Model-Divergence Taxonomy: survey behind ModelDivergenceTypeEnum (2026-09-02)

Purpose. Fix the value set of a typed model-divergence taxonomy from evidence rather than invention, in the way projects/PATHOPHYSIOLOGY_SCALE_FEASIBILITY.md fixed BiologicalScaleEnum at four values.

Method. Every modeled_mechanisms link in kb/disorders/ + kb/modules/ was extracted with its limitations prose, relationship and fidelity. All 50 computational-model limitations were read in full and clustered by hand; the animal and NAM sets were probed by keyword only, to establish which kinds are shared and which are modality-specific. Keyword counts below are a probe, not a classification — they over- and under-count, and are quoted only to show which kinds recur.

---

1. Coverage

Measured 2026-09-02, against the KB as it stood when the 50 computational-model strings below were read and clustered. These are a snapshot of the derivation input, not a live count — the KB grows, and the ratio has since moved (1,662 links / 80% on main at 3b1ba68c3, 2026-09-03). The clustering in section 3 is what the value set rests on; re-measuring the totals does not change it.

	links	with limitations
animal_models	603	577
experimental_models	446	204
computational_models	82	50
total	1131	831 (73%)

So the caveat is nearly always written — it is simply written as unstructured prose, in one slot, with no type, no per-item granularity, and no way to ask "which models are limited by calibration provenance rather than by species".

2. Modality profiles differ sharply

Keyword probe, per section:

candidate kind	comp	exp	animal
species divergence	5	26	276
genotype / allele mismatch	6	33	123
incomplete phenotype	2	10	53
missing compartment / cell type	1	21	29
scale extrapolation	3	27	20
context mismatch	1	27	13
supraphysiological expression	0	14	20
static / no kinetics	6	2	17
proxy / surrogate quantity	5	11	9
parameterisation / calibration	11	1	5

The animal set is dominated by species and allele divergence — the classic translational caveats. The computational set is dominated by calibration provenance, absent dynamics, and proxy quantities, and species divergence is nearly absent. A single undifferentiated fidelity: LOW tier is doing very different work in the two cases, which is the core argument for typing the divergence.

3. The computational taxonomy, from 50 read limitations

Counts are hand-clustered; a limitation usually carries more than one kind, which is why the divergence slot is multivalued.

Kind	n	Representative evidence
BOUNDARY_OMISSION	~20	Asthma: "airflow, heterogeneous airway branching, neural control, mucus obstruction … are outside the single-airway-wall formulation". RA-FLS: "does not represent the immune-cell compartment".
CALIBRATION_PROVENANCE	~14	ASPS: "parameterized from mouse fibroblast experiments rather than human pacemaker neurons". PD: "parameterised from cohort-level miRNA and transcriptomic data rather than from direct measurement of autophagic flux".
PROXY_QUANTITY	~11	MS: "FLAIR lesion fraction is an imaging proxy, not a direct measure of myelin thickness". Glioblastoma: "ADC is an imperfect proxy for tumor cellularity".
CAUSE_UNREPRESENTED	~10	Brugada: "sodium-channel reduction is prescribed phenomenologically rather than generated by an SCN5A allele". ASPS: "does not encode a heterozygous human allele or allele dosage".
STRUCTURAL_IDEALIZATION	~8	Brugada: "a homogeneous one-dimensional generic epicardial ventricular strand, not an anatomically reconstructed RVOT". Asthma: "idealized axisymmetric elastic cylinder".
TEMPORAL_SCOPE	~7	DMD: "only represents a small muscle cross-section over 28 days"; "years-long human disease evolution are not represented". COVID-19: "qualitative reachability … rather than its magnitude or timing".
CONTESTED_ASSUMPTION	~4	ASPS: 2026 knockout evidence "identified MARK2 as physiologically significant", displacing the model's CK1delta/epsilon assignment. Sickle cell: "HbS nucleus structure is not experimentally known".
SCALE_EXTRAPOLATION	~4	RA-FLS: "bone and cartilage loss are represented as output nodes of the synoviocyte network rather than as tissue-level processes". Already derivable from model_scale.
SPECIES_MISMATCH	~4	DMD: "mdx disease is substantially milder than the human disorder". The bridge to the animal set, where it is the dominant kind.
POPULATION_MISMATCH	~4	Glioblastoma: "19 IDH-wildtype and two IDH-mutant tumors, so it is not a pure match to this disease entry". ASPS: "validation cohorts comprised 56 healthy adults … not people with ASPS".

Boundary cases considered and folded in. Confounded perturbation (Cystic fibrosis: "CF required simultaneous parameter changes in chloride, sodium, and potassium transport; CFTR loss is therefore not an isolated disease arm") has one clear instance and is folded into CAUSE_UNREPRESENTED. Unvalidated (8 keyword hits) is not a separate kind: every instance also names what the validation failed to cover, which is one of the kinds above — usually CALIBRATION_PROVENANCE or POPULATION_MISMATCH.

PROXY_QUANTITY vs BOUNDARY_OMISSION is the distinction to get right, and it is the one a scale comparison alone cannot make. In a boundary omission the thing is simply not in the model. In a proxy divergence the thing is in the model, but as a stand-in of a different quantity — transcriptional regulation of dopamine synthesis standing in for striatal dopamine concentration. Both can occur at the same scale, so neither is recoverable from model_scale.

4. Generalization to NAM and animal models

Deliberately out of scope for the first implementation, but the taxonomy was chosen so it extends rather than forks. BOUNDARY_OMISSION (organoid without vasculature or immune compartment), PROXY_QUANTITY, CALIBRATION_PROVENANCE, POPULATION_MISMATCH and SPECIES_MISMATCH all apply unchanged. Extending to the animal set would likely add SUPRAPHYSIOLOGICAL_EXPRESSION, INCOMPLETE_PHENOTYPE, and a genotype-level sibling of CAUSE_UNREPRESENTED — each already visible in the keyword probe at 20–123 hits, so each would be evidenced the same way rather than assumed.

5. Prior art

The ask is standard practice in regulatory computational modeling. ASME V&V 40, an FDA-recognized standard, and the FDA's Assessing the Credibility of Computational Modeling and Simulation guidance anchor every credibility activity in a stated question of interest, context of use, and quantity of interest, and require an applicability analysis — the relevance of the validation evidence to the context of use. A dismech ModelMechanismLink is a context-of-use statement ("this model is informative for this mechanism"), and a typed divergence list is its applicability analysis. Two of the enum values map almost directly: PROXY_QUANTITY is a quantity-of-interest mismatch, and POPULATION_MISMATCH / CALIBRATION_PROVENANCE are applicability gaps.

The framework also grades credibility requirements by model risk = model influence × decision consequence. materiality is the dismech analogue at the level of a single divergence: does this particular gap bear on this particular claim?