Risk register

Core risk is scientific semantic erosion: software can remain executable while input accounting, score meaning, provenance, or workflow claims become wrong.

Risk	Early signal	Consequence	Required control
silent row loss	accepted totals change without matching rejection records	biased downstream evidence	field accounting, rejected rows, reason codes
score-orientation drift	adapter or dependency changes “better” direction	thresholds and FDR invert	orientation fixture, boundaries, reference values
target/decoy policy drift	labels, strata, competition, or denominator moves	error control becomes incomparable	serialized policy and sensitivity tests
missingness collapse	absent, censored, filtered, failed, and zero merge	quantification and contrasts are biased	explicit states and adversarial sparse fixtures
provenance loss	producer, version, parameters, database, or transformation disappears	imported output appears native or irreproducible	lineage invariant and artifact review
circular benchmark	implementation generates its own expected answer	regression evidence is mistaken for independent validation	external reference or independently specified truth
transfer overclaim	one corpus or workflow family sets broader posture	public language exceeds evidence	companion, holdout, and family-specific ceiling
nondeterministic optimization	chunking or parallelism changes order, acceptance, or hash	repeated runs disagree	serial equivalence and stable artifact tests
policy leakage	Runtime, Knowledge, Intelligence, or Lab rules move into Core	scientific functions depend on context outside their contract	dependency and ownership guards
format overclaim	one producer fixture is described as full standard support	unseen dialects parse incorrectly	producer/version scope and mutation corpus

flowchart LR
    I["input accounting drift"] --> S["scientific semantic drift"]
    P["policy or provenance loss"] --> S
    S --> B["benchmark misinterpretation"]
    B --> C["overstated public claim"]

Silent acceptance drift, reversed score meaning, broken error control, lost provenance, and benchmark circularity are release-blocking. A large green suite does not reduce a risk whose relevant negative, reference, transfer, or consumer evidence is absent.