Exp Lens Sensitivity Robustness

Sensitivity & Robustness Experimental Design Lens

Philosophical Mode: Robustness Primary Question: "Which assumptions are load-bearing?" Focus: Ablation Structure, Preprocessing Sensitivity, Metric Sensitivity, Hyperparameter Sensitivity, Distribution Shift

When to Use

• Results may depend on specific preprocessing choices
• Need to verify robustness of conclusions across conditions
• Ablation study seems incomplete or cherry-picked
• User invokes /exp-lens-sensitivity-robustness or /make-experiment-diag sensitivity

Critical Constraints

NEVER:

• Modify any source code or experiment files
• Do not litter the codebase with useless comments, TODO markers, or explanatory annotations — the skill output and diagram speak for themselves
• Treat "untested" as equivalent to "robust"

ALWAYS:

• Build a full sensitivity matrix (choices x perturbation types)
• Classify every analytic choice as load-bearing, minor, or untested
• Flag cases where the most impactful choices are the least tested
• Distinguish between ablations that were run and choices that were simply fixed
• BEFORE creating any diagram, LOAD the /mermaid skill using the Skill tool - this is MANDATORY

---

Analysis Workflow

Step 1: Launch Parallel Exploration Subagents

Spawn Explore subagents to investigate:

Analytic Choices Made

• Find all decision points in the analysis pipeline
• Look for: choice, option, default, parameter, threshold, method, alternative

Ablation Coverage

• Find which factors have been ablated
• Look for: ablation, without, remove, disable, vary, sweep, drop

Preprocessing Variations

• Find preprocessing steps that could be done differently
• Look for: normalize, tokenize, augment, crop, resize, filter, clean, impute

Hyperparameter Sensitivity

• Find which hyperparameters were tuned vs fixed
• Look for: learningrate, batchsize, epochs, dropout, hidden_size, temperature, alpha, beta

Distribution/Environment Variations

• Find evidence of testing under different conditions
• Look for: shift, domain, transfer, cross, outofdistribution, generalize, different

Step 2: Build the Sensitivity Matrix

Rows = analytic choices. Columns = perturbation types (remove, change, stress).

For each cell: Does the conclusion survive the perturbation?

Step 3: Classify Analytic Choices

CRITICAL — Analyze Assumption Load: For every analytic choice:

• What happens if this choice were made differently?
• Is there evidence from ablations, sweeps, or prior literature?
• Are the most impactful choices the least tested?

Classify each choice as:

• Load-bearing: Conclusion changes if this choice changes
• Minor: Conclusion is robust to changes in this choice
• Untested: No evidence either way

Step 4: Create the Optional Perturbation Diagram

If a diagram adds value, create a simplified flowchart. This is OPTIONAL for this hybrid lens — the tables are the primary output.

Direction: TB (choices flow down through perturbation to conclusion stability)

Subgraphs: "ANALYTIC CHOICES", "PERTURBATIONS TESTED", "CONCLUSION STABILITY"

Node Styling:

• stateNode class: analytic choice nodes
• handler class: perturbation type nodes
• output class: stable conclusion nodes
• gap class: load-bearing untested choice nodes
• detector class: sensitivity threshold nodes

Step 5: Write Output

Write the analysis to: temp/exp-lens-sensitivity-robustness/exp_diag_sensitivity_robustness_{YYYY-MM-DD_HHMMSS}.md

---

Output Template

# Sensitivity & Robustness Analysis: {Experiment Name}

**Lens:** Sensitivity & Robustness (Robustness)
**Question:** Which assumptions are load-bearing?
**Date:** {YYYY-MM-DD}
**Scope:** {What was analyzed}

## Sensitivity Matrix

| Analytic Choice | Remove | Change Value | Stress Test | Overall Classification |
|----------------|--------|-------------|-------------|------------------------|
| {choice name} | Stable / Fragile / Untested | Stable / Fragile / Untested | Stable / Fragile / Untested | Load-bearing / Minor / Untested |

## Load-Bearing Assumptions

| Assumption | Evidence Type | Impact if Changed | Tested? |
|------------|--------------|-------------------|---------|
| {assumption} | Ablation / Sweep / Literature / None | High / Medium / Low | Yes / No |

## Ablation Coverage Assessment

| Factor | Ablated? | Result | Interpretation |
|--------|----------|--------|----------------|
| {factor name} | Yes / No | {delta metric if yes} | Conclusion holds / Fragile / Unknown |

