apply-perturbation

SkillDev tools

Lets your agent test how a system reacts by changing one assumption or component at a time and recording the results.

Available today. Use it from your connected AI after setup.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Then ask your AI: use the apply-perturbation skill

About this skill

Apply a controlled change to an assumption, factor, component, parameter, or condition and record response.

What this skill tells your AI

The instructions your AI receives, as published by yogsoth-ai/de-anthropocentric-research-engine in skills/apply-perturbation/SKILL.md and read by ahel’s review.

Purpose

Apply a controlled change to an assumption, factor, component, parameter, or condition and record the response.

Input contract

required: [baseline_artifact, perturbation_target, perturbation_axis, response_metric]
optional: [variation_range, removal_mode, factor_list, uncertainty_model]
constraints: [change one declared target or axis at a time unless the caller explicitly supplies an interaction design; preserve baseline comparability]

Procedure

  1. Record the baseline artifact, target, axis, response metric, and comparison direction.
  2. Generate the caller-specified variation, removal, negation, or ablation conditions.
  3. Re-evaluate the response at each condition and attach evidence and uncertainty.
  4. Identify degradation, flip points, or threshold regions and summarize the effect.

Output contract

produces: [perturbation_series, response_comparison, degradation_or_flip_points, interpretation]
delta_fields: [findings, evidence_updates, decisions, uncertainties]

Quality gates

  • Controlled perturbation records performance at each point along one defined axis and identifies degradation thresholds where applicable.
  • Single-factor removal retains a degradation score from 0.0 (no effect) to 1.0 (collapse) and states conclusion before/after.
  • Ablation removes components one by one; do not combine removals while labeling the result single-factor.

Parameterization

The caller must provide the baseline artifact or system, perturbation target and axis, range or removal mode, response metric, factor list when needed, and degradation/flip classification rule.

Failure and counterexamples

Reject when the baseline is missing, the perturbation is not attributable, response measurements are incomparable, or a degradation score is reported without before/after reasoning.

Provenance map

  • resolved: deep-insight/controlled-perturbation
  • concept: creative-ideation/assumption-perturbation
  • resolved: creative-ideation/ablation-execution
  • resolved: stress-test/single-factor-removal

Verbatim source criteria excerpts

  • single-factor-removal line 35: degradation_score: 0.0 (no effect) to 1.0 (collapse)

Preserved source criteria ledger

sourcephysical linekindsource criterion
stress-test/single-factor-removal24numericDegradation score ranges from 0.0 (no effect) to 1.0 (collapse), with before/after conclusion and reasoning.

Signals

GitHub stars
501
Forks
41
Last commit
Sep 2026
Advanced
Catalog kind
skill
Key
apply-perturbation
Source
github.com/yogsoth-ai/de-anthropocentric-research-engine