanism gaps
Validation requirements & adjustable thresholds

§ 3

Measurement model essentials

Align to experimental columns, not internal custom quantities.

Structure example

measurement_model.yaml

observable: inflammation_proxy
g: "integrate(field.IFN over ROI) / volume"
units: "arbitrary_units_declared"
lod: null
epsilon_model: "declared_noise_model"

§ 4

Default acceptance thresholds (adjustable in COU)

Default acceptance threshold

Metric	Default target	Note
UQ coverage (train/public)	≥ 70%	Configurable
Validation 90% interval coverage	≥ 60%	With sharpness
Sharpness relative width	≤ 2.0	Or WIS/CRPS
Verification-lite endpoint change	≤ 20%	dt/2 & doubled grid
Calibratable parameters	≤ 10 / scenario	Tier disclosure required

Design targets — not measured pass/fail on this site.

§ 5

Audit fields (summary)

Structure example

Field	Purpose
run_id / seed	Reproduction
param_set_id / hashes	Parameter & data integrity
model_version / scenario_version / commit	Code & pack alignment
train_task_id / validate_task_ids	Isolation proof
replay_ref	Report → replay jump

§ 6

Report must-haves

At least: COU summary, measurement-model versions, calibration setup, UQ (coverage+sharpness), L1/L2 definitions, V-lite summary, audit trail and replay_ref.

§ 7

Calibration / validation isolation

Validation conditions never enter the calibration objective; L2 conditions never tune parameters. Digitization error bands enter the audit trail — not post-hoc verbal notes.