H016 Review

• Hypothesis ID: H016
• Name: Structural Recognition
• Current statement: Models may recognize procedural families from topology or structural cues alone.
• Research question: Does recognition persist from structure even when higher-level semantic cues are reduced?
• Prior direction and strength: active / low
• Prior kill-condition status: not_tested

Evidence Supporting

• EXP-001 retained mixed backbone outputs under topology perturbation while recognition categories varied. Source: research/evidence-runs/ECR-000003-representation-sensitivity/experiments/EXP-001-topology-perturbation/comparison/generated-v3.1/exp-001-comparison-summary.md
• Explainability recorded repeated recognition_without_structural_effect cases. Source: research/evidence-runs/ECR-000003-representation-sensitivity/comparison/ECR-000003-explainability-summary.md

Evidence Challenging

• EXP-001 still produced mixed backbone rather than consistently stable topology-only recovery. Source: research/evidence-runs/ECR-000003-representation-sensitivity/experiments/EXP-001-topology-perturbation/comparison/generated-v3.1/exp-001-comparison-summary.md
• No accepted pure graph-only control is part of the final ECR-000003 evidence set.

Mixed Evidence

• ECR-000003 keeps the structural-recognition question active, but does not isolate it cleanly.

Provider, Representation, and Domain Effects

• Some providers may be more sensitive to structural cues than others, but current evidence is not clean enough to separate that from general recognition bias.

Relevant Instrument Findings

• Comparator outputs can show recognition states and structural stability side by side, but the underlying packets were not designed as a decisive structure-only test.

Threats

• topology may still carry recognizable meaning
• perturbations may not be neutral
• missing graph-only control in accepted evidence

Alternative Explanations

• Recognition may be driven more by residual semantic cueing than by topology alone.
• Apparent structural recognition could be a byproduct of generic procedural priors.

Kill Condition

• Current kill condition: graph-only topology variants consistently reduce recognition while preserving extracted structure.
• Status: not_tested

Proposed Direction

waiting

Proposed Evidence Strength

insufficient

Board Recommendation

defer

Rationale

This hypothesis remains live but not reviewable to conclusion from accepted ECR-000003 evidence alone.

Next Evidence Needed

• graph-only recognition-control replication
• low-cue topology controls
• targeted provider comparison

Human Decision

• Decision:
• Reviewer:
• Date: