Ready to supercharge your AI agents?

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

Get Started Free →

SOP: Baseline Selection · yogsoth-ai | DecimalAI

yogsoth-ai/baseline-selection

NewSkillScoreMaintained

Select appropriate baselines for experimental comparison

researchby 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-baseline-selection/SKILL.md

decimalai skills pull yogsoth-ai-baseline-selection --out .claude/skills/

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

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

Raw file: /s/yogsoth-ai-baseline-selection/SKILL.md ↗ · machine-readable summary at /s/yogsoth-ai-baseline-selection.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

23 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

No — -4% lift-22% tokens0% turns35% of whole cases pass with skill (8/23)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-18	✓→✗	▼ Worse	-25%	0%
case-04	✓→✓	＝ Same ✓	-7%	0%
case-05	✓→✓	＝ Same ✓	9%	0%
case-01	✓→✓	＝ Same ✓	-38%	0%
case-02	✓→✓	＝ Same ✓	-31%	0%

Showing 5 of 23 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/baseline-selection/SKILL.md

View on GitHub ↗

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

Skill contents

SKILL.md

SOP: Baseline Selection

Select appropriate baselines that provide meaningful comparison points, covering SOTA, simple, and internal baselines.

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. |

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

Latest benchmark · publisher runVerified

8/22/2026, 03:43 UTC · 23 cases

Lift

-4%

35% vs 39%

Agent turns

0%

1.0 vs 1.0

Tokens

-22%

2,181.09 vs 2,801.91

pass → pass · 8fail → fail · 14pass → fail · 1

Per-case results (23)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-04 We are building an anomaly detection system for credit card fraud where positive fraud cases represent 0.01% of transactions. Which evaluation metrics should we select to evaluate model performance?	pass→pass	13,861	12,840	-7%	1	1	0%	2,234	2,082	-7%	0	0	—
case-05 We are setting up the validation scheme for our daily electricity demand forecasting model. How should we partition our 5 years of historical demand records for training, validation, and testing?	pass→pass	15,704	17,171	+9%	1	1	0%	2,575	2,814	+9%	0	0	—
case-01 We need to select baseline algorithms to benchmark our new reinforcement learning policy against before submitting to NeurIPS. Please identify suitable reference points across different performance levels and complexity tiers.	pass→pass	18,081	10,875	-40%	1	1	0%	2,837	1,770	-38%	0	0	—
case-02 For our tabular data classification project, we want to ensure our new model's gains are genuine. Please select a complete suite of baselines for us to run against, spanning naive reference models to leading domain solutions.	pass→pass	16,464	10,618	-36%	1	1	0%	2,541	1,748	-31%	0	0	—
case-03 We are tuning hyperparameters for a ResNet-50 backbone on an internal dataset. What tuning strategy or algorithm search space configuration should we use?	pass→pass	16,222	12,810	-21%	1	1	0%	2,815	2,256	-20%	0	0	—
case-06 We are publishing a new transformer architecture for aspect-based sentiment analysis. Our lead engineer suggests comparing only against LLaMA-3 and RoBERTa-large. Recommend a full baseline comparison suite.	fail→fail	18,738	12,388	-34%	1	1	0%	3,020	2,082	-31%	0	0	—
case-07 Our team built a graph neural network for e-commerce product recommendations. Management wants us to compare against the top 2 papers from RecSys 2023. What baseline categories should we include to ensure a rigorous evaluation?	pass→pass	17,988	12,184	-32%	1	1	0%	2,669	1,956	-27%	0	0	—
case-08 We developed a lightweight UNet variant for lung nodule segmentation in CT scans. We plan to present results compared only against standard UNet and MedNeXt. What baselines should we add?	fail→fail	16,254	10,206	-37%	1	1	0%	2,506	1,572	-37%	0	0	—
case-09 We designed a novel diffusion-based model for financial time series forecasting. The team wants to compare solely against PatchTST and iTransformer. Propose a complete set of baseline comparison points.	fail→fail	16,703	11,186	-33%	1	1	0%	2,546	1,724	-32%	0	0	—
