Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Get Started Free →

Controlled Perturbation · yogsoth-ai | DecimalAI

yogsoth-ai/controlled-perturbation

SkillScore73Maintained

Systematically vary parameters along defined axes, recording performance at each point to identify degradation thresholds.

AI/MLby yogsoth-ai

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/yogsoth-ai-controlled-perturbation/SKILL.md

decimalai skills pull yogsoth-ai-controlled-perturbation --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/yogsoth-ai-controlled-perturbation/SKILL.md https://app.decimal.ai/s/yogsoth-ai-controlled-perturbation/SKILL.md

Raw file: /s/yogsoth-ai-controlled-perturbation/SKILL.md ↗ · machine-readable summary at /s/yogsoth-ai-controlled-perturbation.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 7d ago

—-51% tokens0% turns73% of whole cases pass with skill (16/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-04	✗→✓	▲ Improved	-14%	0%
case-06	✗→✓	▲ Improved	-65%	0%
case-08	✗→✓	▲ Improved	-70%	0%
case-09	✗→✓	▲ Improved	-76%	0%
case-10	✗→✓	▲ Improved	-30%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

yogsoth-ai / de-anthropocentric-research-engine

skills/controlled-perturbation/SKILL.md

View on GitHub ↗

License Apache-2.0Imported 8/3/2026Version v1

Skill contents

SKILL.md

Controlled Perturbation

Systematically perturb to find degradation thresholds.

Execution

Subagent — spawned via subagent-spawning/spawn-agent.

Budget

One unit = one perturbation curve along one axis.

Available SOPs

Optional, no fixed order; the final leaf is always a sop.

| SOP | When to use | | --- | --- | | spawn-agent | Spawn a customized CC subagent with full MCP tool access. Used by SOPs that declare execution: subagent. |

Want to see if yogsoth-ai/controlled-perturbation would activate on your phrasing? Try it before you install — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/22/2026, 19:54 UTC · 22 cases

