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Historical Status: Superseded as current-state guidance by Comparator v3.1 implementation and regression results.

Structural Agreement Audit

Scope

This audit reviews the structural disagreement behavior reported in:

• comparison/generated/structural-stability-report.md
• comparison/generated/exp-001-comparison-summary.md
• canonical response files used by EXP-001

Headline Assessment

The structural comparator is measuring something real, but the disagreement presentation is not calibrated to the intended construct.

Observed problem:

• every provider-scope and variant-scope row is labeled disagreement-heavy
• the overall structural profile is still mostly_stable

That combination strongly suggests that literal structural comparison is over-counting wording and decomposition differences as structural disagreement.

Disagreement Categories

A. Literal only

These are the dominant cases.

Examples:

• entry_conditions
  • "Procedure begins at START and enters N1"
  • "START node is reached"
  • "Execution starts at START node leading directly to A7"
• exit_conditions
  • some providers merge both exits into one sentence
  • others split them into separate items
• required_steps
  • some responses list graph nodes
  • others list graph actions
  • others combine node plus action

Interpretation:

• The underlying topology is usually preserved.
• The comparator is mostly reacting to phrasing and decomposition choices.

B. Dimensional only

Few clear examples were observed.

Interpretation:

• Dimensional normalization is not the main source of false structural disagreement here.
• Most structural instability is created before dimensional comparison, at the field/value representation level.

C. True structural disagreement

Some real structural divergence does exist, but it is narrower than the current report implies.

Examples:

• P001D introduces pass-through nodes N1a/N1b/N1c and N5a/N5b/N5c.
  • Some responses treat them as explicit structural steps.
  • Others compress them into the higher-level expansion and loop.
• Claude P001B introduces an extra inferred exit interpretation:
  • "Procedure may exit at END via X1 when threshold is not met"
  • That adds structural semantics beyond the packet's explicit topology.

Interpretation:

• These are real disagreements, but they are localized.
• They do not justify the blanket impression that all provider or variant comparisons are structurally unstable.

D. Possible ontology deficiency

Examples:

• entry_conditions and exit_conditions show persistent dimensional weakness even when all responses obviously refer to start/end topology.
• The role-first normalization does not appear strong enough to collapse:
  • begin / enter / start / initiate
  • exit / terminate / close / reach END

Interpretation:

• Structural disagreement is inflated partly because the dimensional ontology is not yet strong on boundary-condition language.

E. Comparator defect

Most important defect:

• Scope-level summary language (disagreement-heavy) does not reflect the actual research question for EXP-001.

Why:

• EXP-001 is about topology perturbation stability.
• The current structure comparator gives little weight to backbone preservation and too much weight to field-level wording divergence.

Audit Summary

Conclusion

Structural comparison is directionally useful, but not yet calibrated for scientific freezing.

What is trustworthy:

• the comparator does detect malformed or missing data
• it preserves the distinction between literal and dimensional comparison
• it surfaces real pass-through-node and inferred-exit differences

What is not yet trustworthy:

• the severity implied by disagreement-heavy
• the use of field-level textual divergence as a proxy for structural instability

Recommended calibration target before v3.1:

• separate backbone preservation from explanatory phrasing variance
• keep literal reporting, but do not let it dominate the structural headline