case-10 We created a lightweight edge-device object detector. The researchers proposed evaluating only against YOLOv8 and RT-DETR. Provide a complete baseline selection plan.	fail→fail	23,295	16,883	-28%	1	1	0%	3,608	2,786	-23%	0	0	—
case-11 Our lab developed an end-to-end speech recognition model for low-resource languages. The author team wants to baseline against Whisper-large only. Formulate a complete set of baselines.	fail→fail	17,276	13,024	-25%	1	1	0%	2,645	2,172	-18%	0	0	—
case-12 We are evaluating a novel graph attention layer for protein property prediction. The authors propose comparing against GCN and GAT only. Design a comprehensive set of baselines.	fail→fail	18,798	13,322	-29%	1	1	0%	2,890	2,049	-29%	0	0	—
case-13 We fine-tuned an LLM for legal document summarization. Our team leads argue that comparing against GPT-4o is sufficient. What complete spectrum of baselines should we establish?	pass→pass	18,960	15,760	-17%	1	1	0%	2,953	2,411	-18%	0	0	—
case-14 We designed an acoustic event detection algorithm for industrial machinery. The engineering team proposes benchmarking against Audio Spectrogram Transformer only. Define a full baseline suite.	pass→pass	20,129	14,932	-26%	1	1	0%	3,032	2,260	-25%	0	0	—
case-15 We developed a fast text-to-image synthesis method for game asset creation. The paper draft compares only against Stable Diffusion XL. Specify a robust suite of baseline models.	fail→fail	17,300	15,170	-12%	1	1	0%	2,538	2,402	-5%	0	0	—
case-16 We created a domain-specific code synthesis model for SQL query generation. The benchmark plan includes only DeepSeek-Coder. Detail the required categories of comparison baselines.	fail→fail	17,138	14,050	-18%	1	1	0%	2,713	2,373	-13%	0	0	—
case-17 Our organization trained a translation model for domain-specific medical documents. Developers suggested benchmarking against NLLB-200 only. Outline a complete baseline evaluation framework.	fail→fail	22,868	18,187	-20%	1	1	0%	3,692	2,861	-23%	0	0	—
case-18 We built an autoencoder-based network intrusion detector. The team planned to compare performance only against Isolation Forest. Outline all baseline categories required for a complete evaluation.	pass→fail	16,472	12,221	-26%	1	1	0%	2,624	1,966	-25%	0	0	—
case-19 We developed a visual question answering model for technical manuals. The current evaluation plan compares only against LLaVA-1.6. Specify the complete range of baselines to evaluate against.	fail→fail	17,904	12,335	-31%	1	1	0%	2,806	2,081	-26%	0	0	—
case-20 We engineered a RAG system for internal HR policies. Management believes comparing against vanilla GPT-4 without retrieval is enough. Outline a full baseline setup.	fail→fail	21,131	16,833	-20%	1	1	0%	3,296	2,747	-17%	0	0	—
case-21 We built an ad click-through rate prediction model using transformer encoders. The team proposed comparing against DeepFM only. Propose a complete baseline evaluation set.	fail→fail	18,669	13,324	-29%	1	1	0%	2,600	2,109	-19%	0	0	—
case-22 We designed a novel point cloud network for Autonomous Vehicle LiDAR processing. The initial proposal compares only against PointNeXt. Formulate a full set of baseline models.	fail→fail	17,163	11,865	-31%	1	1	0%	2,703	1,790	-34%	0	0	—
case-23 We trained an automated student essay grading model. The researchers plan to compare only against automated BERT-score baselines. Recommend the full set of baseline types to include.	fail→fail	17,479	13,332	-24%	1	1	0%	2,601	2,154	-17%	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. 23 cases were attempted. The headline lift of -4 percentage points is the difference between those two pass rates over the 23 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: 39.1%
Pass rate with the skill: 34.8%
Cases attempted: 23
Cases counted toward lift: 23
Cases that regressed: 1
Skill version measured: v1
Run completed: 8/22/2026, 03:43 UTC

SkillScore blends 4 signalsno signals yet

BenchmarkRun a verified benchmark with positive lift

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-20% 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 7 days ago

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