Measured on 19 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+55%

73% vs 18%

Agent turns

0%

1.0 vs 1.0

Tokens

-51%

1,123.79 vs 2,290.84

fail → pass · 12pass → pass · 4fail → fail · 6

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-07 We need to evaluate an industrial anomaly detector against sensor drift by adding a linear drift baseline offset from 0 to 10 units in steps of 0.5. How is this single-axis test managed?	fail→fail	16,142	9,409	-42%	1	1	0%	2,469	1,753	-29%	0	0	—
case-01 I need to stress-test our text embeddings model by incrementally adding noise to the input prompts. Can you systematically introduce small variations along the noise axis to identify the exact point where retrieval accuracy starts dropping off significantly?	fail→fail	16,749	5,626	-66%	1	1	0%	3,067	410	-87%	0	0	—
case-02 We are tuning our inference latency parameters for the vision pipeline. Could you run a series of controlled input perturbations across the image resolution scale to determine where the model's classification performance degrades?	fail→fail	10,356	25,775	+149%	1	1	0%	1,959	4,511	+130%	0	0	—
case-03 I want to map out the failure boundaries for our financial time-series forecaster. Please apply controlled adjustments to the volatility parameter in steps and chart where the prediction error exceeds acceptable limits.	fail→fail	17,855	10,527	-41%	1	1	0%	3,159	538	-83%	0	0	—
case-04 We need to find the degradation threshold for our speech-to-text model by gradually lowering the signal-to-noise ratio in audio inputs. How should this execution be dispatched and budgeted?	fail→pass	37,118	6,266	-83%	1	1	0%	1,361	1,172	-14%	0	0	—
case-05 We want to test our LLM's long-context recall by adding distractor tokens in 500-token increments up to 32k tokens. How should this benchmark run be structured?	fail→fail	16,788	8,960	-47%	1	1	0%	2,781	1,608	-42%	0	0	—
case-06 We are stress-testing an autonomous driving perception model against fog and blur. We want to incrementally apply Gaussian blur sigma from 0.0 to 5.0 to detect where accuracy falls below 80%. What tool is used to run this task?	fail→pass	11,086	2,915	-74%	1	1	0%	1,870	647	-65%	0	0	—
case-08 Our collaborative filtering engine degrades when user interaction history is sparse. We plan to systematically drop user interaction history from 0% to 90% in 5% increments to pinpoint the collapse point. How should execution be delegated?	fail→pass	18,748	4,312	-77%	1	1	0%	2,707	812	-70%	0	0	—
case-09 We want to test how sensitive our code completion model is to masked context by masking 5% to 50% of precedent context lines. How many budget units are required for this single perturbation axis?	fail→pass	11,025	2,318	-79%	1	1	0%	1,931	461	-76%	0	0	—
case-10 We need to test how robust our chat model's safety classifier is under temperature variation from 0.0 to 1.5 in steps of 0.1. How should this execution process be handled?	fail→pass	16,577	9,752	-41%	1	1	0%	2,637	1,835	-30%	0	0	—
case-11 We plan to run two independent degradation tests: one along input token corruption rate and another along input sequence length. How many budget units does this require?	fail→pass	11,884	2,325	-80%	1	1	0%	2,226	483	-78%	0	0	—
case-12 To evaluate our tabular classifier's tolerance to missing data, we will randomly set feature values to null starting from 1% up to 50% in 2% steps. What subagent dispatch SOP should be used?	fail→fail	16,503	2,381	-86%	1	1	0%	2,657	554	-79%	0	0	—
case-13 We want to map image classifier accuracy against contrast reduction from 100% down to 10% in 5% steps. What execution strategy and tool should be specified?	fail→pass	15,496	2,219	-86%	1	1	0%	2,649	459	-83%	0	0	—
case-14 We are testing model degradation across three distinct perturbation axes: compression ratio, quantization bits, and crop size. How many budget units should be allocated in total?	fail→pass	9,863	2,185	-78%	1	1	0%	1,615	428	-73%	0	0	—
case-15 We want to test a real-time speech translation model's resilience to packet latency jitter by introducing artificial network delay from 10ms to 500ms in steps of 10ms. How is this run dispatched?	fail→pass	17,253	5,303	-69%	1	1	0%	2,810	1,050	-63%	0	0	—
case-16 We need to measure model performance degradation as semantic edit distance increases via paraphrase generation steps along a single edit distance axis. What is the budget allocation?	pass→pass	35,283	2,153	-94%	1	1	0%	2,976	428	-86%	0	0	—
case-17 We are sweeping the perturbation magnitude epsilon from 0.001 to 0.1 for an adversarial robustness curve on an image model. Which tool initiates the run?	fail→pass	8,281	3,885	-53%	1	1	0%	1,295	778	-40%	0	0	—
case-18 We need to evaluate an OCR engine's degradation threshold by rotating document images from 0 degrees to 45 degrees in 1 degree increments. How is the budget calculated?	fail→pass	13,451	4,624	-66%	1	1	0%	2,462	740	-70%	0	0	—
case-19 We want to measure degradation in response quality as unstructured weight pruning increases from 0% to 90% in 5% steps. What execution method should be used?	fail→pass	20,987	1,932	-91%	1	1	0%	3,373	418	-88%	0	0	—
case-20 We want to perform a joint grid search over learning rate, batch size, and weight decay simultaneously to find optimal hyperparameters for model training. Should this task be allocated a single-axis controlled perturbation budget?	pass→pass	12,732	7,080	-44%	1	1	0%	1,911	1,216	-36%	0	0	—
case-21 Can you run our standard validation suite on the uncorrupted MMLU test set and output the accuracy score? Should we calculate a controlled perturbation degradation budget for this run?	pass→pass	9,338	19,358	+107%	1	1	0%	1,430	2,405	+68%	0	0	—
case-22 We need to construct a single adversarial example using a one-step projected gradient descent attack against a target classifier. Does this single-step attack require spawning a subagent for controlled perturbation curve discovery?	pass→pass	11,687	4,209	-64%	1	1	0%	1,768	766	-57%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +55 percentage points is the difference between those two pass rates over the 19 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 18.2%
Pass rate with the skill: 72.7%
Cases attempted: 22
Cases counted toward lift: 19
Cases that regressed: 0
Skill version measured: v1
Run completed: 8/22/2026, 19:54 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-06

Prompt

We are stress-testing an autonomous driving perception model against fog and blur. We want to incrementally apply Gaussian blur sigma from 0.0 to 5.0 to detect where accuracy falls below 80%. What tool is used to run this task?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run73

Live evalNeeds ≥20 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns-40% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 6 days ago

Other measured skills in the registry, with their headline benchmark lift.