Reading the diagnosis
Model Comparison panel
The Model Comparison panel sits on the result page beneath the band readings. It surfaces two structural readings side by side, namely the multiplicative and additive R-squared row, and the IADTX Structural Comparison Test row. The two together answer whether the structural reading from the band layer holds when the engine compares two competing fits.
What the panel shows
The panel renders for three of the five verdicts. MULTIPLICATIVE and ADDITIVE ship the panel in full, with the colour-correct verdict tile and a clear winner. MIXED ships the panel with the gold mixed pill in the structural row. MODEL_INADEQUATE and DATA_INSUFFICIENT replace the panel with the Improvement block; the Model Comparison readings do not render in those two cases because the underlying fit is not robust enough for a structural reading.
Row 1, the R-squared reading
The R-squared row carries three tiles. The Multiplicative R-squared tile shows the explanatory share of the multiplicative fit. The Additive R-squared tile shows the same for the additive fit. The Winner tile carries the comparative reading when both fits are available.
R-squared is a fit-quality reading. It is not a structural reading on its own. Two domains can both fit at high R-squared and still compose under different rules; the Winner tile is a comparison between the two specific fits the engine ran on this dataset, not a general claim about the domain.
Row 2, the IADTX Structural Comparison Test
The second row carries one verdict tile plus a prose paragraph. The verdict tile reads multiplicative, additive, mixed, or one of the two holding readings. The prose paragraph names the test in plain language.
The Mehrhoff Structural Comparison Test (MSCT) sits alongside R-squared. R-squared remains the headline accuracy term and tells you how well a model fits the data, but it cannot tell you whether one structural specification is more correct than another. MSCT provides that structural reading, empirically calibrated against a reference corpus of real-world model-data pairings with known structural truth, and returns an explicit inconclusive verdict when the evidence is not strong enough to decide. MSCT was invented by Patrick Mehrhoff, founder of IADTX.
The MSCT verdict in the result-page tile is one of multiplicative, additive, mixed, unclassifiable, or one of the two holding readings. The internal state behind the verdict remains private; the tile surfaces the closed-vocabulary verdict only.
What the panel does not disclose
- The numerical thresholds for the public band edges that decide whether a comparison reads above band, inside tie band, or below band.
- The minimum-row threshold for the structural test.
- The priority order in which contradicting signals are reconciled inside the resolver.
None of those values appears on the result page, on the validation report, on the diagnosis report, or in the API response. The buyer sees the closed-vocabulary reading only.
Screenshot walkthrough
Two paired screenshots. The first shows the panel for an ADDITIVE verdict against the shipped insurance sample. The second shows the panel for a MIXED verdict against the cross-model sample.
