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Historical Status: Superseded as current-state guidance by Comparator v3.1 implementation and regression results.

Comparator Strength Register

Confirmed Strengths

CS-001 Role-first normalization remains directionally useful

Evidence:

• Despite severe literal disagreement, the overall dimensional structural profile is still mostly_stable.

Why it matters:

• This suggests the dimensional layer is preserving some intended construct separation rather than collapsing immediately into lexical mismatch.

CS-002 Literal versus dimensional separation is conceptually sound

Evidence:

• The comparator preserves distinct literal and dimensional outputs instead of hiding one behind the other.

Why it matters:

• This makes calibration possible.
• It prevents semantic smoothing from erasing raw disagreement.

CS-003 Recognition persistence can be tracked by provider and variant

Evidence:

• recognition-persistence-results.csv and recognition-persistence-report.md do capture provider-specific patterns across P001A-P001D.

Why it matters:

• EXP-001 specifically needs variant-by-variant recognition behavior.

CS-004 Packet version exclusion is scientifically correct

Evidence:

• Pre-fix P001D responses were excluded from the primary comparison.

Why it matters:

• This prevents packet-version contamination from being mistaken for model drift.

CS-005 Malformed response handling is visible

Evidence:

• The current data-quality report surfaces malformed P001A GPT and P001C Claude responses.

Why it matters:

• Invalid inputs are not silently treated as normal evidence.

CS-006 Observation-only EDR draft generation is appropriate

Evidence:

• comparison/generated/edr-draft.md carries metadata, observations, evidence inputs, and unresolved uncertainty without updating confidence.

Why it matters:

• This preserves the separation between evidence collection and hypothesis review.

CS-007 Domain leakage is tracked separately from recognition persistence

Evidence:

• leakage findings appear in their own report instead of overwriting recognition classification.

Why it matters:

• This helps preserve analytical separation between "what was recognized" and "what outside language was introduced."