Get started

What IADTX is

IADTX is a structural diagnostic. It tests whether a decision domain composes additively or multiplicatively.

In one line

Upload tabular data, name the outcome column, and the engine returns a verdict on the structural composition of the domain. The verdict is one of five fixed readings, accompanied by a band reading per variable and a model-comparison panel.

What it is not

  • It is not a regulator-issued certification.
  • It is not a substitute for model validation under any of the frameworks listed below.
  • It does not produce predictions or scores for individual cases.

Why this matters

Every model review starts with a structural assumption. If the domain is additive, linear methods and additive ensembles are defensible. If the domain is multiplicative, the same methods systematically understate tail risk and the review needs a different toolkit. Picking the wrong assumption is the single largest source of avoidable model risk.

IADTX surfaces structural validation evidence that is 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 evidence supports the reviewer in stating, with citation, why a chosen modelling approach matches the domain it operates in.

What the engine does

  • Reads the file. Parses CSV, TSV, and Excel inputs against the upload contract.
  • Computes a band reading for every variable. Each reading is one of three closed values, namely below band, inside tie band, or above band.
  • Composes the readings into a verdict and surfaces a Model Comparison panel and the relevant lenses for the buyer's tier.

What the engine does not do

  • It does not store the uploaded file. Zero data retention is the operating contract; raw uploads are processed in memory and discarded after the response is sent.
  • It does not retrain a model. The submitted model is the unit under test, not a candidate for replacement.
  • It does not advise on remediation. When the verdict is MODEL_INADEQUATE or DATA_INSUFFICIENT, the engine surfaces an Improvement block describing what is missing, not what to do next.

Further reading

The First diagnosis walkthrough is the next page if the goal is to run the engine on a real file. The Bands reference is the next page if the goal is to understand what the engine reads. The Operations section, pending in a later commit, will host the public attestation language alongside the data retention contract.