MechInterp Labeler

Manage the feature labeling workflow. This skill provides tools for:

• Priority queue management
• Setting and syncing labels
• Finding similar features
• Tracking labeling progress

Purpose

The labeler skill enables interactive feature labeling sessions:

1. Get the next feature to label from a priority queue
2. Use overview and experiments to understand the feature
3. Save labels with categories and notes
4. Find similar features to label next
5. Track overall progress

Commands

Get Next Feature

cd /root/dev/SplatNLP

# Get next feature from queue
poetry run python -m splatnlp.mechinterp.cli.labeler_cli next --model ultra

# Don't auto-build queue if empty
poetry run python -m splatnlp.mechinterp.cli.labeler_cli next --model ultra --no-build

Set a Label

IMPORTANT: Always use

--source

Source Options:

• claude code

  — Label created through Claude Code CLI investigation

• codex

  — Label created through Codex (OpenAI) agent

• codex/claude

  — Label created through Codex orchestrating Claude

• manual

  — Label created by human manually

• dashboard

  — Label created through dashboard UI (default)

# Label from Claude Code investigation
poetry run python -m splatnlp.mechinterp.cli.labeler_cli label \
    --feature-id 18712 \
    --name "Special Charge Stacker" \
    --model ultra \
    --source "claude code"

# With category and notes
poetry run python -m splatnlp.mechinterp.cli.labeler_cli label \
    --feature-id 18712 \
    --name "SCU Detector" \
    --category tactical \
    --notes "Responds to Special Charge Up presence, stronger at high AP" \
    --source "claude code"

# Manual labeling by human
poetry run python -m splatnlp.mechinterp.cli.labeler_cli label \
    --feature-id 18712 \
    --name "My Label" \
    --source "manual"

Categories:

• mechanical

  : Low-level patterns (token presence, combinations)

• tactical

  : Mid-level patterns (build strategies, weapon synergies)

• strategic

  : High-level patterns (playstyle, meta concepts)

• none

  : Uncategorized

Required Label Fields

Every label in

consolidated_ultra.json

feature_id

model_type

dashboard_name

dashboard_category

dashboard_notes

display_name

last_updated

source

hypothesis_confidence

importance_percentile

interpretability_confidence

stability_score

research_label

research_state_path

Separating Importance from Interpretability

These three fields capture distinct dimensions:

importance_percentile

interpretability_confidence

stability_score

Common combinations:

Rule of thumb: Don't conflate these. A feature with 9th percentile importance but 0.85 interpretability confidence is "weak but understood" - useful for pattern recognition but not a major model component.

Example complete label:

{
  "feature_id": 10938,
  "model_type": "ultra",
  "dashboard_name": "Positional Survival - Midrange",
  "dashboard_category": "strategic",
  "dashboard_notes": "Survival through positioning, not stealth/trading. Decoder promotes: SSU, BRU (all levels), ISS, IA, IRU. Suppresses: BPU, RSU, QR, SS. Weapons: Midrange with NO/BAD NS fit, LOW death tolerance. NS 0.84x depleted, QR 0.66x suppressed.",
  "display_name": "Positional Survival - Midrange",
  "last_updated": "2025-12-14T01:30:00.000000",
  "source": "claude code (full investigation)",
  "hypothesis_confidence": 0.85,
  "importance_percentile": 9.3,
  "interpretability_confidence": 0.85,
  "stability_score": null,
  "research_label": "Positional Survival - Midrange",
  "research_state_path": "/mnt/e/mechinterp_runs/state/feature_10938_ultra.json"
}

⚠️ Super-Stimuli Warning

High activations may be "flanderized" versions of the true concept!

When labeling features, don't only examine extreme activations. High activation builds can be:

• Super-stimuli: Extreme, exaggerated versions of the core concept
• Weapon-gated: Only achievable on specific niche weapons
• Unrepresentative: Missing the general pattern that applies across weapons

How to Detect Super-Stimuli

1. Examine activation regions (as % of effective max = 99.5th percentile):

   • Floor (≤1%), Low (1-10%), Below Core (10-25%)
   • Core (25-75%), High (75-90%), Flanderization Zone (90%+)
   • Use effective max to prevent outliers from distorting region boundaries

2. Look for weapons that span ALL levels continuously:

   • If Splattershot appears in every region → feature encodes a general concept
   • If only niche weapons reach 90%+ → those are "super-stimuli"

3. Compare core (25-75%) vs flanderization zone (90%+):

   • Core region: diverse weapons, general builds = TRUE CONCEPT
   • Flanderization zone: concentrated on 3-4 special-dependent weapons = SUPER-STIMULI

Example: Feature 9971

Initial label (wrong): "Death-Averse SCU Stacker"
- Only looked at 90%+ activations (SCU_57 + special-dependent weapons)

Better label: "Offensive Intensity (Death-Averse)"
- Core region (25-75%) showed diverse weapons (Splattershot family, Sploosh, Hydra)
- Feature tracks general offensive investment, not specifically SCU
- Flanderization zone (90%+) with Bloblobber, Glooga are "super-stimuli" not the core concept

Key insight: The core region (25-75% of effective max) reveals the TRUE feature concept. High activations (90%+ of effective max) show what happens when that concept is pushed to flanderized extremes.

Core Coverage Validation (BEFORE LABELING)

Before finalizing any label, verify core coverage of the proposed signature.

A label based on a token/ability that only appears in <30% of core examples is labeling the TAIL, not the concept.

from splatnlp.mechinterp.skill_helpers import load_context
import polars as pl
import numpy as np

ctx = load_context('ultra')
df = ctx.db.get_all_feature_activations_for_pagerank(FEATURE_ID)

# Define core region
acts = df['activation'].to_numpy()
nonzero_acts = acts[acts > 0]
effective_max = np.percentile(nonzero_acts, 99.5)
core_df = df.filter(
    (pl.col('activation') > 0.25 * effective_max) &
    (pl.col('activation') <= 0.75 * effective_max)
)

# Check coverage of proposed label driver
driver_id = ctx.vocab['YOUR_TOKEN_HERE']  # e.g., 'respawn_punisher'
core_with_driver = core_df.filter(
    pl.col('ability_input_tokens').list.contains(driver_id)
)

coverage = len(core_with_driver) / len(core_df) * 100
print(f"Core coverage: {coverage:.1f}%")

Red flags that indicate wrong labeling:

• Binary ability with >5x tail enrichment but <20% core presence → tail marker
• Weapon with >40% in top-100 but <15% in core → flanderized
• Proposed signature covers <30% of core examples → incomplete interpretation

Example (Feature 13934):

Wrong approach: See RP with 8.57x enrichment → label as "RP Backline Anchor"
Reality: RP only in 12% of core → RP is super-stimulus, not concept

Right approach: Check core coverage FIRST
→ RP at 12% means it's a tail marker
→ Split by RP presence to find true concept
→ Label the commonality across modes

Label Quality Examples

Evolution from Mechanical to Strategic

Good vs Bad Labels

The Strategic Label Test

Before saving a label, ask:

1. "Would a competitive Splatoon player recognize this playstyle?"

   • If no → too mechanical or wrong terminology

2. "Does this explain WHY the model learned this pattern?"

   • If no → you're describing correlation, not causation

3. "Could I explain this to someone who doesn't know the tokens?"

   • If no → label is too technical

Mandatory Label Components

Every strategic/tactical label should have:

1. Core concept - The gameplay behavior (e.g., "Positional Survival")
2. Role qualifier - Where/how it's played (e.g., "- Midrange")
3. Notes with evidence - Decoder groups, weapon classification, key enrichments

Label Specificity by Category

Match label specificity to concept level:

• Mechanical = low-level pattern → precise, token-focused
• Tactical = build strategy → names the combo + weapon/class
• Strategic = gameplay philosophy → high-concept + role qualifier

Skip a Feature

# Skip the next feature
poetry run python -m splatnlp.mechinterp.cli.labeler_cli skip --model ultra

# Skip specific feature with reason
poetry run python -m splatnlp.mechinterp.cli.labeler_cli skip \
    --feature-id 18712 \
    --reason "ReLU floor too high, hard to interpret"

