H013 Review

• Hypothesis ID: H013
• Name: Recognition Bias
• Current statement: Models may reconstruct familiar frameworks from prior knowledge rather than from packet structure alone.
• Research question: Is recognition materially driving recovered structure?
• Prior direction and strength: active / moderate
• Prior kill-condition status: not_tested

Evidence Supporting

• EXP-003 produced leakage findings and persistent recognition for some providers. Source: research/evidence-runs/ECR-000003-representation-sensitivity/experiments/EXP-003-isomorphic-procedures/comparison/generated-v3.1/comparison-summary.md
• Explainability recorded repeated recognition_without_structural_effect and terminology_leakage_only cases, confirming recognition pressure remains active in the dataset. Source: research/evidence-runs/ECR-000003-representation-sensitivity/comparison/ECR-000003-explainability-summary.md

Evidence Challenging

• Recognition varied while EXP-003 backbone remained mostly_stable, which weakens the claim that recognition is always required for structural recovery. Source: research/evidence-runs/ECR-000003-representation-sensitivity/comparison/ECR-000003-explainability-summary.md
• EXP-001 included not_recognized, partial, and unknown states while still yielding comparator outputs. Source: research/evidence-runs/ECR-000003-representation-sensitivity/pipeline/generated/ecr-000003-comparison-summary.md

Mixed Evidence

• Recognition pressure is real, but its effect on recovered structure appears inconsistent rather than universally dominant.

Provider, Representation, and Domain Effects

• Provider differences appear large enough that recognition bias may not be a single unified phenomenon.
• Domain language and representation cues likely both contribute to recognition.

Relevant Instrument Findings

• Comparator 3.1.0 can distinguish recognition category differences from structural agreement categories.
• Explainability suggests recognition should be analyzed separately from structural recovery rather than collapsed into one signal.

Threats

• topology may still be recognizable
• domain terms may cue known frameworks
• no fully accepted recognition-control replication yet

Alternative Explanations

• Some apparent recognition effects may actually be general reasoning priors rather than artifact recognition.
• Shared procedural schemas may make some framework families naturally easier to reconstruct even without recognition.

Kill Condition

• Current kill condition: better-blinded or novel artifacts preserve agreement without recognition.
• Status: tested_not_met

Proposed Direction

mixed

Proposed Evidence Strength

moderate

Board Recommendation

split

Rationale

The evidence supports keeping recognition bias active, but the phenomenon is broad enough that lexical recognition, structural recognition, and structural import should be reviewed separately.

Next Evidence Needed

• recognition-control replication
• negative non-isomorphic controls
• explicit graph-only or low-cue controls

Human Decision

• Decision:
• Reviewer:
• Date: