Ready to supercharge your AI agents?

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

Get Started Free →

SOP: Level Specification · yogsoth-ai | DecimalAI

yogsoth-ai/level-specification

NewSkillScoreMaintained

Determine appropriate levels for each experimental factor

scienceby yogsoth-ai

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

Automatic

Research pack

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

pip install decimalai

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

decimalai skills pull yogsoth-ai-level-specification --out .claude/skills/

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

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

Raw file: /s/yogsoth-ai-level-specification/SKILL.md ↗ · machine-readable summary at /s/yogsoth-ai-level-specification.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 18d ago

No — -100% lift+7% tokens0% turns9% of whole cases pass with skill (2/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-21	✓→✗	▼ Worse	20%	0%
case-20	✓→✓	＝ Same ✓	31%	0%
case-22	✓→✓	＝ Same ✓	39%	0%
case-01	✗→✗	＝ Same ✗	10%	0%
case-02	✗→✗	＝ Same ✗	32%	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 — 8/10/2026

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/level-specification/SKILL.md

View on GitHub ↗

License Apache-2.0Imported 8/10/2026Version v1

Skill contents

SKILL.md

SOP: Level Specification

Determine the specific values (levels) at which each factor will be tested, including spacing strategy and justification.

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

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. |

Illustrative Want to see what yogsoth-ai/level-specification would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

9/3/2026, 00:15 UTC · 22 cases

Lift

-100%

0% vs 100%

Agent turns

0%

1.0 vs 1.0

Tokens

+7%

2,843.23 vs 2,665.59

pass → pass · 2fail → fail · 19pass → fail · 1

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-01 We are setting up a grid search experiment for a transformer model training pipeline where the independent factors are learning rate, batch size, and dropout rate. Could you outline the concrete settings for each variable, explain the spacing methodology used between points, and provide the underlying rationale for selecting these specific test points?	fail→fail	27,087	23,469	-13%	1	1	0%	3,087	3,400	+10%	0	0	—
case-02 We need to select 4 test values for learning rate spanning 1e-5 to 1e-2 for a ResNet fine-tuning experiment. A colleague suggested equal arithmetic spacing (0.00001, 0.00334, 0.00667, 0.01000). Specify the factor levels and explain the spacing strategy.	fail→fail	15,616	12,777	-18%	1	1	0%	2,034	2,684	+32%	0	0	—
case-03 For an LLM inference benchmark across GPU memory bounds, we want to test batch size between 4 and 64 across 5 levels. Provide the level values, spacing strategy, and justification.	fail→fail	10,149	13,683	+35%	1	1	0%	1,823	2,413	+32%	0	0	—
case-04 We are configuring weight decay test points across 1e-4 to 1e-1 for an AdamW optimizer on a vision transformer. Recommend 4 specific values, define the spacing strategy, and provide the rationale for the boundary points.	fail→fail	11,148	11,371	+2%	1	1	0%	2,246	2,230	-1%	0	0	—
case-10 For an SGD with momentum trial on ImageNet, choose 4 test levels for momentum between 0.85 and 0.99. Specify the levels, spacing method, and justification.	fail→fail	19,961	17,540	-12%	1	1	0%	2,691	2,493	-7%	0	0	—
case-05 For evaluating an LLM text generation quality metric, temperature will be varied between 0.1 and 1.0 across 4 settings. Provide the exact temperature levels, spacing strategy, and selection rationale.	fail→fail	16,682	8,859	-47%	1	1	0%	2,046	1,699	-17%	0	0	—
case-06 We are setting up an ablation study measuring perplexity vs model weight precision on edge devices. Select the test levels for bit precision between 2-bit and 16-bit, noting spacing strategy and hardware constraint rationale.	fail→fail	25,678	22,213	-13%	1	1	0%	3,303	2,917	-12%	0	0	—
case-07 A context window evaluation requires testing prompt lengths from 512 tokens to 32768 tokens across 5 test points. Output the level values, spacing method, and justification.	fail→fail	17,572	23,301	+33%	1	1	0%	3,322	3,402	+2%	0	0	—
case-08 We need 5 test points for dropout rate ranging from 0.0 to 0.5 in a dense classification network. Define the concrete levels, spacing strategy, and justification.	fail→fail	19,439	14,965	-23%	1	1	0%	2,630	2,715	+3%	0	0	—
case-09 In a pre-training experiment schedule of 100,000 total steps, set up 4 levels for warmup steps ranging from 500 to 10,000 steps. List the level values, spacing strategy, and rationale.	fail→fail	12,312	16,846	+37%	1	1	0%	2,303	2,325	+1%	0	0	—
case-11 We are defining gradient norm clipping threshold levels from 0.1 to 10.0 for stable RNN training across 4 levels. Outline the level values, spacing strategy, and rationale.	fail→fail	20,083	33,696	+68%	1	1	0%	3,061	1,894	-38%	0	0	—
case-12 For parameter-efficient fine-tuning of Llama-3 using LoRA, select 5 test levels for rank (r) between 4 and 64. Output the exact values, spacing strategy, and justification.	fail→fail	13,356	11,375	-15%	1	1	0%	1,495	1,132	-24%	0	0	—
case-13 When testing LoRA scaling factor alpha alongside rank r=16, determine 4 test values for alpha between 8 and 64. State the levels, spacing methodology, and structural rationale.	fail→fail	17,615	19,355	+10%	1	1	0%	2,265	2,625	+16%	0	0	—
case-14 Define 5 factor levels for top_p nucleus sampling between 0.50 and 1.00 for a creative writing generation task. List the specific levels, spacing strategy, and justification.	fail→fail	16,547	18,262	+10%	1	1	0%	2,063	2,409	+17%	0	0	—
case-15 For a data efficiency scaling law experiment, select 5 dataset size fraction levels between 1% and 100%. Detail the exact levels, spacing methodology, and rationale.	fail→fail	19,994	20,744	+4%	1	1	0%	2,579	3,809	+48%	0	0	—
case-16 We need 4 factor levels for hidden embedding size between 128 and 1024 for a lightweight retrieval model. Detail the chosen values, spacing strategy, and architectural justification.	fail→fail	20,956	19,286	-8%	1	1	0%	2,757	2,465	-11%	0	0	—
case-17 Set up 4 factor levels for beam search width between 1 and 10 for a machine translation evaluation. Output the exact levels, spacing strategy, and runtime complexity justification.	fail→fail	19,630	12,919	-34%	1	1	0%	2,507	2,377	-5%	0	0	—
case-18 We want to evaluate label smoothing epsilon between 0.0 and 0.2 across 4 levels in a cross-entropy loss classifier. State the concrete levels, spacing strategy, and statistical rationale.	fail→fail	14,026	15,634	+11%	1	1	0%	2,506	1,911	-24%	0	0	—
case-19 For a Beta-VAE disentanglement experiment, choose 5 factor levels for beta between 0.1 and 10.0. Provide the levels, spacing strategy, and loss balancing justification.	fail→fail	19,380	19,460	+0%	1	1	0%	2,426	2,385	-2%	0	0	—
case-20 We are preparing a machine learning experiment on tabular data. Identify which hyperparameter factors (e.g. learning rate, tree depth, subsample ratio) should be included in our study based on feature importance screening, before any level values are chosen.	pass→pass	22,591	27,368	+21%	1	1	0%	2,934	3,833	+31%	0	0	—
case-21 Here is a completed level specification matrix for our grid search (learning_rate: [1e-4, 1e-3], batch_size: [16, 32]). Execute the PyTorch model training runs for these 4 combinations on CUDA GPU 0 and report final validation losses.	pass→fail	18,814	20,346	+8%	1	1	0%	2,510	3,019	+20%	0	0	—
case-22 We ran a 3x3 factorial experiment testing learning rate and batch size against validation accuracy. Here are the 9 measured accuracy outputs: [0.72, 0.78, 0.81, 0.75, 0.83, 0.86, 0.70, 0.80, 0.84]. Perform an ANOVA statistical test to calculate p-values for main effects.	pass→pass	30,936	91,708	+196%	1	1	0%	6,055	8,414	+39%	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. The headline lift of -100 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 100%
Pass rate with the skill: 0%
Cases attempted: 22
Cases counted toward lift: 22
Cases that regressed: 1
Skill version measured: v1
Run completed: 9/3/2026, 00:15 UTC

SkillScore blends 4 signalsno signals yet

BenchmarkRun a verified benchmark with positive lift

Live evalNeeds ≥10 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+36% tokens

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

No signals yet — run a benchmark or collect evaluated traces to earn a score.

Verified on gemini-3.6-flash · re-tested 17 days ago

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