H016 Hypothesis Review Record

Hypothesis ID: H016
Title: Structural Recognition
Evidence Run: ECR-000003 – Representation Sensitivity
Review Board: ECR-000003 Hypothesis Review Board
Status: Proposed – Pending Human Approval

Original Hypothesis

Language models are capable of recovering meaningful procedural structure from knowledge artifacts, and this structural recovery remains substantially stable despite changes in representation, topology, or domain.

Prediction

If H016 is correct, we expect to observe:

• Stable recovery of procedural backbone across representation changes.
• Stable recovery across topology-preserving transformations.
• Recovery of similar procedural organization despite domain changes.
• Greater agreement on high-level procedural structure than on wording or formatting.
• Structural similarities that persist even when recognition behavior differs.

Supporting Evidence

The review board identified the following observations as supporting H016.

EXP-001 – Topology Perturbation

Observed:

• Structural recovery remained largely stable despite topology-preserving perturbations.
• Node renumbering did not substantially alter recovered procedural organization.
• Edge ordering changes produced limited structural impact.
• Identity-node expansion primarily affected representation rather than backbone.

Interpretation:

These observations are consistent with recovery of procedural structure beyond literal graph representation.

EXP-002 – Cross Representation Stability

Observed:

• Multiple representations of equivalent procedures produced similar higher-level procedural recovery.
• Representation changes affected wording and decomposition more than overall procedural organization.

Interpretation:

Evidence is consistent with structural recognition that is partially independent of representation.

EXP-003 – Isomorphic Procedures

Observed:

• Procedural backbones remained broadly similar across distinct domains.
• Primitive role agreement exceeded complete procedural agreement.
• Domain vocabulary varied more than overall procedural organization.

Interpretation:

Evidence is consistent with partial recovery of shared procedural structure.

Cross-Experiment Pattern

Across all three experiments:

• Structural similarities appeared more stable than wording.
• Higher-level organization appeared more stable than detailed decomposition.
• Procedural recovery persisted despite multiple controlled perturbations.

Challenging Evidence

The review board identified several observations that limit confidence.

• Primitive sequence disagreements remained.
• Constraint disagreement remained significant.
• Backbone agreement was not uniformly complete across all providers.
• Domain effects were not completely isolated from structural effects.
• Human baseline comparisons are unavailable.
• Comparator calibration continues to evolve.

These observations reduce confidence but do not directly falsify the hypothesis.

Alternative Explanations Considered

Alternative A

Models recover semantic similarity rather than procedural structure.

Status:

Plausible.

Current experiments do not completely separate semantic similarity from structural recovery.

Alternative B

Shared training data produces similar outputs without requiring genuine structural recognition.

Status:

Possible.

Future negative-control experiments are needed.

Alternative C

Comparator ontology favors apparent structural agreement.

Status:

Possible.

Requires future calibration and independent validation.

Kill Condition Review

The current kill conditions are considered too broad.

The review board recommends distinguishing backbone-level failure from lower-level disagreement.

Recommended Kill Conditions

Substantially weaken or reject H016 if any of the following occur:

• Topology-preserving transformations consistently destroy recovered procedural backbone.
• Representation changes consistently eliminate structural agreement.
• Isomorphic procedures consistently produce unrelated procedural structures.
• Structural agreement collapses after controlling for domain vocabulary.
• Human analysts fail to recover comparable procedural organization under equivalent conditions.

Primitive-sequence disagreement alone should not constitute failure of H016.

Constraint disagreement alone should not constitute failure of H016.

These observations may instead identify limitations in structural recovery.

Board Assessment

Recommendation

Retain H016.

Evidence Direction

Supports the hypothesis, but only provisionally.

Suggested Evidence Strength

Weak to Moderate.

Confidence Change

Increase slightly.

The current evidence suggests structural recognition is plausible, but it does not yet isolate structural recognition from semantic familiarity or training effects.

Remaining Uncertainty

Outstanding questions include:

• How much recovery depends on semantic familiarity.
• Whether structural recovery persists in stronger negative controls.
• Whether humans exhibit similar recovery patterns.
• Whether comparator assumptions influence apparent backbone agreement.

Recommended Next Experiment

Design a negative-control experiment in which:

• semantic similarity is minimized,
• procedural topology is preserved,
• domain vocabulary is deliberately misleading.

The objective is to determine whether models recover underlying procedural structure or primarily respond to semantic familiarity.

Human Decision

Decision:

• Retain
• Strengthen
• Split
• Revise
• Suspend
• Retire

Reviewer:

Date:

Comments:

Final Notes

The review board concludes that ECR-000003 provides evidence consistent with structural recognition, but not sufficient evidence to establish structural recognition as an independent phenomenon.

Future experiments should focus on separating structural recognition from semantic familiarity, representation effects, and comparator bias before stronger conclusions are drawn.