Historical calibration artifact. This document describes Comparator v3.0 before the Comparator v3.1 implementation. Current instrument status is recorded in comparator-v3.1-approval-record.md, comparator-v3.1-regression-report.md, and the generated-v3.1 outputs.

Historical Status: Superseded as current-state guidance by Comparator v3.1 implementation and regression results.

Constraint Calibration Audit

Scope

This audit reviews:

• comparison/generated/constraint-stability-report.md
• canonical EXP-001 response files

Current output marks every constraint field as disagreement.

Headline Assessment

The current constraint comparator is not yet measuring the intended construct consistently.

Most of the reported disagreement appears to come from:

• schema differences
• wording differences
• field-placement differences
• comparator strictness

Not primarily from genuine procedural disagreement.

Constraint Field Review

Invariants

Observed behavior:

• some models treat pass-through-node properties as invariants
• others treat reconvergence at N3 as the invariant
• others treat ordering of differentiating input before re-sorting as the invariant

Audit judgment:

• mostly schema and wording driven
• not strong evidence of true disagreement

Preconditions

Observed behavior:

• some responses include only START -> N1
• others add "differentiating input appears before re-sorting"
• others move those conditions into validity conditions instead

Audit judgment:

• mostly field-placement and schema difference

Postconditions

Observed behavior:

• some responses describe termination through END
• others describe strongest-path closure
• others merge the two

Audit judgment:

• mostly wording and comparator strictness

Stopping criteria

Observed behavior:

• some responses treat continuation not justified as a stopping criterion
• some treat one path remains strongest as a stopping criterion
• some list both

Audit judgment:

• partly genuine modeling difference
• mostly schema sensitivity because the packet naturally supports both criteria

Validity conditions

Observed behavior:

• some responses use this field for "re-sorting occurs after differentiating input"
• others place that same idea under invariants or preconditions

Audit judgment:

• mostly schema difference and comparator strictness

Cause Classification

Conclusion

Constraint comparison should not be frozen in its current scoring form.

What is trustworthy:

• the comparator correctly exposes that models distribute similar ideas across different constraint fields

What is not trustworthy:

• the conclusion that constraint preservation is broadly poor

Calibration recommendation:

• constraint metrics should remain visible
• but current disagreement labels should not yet be interpreted as reliable evidence of construct-level divergence