Reading the diagnosis
Verdicts
The verdict is one of five fixed readings, namely multiplicative, additive, mixed, model inadequate, or data insufficient. The resolver consumes three engine signals plus two precondition flags and returns one of these five labels with a locked headline.
The five verdicts
| Verdict | Locked headline | Buyer reading |
|---|---|---|
MULTIPLICATIVE |
The domain composes multiplicatively. | The structural test and the model fit agree that the variables compose multiplicatively. Any variable approaching zero collapses the outcome. |
ADDITIVE |
The domain composes additively. | The variables compose additively. Each contributes a near-independent share of the outcome and no single variable collapses the result. |
MIXED |
The domain has both additive and multiplicative substructure. | Different subsets of the variables compose under different rules. The reading is real and ships the Model Comparison panel with a mixed pill rather than a green or red one. |
MODEL_INADEQUATE |
The submitted model does not fit the data closely enough for a structural verdict. The items below indicate what is missing. | The fit quality is below the level needed for a structural reading. The Improvement block lists the closed labels under What is missing and the fixed-order suggestions under What would help. |
DATA_INSUFFICIENT |
The submitted dataset does not contain enough structural signal for a verdict. The items below indicate what is missing. | The dataset does not carry enough signal for the resolver to reach a verdict. The Improvement block lists the closed labels under What is missing and the fixed-order suggestions under What would help. |
How the resolver decides
The resolver consumes three signals from the engine. The domain classification from the structural test reads multiplicative, additive, mixed, or neutral. The primary statistic carries the model comparison signal. The fitness label carries the goodness-of-fit reading from the engine.
The resolver also reads two precondition flags. Sample size below the public minimum-row threshold short-circuits to DATA_INSUFFICIENT. Design-matrix rank deficiency does the same. The closed-vocabulary public surface for the resolver is primary_statistic, information_criterion_delta, explained_variance_lift, and the IADTX Structural Comparison Test verdict.
The numerical thresholds the resolver uses are private to the engine. The order in which contradicting signals are reconciled is also private. The page does not name either of those.
Nine resolver paths
The resolver enumerates nine outcome paths. Six of them are reachable through the live upload pipeline. The other three exist in the resolver but are guarded against on every uploaded path; they remain covered by the internal regression suite at tests/test_verdict_resolver.py.
| Path | Reachability |
|---|---|
MULTIPLICATIVE.three_way_agreement | Reachable |
ADDITIVE.three_way_agreement | Reachable |
MIXED.domain_mixed | Reachable |
MODEL_INADEQUATE.direction_contradiction | Reachable |
DATA_INSUFFICIENT.below_min_n | Reachable |
DATA_INSUFFICIENT.partial_neutral | Reachable |
MODEL_INADEQUATE.fitness_contradicting | Unreachable through the live pipeline |
DATA_INSUFFICIENT.rank_deficient | Unreachable through the live pipeline |
DATA_INSUFFICIENT.all_neutral | Unreachable through the live pipeline |
What each verdict implies for the review
The IADTX validation report and the diagnosis report are structured for review under SR 11-7, PRA SS1/23 Principle 3.4, EU AI Act Article 9 (high-risk per Annex III, applicable from August 2026), the ECB IM Guide, and APRA CPG 220. The framing is structured for review under, never satisfies, and never ready for submission to.
Multiplicative. The structural reading has a clear direction and the Model Comparison panel ships in full. Reviewers reading the report under SR 11-7 receive a Model Comparison panel with both R-squared values, the winner tile, and the IADTX Structural Comparison Test verdict.
Additive. The structural reading has a clear direction and the Model Comparison panel ships in full. The verdict pill reads in the additive token (red) by design; this is the locked semantic discipline and is not an error state for the report.
Mixed. The structural reading is real but composite. The Model Comparison panel ships with a mixed pill in the gold token. The reviewer reads this as a structurally meaningful reading that nonetheless requires sub-population disaggregation to act on under PRA SS1/23 Principle 3.4.
Model inadequate. The fit quality is below the structural threshold. The Model Comparison panel is replaced by the Improvement block. The reviewer reads this as a request for a tighter model fit before any structural reading is defensible under EU AI Act Article 9 documentation.
Data insufficient. The dataset does not carry enough structural signal. The reviewer reads this as a request for either more rows, more features, or a feature-correlation review before the engine can return a verdict. APRA CPG 220 Section 4 review evidence requirements are not met by an insufficient-data verdict.