Mechinterp State

MechInterp State Management

Manage the research state for mechanistic interpretability analysis of SAE features. Track hypotheses, link evidence, maintain history, and generate summaries.

Purpose

This skill provides persistent research state management:

• Create and track hypotheses about feature behavior
• Link experimental evidence to hypotheses
• Maintain a chronological history of research actions
• Generate summaries and export notes

When to Use

Use this skill to:

• Start a new research investigation on a feature
• Add or update hypotheses based on observations
• Record evidence from experiments
• Get a summary of current research progress
• Export notes for documentation

State Location

Research state is stored at:

/mnt/e/mechinterp_runs/state/feature_{id}_{model}.json

Notes are exported to:

/mnt/e/mechinterp_runs/notes/feature_{id}_{model}.md

Operations

Initialize or Load State

from splatnlp.mechinterp.state import ResearchStateManager

# Load existing state or create new
manager = ResearchStateManager(feature_id=18712, model_type="ultra")

# Check current state
print(manager.get_summary())

Add Hypothesis

# Add a new hypothesis
h = manager.add_hypothesis(
    statement="Feature 18712 detects high SCU investment (>= 41 AP)",
    confidence=0.6,
    tags=["family-specific", "threshold-based"]
)
print(f"Created hypothesis {h.id}")

Update Hypothesis

# Update confidence based on evidence
manager.update_hypothesis(
    h_id="h001",
    confidence_delta=+0.1,  # Increase by 10%
    status=HypothesisStatus.TESTING
)

# Or set absolute confidence
manager.update_hypothesis(
    h_id="h001",
    confidence_absolute=0.8,
    status=HypothesisStatus.SUPPORTED
)

Add Evidence

# Link evidence from an experiment
from splatnlp.mechinterp.schemas.research_state import EvidenceStrength

evidence = manager.add_evidence(
    experiment_id="20250607_142531",
    result_path="/mnt/e/mechinterp_runs/results/20250607_142531__result.json",
    summary="SCU family sweep shows +0.35 mean delta at rungs >= 41",
    strength=EvidenceStrength.STRONG,
    supports=["h001"],  # Hypothesis IDs this supports
    key_metrics={"mean_delta": 0.35, "effect_size": 1.2}
)

Add from Experiment Result

# Directly from ExperimentResult object
from splatnlp.mechinterp.schemas import ExperimentResult

result = ExperimentResult.model_validate_json(result_path.read_text())
evidence = manager.add_evidence_from_result(
    result=result,
    supports=["h001"],
    strength=EvidenceStrength.MODERATE
)

Record Pitfalls

# Note things to avoid in future experiments
manager.add_pitfall("ReLU floor detected at low activation - avoid weapon gating")
manager.add_pitfall("Multi-rung SCU already present in 30% of base contexts")

Get Summary

# Get current research summary
summary = manager.get_summary()
print(summary)

# Example output:
# # Research State: Feature 18712
# Model: ultra
# Label: unlabeled
#
# ## Hypotheses (2 active)
# - [h001] (testing, 70%) Feature 18712 detects high SCU investment
# - [h002] (proposed, 50%) Secondary response to ISS at high SCU
#
# ## Evidence (3 items)
# - [e001] SCU family sweep shows +0.35 mean delta...
# ...

Export Notes

# Export to Markdown file
notes_path = manager.export_notes()
print(f"Notes exported to {notes_path}")

Get Next Experiment Suggestions

# Get suggestions based on current state
suggestions = manager.get_next_experiment_suggestions()
for s in suggestions:
    print(f"- {s}")

CLI Usage

# View state summary
cd /root/dev/SplatNLP
poetry run python -c "
from splatnlp.mechinterp.state import ResearchStateManager
m = ResearchStateManager(18712, 'ultra')
print(m.get_summary())
"

# List all states
poetry run python -c "
from splatnlp.mechinterp.state.io import list_states
for fid, model, path in list_states():
    print(f'{model}/{fid}: {path}')
"

State Schema

{
  "feature_id": 18712,
  "model_type": "ultra",
  "feature_label": "SCU threshold detector",
  "hypotheses": [
    {
      "id": "h001",
      "statement": "Feature detects SCU >= 41 AP",
      "status": "testing",
      "confidence": 0.7,
      "supporting_evidence": ["e001", "e002"],
      "refuting_evidence": [],
      "tags": ["family-specific"]
    }
  ],
  "evidence_index": [
    {
      "id": "e001",
      "experiment_id": "20250607_142531",
      "result_path": "/mnt/e/...",
      "summary": "SCU sweep shows threshold at 41 AP",
      "strength": "strong",
      "key_metrics": {"mean_delta": 0.35}
    }
  ],
  "active_constraints": ["one_rung_per_family"],
  "known_pitfalls": ["relu_floor_at_low_activation"],
  "history": [...],
  "notes": "Free-form research notes..."
}

Hypothesis Status Flow

proposed -> testing -> supported
                   \-> refuted
                   \-> superseded (by new hypothesis)

See Also

• mechinterp-glossary-and-constraints: Domain knowledge reference
• mechinterp-next-step-planner: Propose experiments
• mechinterp-runner: Execute experiments
• mechinterp-summarizer: Generate research notes from results