## Robustness Profile

| Dimension | Status | Notes |
|-----------|--------|-------|
| Preprocessing choices | Robust / Fragile / Untested | {detail} |
| Hyperparameter choices | Robust / Fragile / Untested | {detail} |
| Metric choices | Robust / Fragile / Untested | {detail} |
| Distribution shift | Robust / Fragile / Untested | {detail} |

## Perturbation Diagram (Optional)

```mermaid
%%{init: {'flowchart': {'nodeSpacing': 50, 'rankSpacing': 60, 'curve': 'basis'}}}%%
flowchart TB
    %% CLASS DEFINITIONS %%
    classDef cli fill:#1a237e,stroke:#7986cb,stroke-width:2px,color:#fff;
    classDef stateNode fill:#004d40,stroke:#4db6ac,stroke-width:2px,color:#fff;
    classDef handler fill:#e65100,stroke:#ffb74d,stroke-width:2px,color:#fff;
    classDef phase fill:#6a1b9a,stroke:#ba68c8,stroke-width:2px,color:#fff;
    classDef newComponent fill:#2e7d32,stroke:#81c784,stroke-width:2px,color:#fff;
    classDef output fill:#00695c,stroke:#4db6ac,stroke-width:2px,color:#fff;
    classDef detector fill:#b71c1c,stroke:#ef5350,stroke-width:2px,color:#fff;
    classDef gap fill:#ff6f00,stroke:#ffa726,stroke-width:2px,color:#000;
    classDef integration fill:#c62828,stroke:#ef9a9a,stroke-width:2px,color:#fff;

    subgraph Choices ["ANALYTIC CHOICES"]
        CHOICE1["Preprocessing Choice━━━━━━━━━━{e.g., normalization}"]
        CHOICE2["Model Choice━━━━━━━━━━{e.g., architecture}"]
        LOAD_BEARING["Load-Bearing Choice━━━━━━━━━━{untested critical choice}"]
    end

    subgraph Perturbations ["PERTURBATIONS TESTED"]
        PERTURB1["Remove━━━━━━━━━━Ablation"]
        PERTURB2["Change Value━━━━━━━━━━Sensitivity sweep"]
        THRESHOLD["Sensitivity Threshold━━━━━━━━━━{delta that changes conclusion}"]
    end

    subgraph Stability ["CONCLUSION STABILITY"]
        STABLE["Stable Conclusion━━━━━━━━━━{holds under perturbation}"]
    end

    CHOICE1 --> PERTURB1
    CHOICE2 --> PERTURB2
    PERTURB1 --> THRESHOLD
    PERTURB2 --> THRESHOLD
    THRESHOLD --> STABLE
    LOAD_BEARING -.->|"untested risk"| THRESHOLD

    class CHOICE1,CHOICE2 stateNode;
    class LOAD_BEARING gap;
    class PERTURB1,PERTURB2 handler;
    class THRESHOLD detector;
    class STABLE output;

Color Legend: | Color | Category | Description | |-------|----------|-------------| | Teal | Analytic Choices | Decision points in the pipeline | | Yellow | Load-Bearing Untested | Critical choices with no perturbation evidence | | Orange | Perturbations | Types of tests applied | | Red | Sensitivity Thresholds | Points where conclusion may change | | Dark Teal | Stable Conclusions | Results robust to perturbation |

Recommendations

1. {Most urgent ablation to run — highest impact untested choice}
2. {Preprocessing sensitivity test needed}
3. {Distribution shift or domain generalization test needed}

---

## Pre-Diagram Checklist

Before creating the diagram, verify:

- [ ] LOADED `/mermaid` skill using the Skill tool
- [ ] Using ONLY classDef styles from the mermaid skill (no invented colors)
- [ ] Diagram will include a color legend table

---

## Related Skills

- `/make-experiment-diag` - Parent skill for lens selection
- `/mermaid` - MUST BE LOADED before creating diagram
- `/exp-lens-estimand-clarity` - For clarifying which conclusions are being stress-tested
- `/exp-lens-iterative-learning` - For tracking robustness improvements across experiment iterations

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [Trecek](https://github.com/Trecek)
- **Source:** [Trecek/useful-claude-skills](https://github.com/Trecek/useful-claude-skills)
- **License:** MIT

Install and usage instructions live in the source repository linked above.