Reading the diagnosis
Improvement blocks
The Improvement block sits in the panel area of the result page when the verdict is model inadequate or data insufficient. It carries two heads, What is missing and What would help, each populated with closed-vocabulary labels from the resolver.
When the block renders
The Improvement block renders only for two of the five verdicts, namely MODEL_INADEQUATE and DATA_INSUFFICIENT. For the other three verdicts (MULTIPLICATIVE, ADDITIVE, MIXED) the panel area is occupied by the Model Comparison panel instead. The Improvement block and the Model Comparison panel never render at the same time.
The two heads of the block
The block carries two heads. The What is missing head reads from the closed-vocabulary keys in WHAT_IS_MISSING_LABELS, exhaustive over the structural reading vocabulary. The What would help head reads from WHAT_WOULD_HELP_KEYS in the fixed order defined by _WHAT_WOULD_HELP_ORDER in the engine.
What is missing
- Information criterion delta below the public lower band
- Information criterion delta above the public upper band
- Information criterion delta inside the public tie band
- Sample size below the public minimum-row threshold
- Feature matrix is rank-deficient (collinear features)
- All variables read as neutral on the structural test
What would help
- Sample size
- Feature coverage
- Feature correlation structure
- Derived-covariate definitions
- Outcome variable definition
- Complete-case sample
The order under What would help is fixed by the engine. The order under What is missing is the closed-vocabulary order from READING_VOCAB followed by the precondition flags. The buyer-legible labels above are exhaustive over the resolver vocabulary; nothing else is rendered under either head.
What the block does not do
- The block does not advise on remediation specific to the buyer's domain or the buyer's particular dataset.
- The block does not recommend any vendor, library, or product.
- The block does not promise that following the suggestions will produce a multiplicative verdict on the next run.
The block is diagnostic, not prescriptive. The defensibility envelope under SR 11-7 and EU AI Act Article 9 documentation requires that the engine not give actionable advice that would amount to consulting; the closed-vocabulary labels stay neutral nouns for this reason.
Screenshot walkthrough
Two paired screenshots. The first shows the block for a DATA_INSUFFICIENT verdict, generated by truncating the insurance sample below the public minimum-row threshold. The second shows the block for a MODEL_INADEQUATE verdict, generated by replacing the outcome column of the insurance sample with a random vector.
