pub const DISCLOSURE: &str = "\
DISCLOSURE (pre-committed, System One program).
The calibration and agreement numbers in this report for {subject} are scoped as
follows, and no stronger claim is made:
1. MECHANICALLY-CHECKABLE FAMILIES. Absolute calibration anchors (ECE, Brier, SCE/ACE,
RPS, conformal coverage) are measured on procedurally generated verifiable-outcome
benchmark families whose ground truth is computed, not judged. They characterize
behavior on those families and their stated difficulty strata only.
2. TEACHER-AGREEMENT SPLITS. Where ground truth is not mechanically checkable, numbers
measure agreement with a corrected teacher ensemble. Calibration-to-teacher is NOT
calibration-to-reality: a model can match its teachers' probabilities exactly and
inherit their shared errors.
3. REALIZED TELEMETRY. Production-outcome calibration, where reported, covers realized
traffic since {telemetry_since} only, under the version triple (schema, model,
calibration) recorded with each batch. Past calibration does not guarantee future
calibration under distribution shift; post-hoc calibration is re-fit online and each
re-fit mints a new calibration version.
4. MARGINAL VS JOINT COVERAGE. Conformal coverage guarantees are marginal per question.
Cross-question joint coverage is measured and reported separately; it is lower than
the marginal guarantee whenever per-question errors correlate.
5. UNSEEN FAMILIES. Zero-shot transfer to never-synthesized question families is
measured on the held-out gate families and published with the same protocol. Transfer
is a measured result, not an asserted property.";Expand description
The pre-committed disclosure text, with {telemetry_since} and {subject}
placeholders for disclosure_with.
Claims are scoped to (a) mechanically-checkable verifiable-outcome families, (b) teacher-agreement splits, and (c) realized production telemetry from a stated date — and calibration-to-teacher is explicitly not calibration-to-reality.