Add Features to Queue

# Add single feature
poetry run python -m splatnlp.mechinterp.cli.labeler_cli add 18712 --model ultra

# Add multiple with priority
poetry run python -m splatnlp.mechinterp.cli.labeler_cli add 18712,18890,19042 \
    --priority 0.8 \
    --reason "SCU-related cluster"

Find Similar Features

poetry run python -m splatnlp.mechinterp.cli.labeler_cli similar \
    --feature-id 18712 \
    --top-k 5 \
    --model ultra

Check Status

poetry run python -m splatnlp.mechinterp.cli.labeler_cli status --model ultra

Output example:

## Labeling Status (ultra)

### Labels
- Total labeled: 45
- From dashboard: 30
- From research: 10
- Merged: 5

### Categories
- tactical: 20
- mechanical: 15
- strategic: 5
- uncategorized: 5

### Queue
- Pending: 25
- Completed: 40
- Skipped: 5

Sync Labels

Pull labels from all sources (dashboard, research states):

poetry run python -m splatnlp.mechinterp.cli.labeler_cli sync --model ultra

Export Labels

poetry run python -m splatnlp.mechinterp.cli.labeler_cli export \
    --model ultra \
    --output /mnt/e/mechinterp_runs/labels/export.csv

Build Priority Queue

# By activation count (features with most data)
poetry run python -m splatnlp.mechinterp.cli.labeler_cli build-queue \
    --model ultra \
    --method activation_count \
    --top-k 50

# From cluster (similar to a seed feature)
poetry run python -m splatnlp.mechinterp.cli.labeler_cli build-queue \
    --model ultra \
    --method cluster \
    --seed 18712 \
    --top-k 10

Typical Labeling Session

User: Let's label some features

Claude: [runs: labeler_cli next --model ultra]
        Next feature: 18712 (priority: 0.85)

        [runs: overview_cli --feature-id 18712]
        ## Feature 18712 Overview
        - Top token: special_charge_up (27%)
        - Family: SCU 31%
        ...

        Based on the overview, this feature appears to detect
        Special Charge Up stacking. Want me to run a sweep?

User: Yes, confirm with an SCU sweep

Claude: [runs: runner_cli with family_1d_sweep]
        Results confirm monotonic increase with SCU.

        Proposed label: "SCU Detector - High AP"

User: Call it "Special Charge Stacker"

Claude: [runs: labeler_cli label --feature-id 18712
               --name "Special Charge Stacker" --category tactical
               --source "claude code"]
        Label saved (source: claude code). Finding similar features...

        [runs: labeler_cli similar --feature-id 18712]
        Similar features:
        - 19042 (sim=0.82)
        - 18890 (sim=0.75)

        Want to add these to the queue?

Label Storage

Labels are stored in three places (kept in sync):

1. Dashboard:

   src/splatnlp/dashboard/feature_labels_{model}.json

2. Research State:

   /mnt/e/mechinterp_runs/state/{model}/f{id}.json

3. Consolidated:

   /mnt/e/mechinterp_runs/labels/consolidated_{model}.json

The consolidator merges all sources and resolves conflicts.

Queue Storage

Queue state is persisted at:

• /mnt/e/mechinterp_runs/labels/queue_{model}.json

Contains:

• Pending entries with priorities
• Completed feature IDs
• Skipped feature IDs

Programmatic Usage

from splatnlp.mechinterp.labeling import (
    LabelConsolidator,
    LabelingQueue,
    QueueBuilder,
    SimilarFinder,
)

# Queue management
queue = LabelingQueue.load("ultra")
entry = queue.get_next()
queue.mark_complete(entry.feature_id, "My Label")

# Set labels
consolidator = LabelConsolidator("ultra")
consolidator.set_label(
    feature_id=18712,
    name="SCU Detector",
    category="tactical",
    notes="Responds to SCU presence",
)

# Find similar
finder = SimilarFinder("ultra")
similar = finder.find_by_top_tokens(18712, top_k=5)

# Build queue
builder = QueueBuilder("ultra")
queue = builder.build_by_activation_count(top_k=50)

See Also

• mechinterp-overview: Quick feature overview before labeling
• mechinterp-runner: Run experiments to validate hypotheses
• mechinterp-state: Track detailed research progress
• mechinterp-summarizer: Generate notes from experiments