Comparator 3.2 Explainability

Purpose

Comparator 3.2 is a post-processing explainability layer for Comparator v3.1 outputs.

Comparator v3.1 answers:

Do these outputs agree?

Comparator 3.2 answers:

Why do they agree or disagree, and which differences matter?

Guardrail

Comparator 3.2 does not:

• rescore Comparator v3.1 outputs
• overwrite official v3.1 files
• update hypotheses automatically
• replace human review

Inputs

Comparator 3.2 requires:

• official Comparator v3.1 outputs
• normalized response files
• a valid normalization certificate

It refuses to run when official inputs are incomplete or when response files appear to have changed since the official v3.1 comparison.

Outputs

Per experiment, Comparator 3.2 writes:

• machine-readable pairwise difference records
• layer-specific explanation reports
• compression and elaboration analysis
• review-priority guidance
• EDR explainability input

At the ECR level it writes:

• a cross-experiment explainability summary
• an implementation report
• closure readiness

Commands

From the ECR root:

npm run explain:all
npm run reports:explainability

From an experiment directory:

npm run explain:v3.2

Limitations

• Explainability classifications are heuristic.
• Comparator 3.2 inherits Comparator v3.1 ontology limitations.
• Explanations may identify review targets without resolving them automatically.