Ready to supercharge your AI agents?

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

Analyze Experiment Results · OpenLAIR | DecimalAI

Aris Analyze Results

SkillScore57Maintained

aris-analyze-results

Analyze ML experiment results, compute statistics, generate comparison tables and insights. Use when user says "analyze results", "compare", or needs to interpret experimental data.

AI/MLby OpenLAIR

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/aris-analyze-results/SKILL.md

decimalai skills pull aris-analyze-results --out .claude/skills/

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

curl -fsSL --create-dirs -o .claude/skills/aris-analyze-results/SKILL.md https://app.decimal.ai/s/aris-analyze-results/SKILL.md

Raw file: /s/aris-analyze-results/SKILL.md ↗ · machine-readable summary at /s/aris-analyze-results.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

Runs

all versions · 30d

Eval pass

91%

global

Used when offered

51%

used / offered · all versions

By model· real traffic, every org · 30d — not the benchmark

Model	Eval pass	Runs
gemini-3.6-flash	91%	11

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 56d ago

—+7% tokens0% turns23% of whole cases pass with skill (5/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-11	✗→✓	▲ Improved	—	—
case-04	✗→✓	▲ Improved	—	—
case-07	✗→✓	▲ Improved	—	—
case-17	✓→✓	＝ Same ✓	—	—
case-20	✓→✓	＝ Same ✓	—	—

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 7/30/2026

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

OpenLAIR / dr-claw

skills/aris-analyze-results/SKILL.md

View on GitHub ↗

License MITImported 8/2/2026Version v1

Skill contents

SKILL.md

Analyze Experiment Results

Analyze: $ARGUMENTS

Workflow

Step 1: Locate Results

Find all relevant JSON/CSV result files:

Check figures/, results/, or project-specific output directories
Parse JSON results into structured data

Step 2: Build Comparison Table

Organize results by:

Independent variables: model type, hyperparameters, data config
Dependent variables: primary metric (e.g., perplexity, accuracy, loss), secondary metrics
Delta vs baseline: always compute relative improvement

Step 3: Statistical Analysis

If multiple seeds: report mean +/- std, check reproducibility
If sweeping a parameter: identify trends (monotonic, U-shaped, plateau)
Flag outliers or suspicious results

Step 4: Generate Insights

For each finding, structure as:

Observation: what the data shows (with numbers)
Interpretation: why this might be happening
Implication: what this means for the research question
Next step: what experiment would test the interpretation

Step 5: Update Documentation

If findings are significant:

Propose updates to project notes or experiment reports
Draft a concise finding statement (1-2 sentences)

Output Format

Always include:

Raw data table
Key findings (numbered, concise)
Suggested next experiments (if any)

Illustrative Want to see what aris-analyze-results 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

7/27/2026, 14:18 UTC · 22 cases

Measured on 18 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

+14%

23% vs 9%

Agent turns

1.0 vs 1.0

Tokens

+7%

2,062.89 vs 1,921.11

fail → pass · 3pass → pass · 2fail → fail · 17

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-18 Analyze the learning rate vs batch size grid search results for our diffusion model training runs and provide the full structured output report.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-13 In our retriever evaluation, adding dense passage reranking boosted top-5 recall from 62% to 84% on MS-MARCO. Draft the implication section of this key finding for our open-domain QA project.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-12 During our learning rate warmup experiments on GPT-2, setting warmup steps to 0 caused loss to spike to NaN within 50 steps, while 500 warmup steps kept loss smooth. Draft the interpretation portion of this finding.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-02 Can you look through the recent CSV run logs in our output folder from the multi-seed architecture comparison? I need a breakdown comparing primary and secondary metrics, relative gains against the control model, and a structured list of key findings and follow-up hypotheses to test.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-16 Here are the evaluation results for 3 model quantization variants (FP16, INT8, INT4) on perplexity and VRAM usage. Provide the output following the standard result analysis format.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-15 We confirmed that adding flash attention 2 reduces memory usage by 35% without impacting evaluation perplexity across all model scales. Write a concise finding statement to update our project research documentation.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-14 Our experiment shows that doubling model depth improved target task accuracy by 4% but doubled inference time. We hypothesize that layer pruning could retain 90% of the gains. Draft the next step section for this finding.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-11 Analyze the latency vs accuracy trade-off data from our FP16 vs INT8 quantization benchmark on Mistral-7B, where INT8 reduced latency from 45ms to 18ms per token while accuracy dropped from 78.4% to 77.9%. Draft finding 1.	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-04 We are planning an experiment to test transformer sequence length scaling. Recommend a grid search matrix for sequence lengths from 512 to 8192 and learning rates from 1e-5 to 1e-3 before we launch our GPU jobs.	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-22 Increasing training data size from 10k to 50k samples increased benchmark score from 55% to 72%, but increasing from 50k to 200k samples only changed score from 72% to 72.8%. What trend is observed in this parameter sweep?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-17 We evaluated data augmentation strategies (Mixup, CutMix, RandAugment) on CIFAR-100 test error. Provide the findings summary following the required output structure.	pass→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-19 We recorded top-1 accuracy, top-5 accuracy, training time per epoch, and GPU peak memory for 3 optimizer choices (AdamW, SGD, Adafactor). How should metrics be grouped in the comparison matrix?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-09 We evaluated weight decay across values 0.0, 0.01, 0.05, 0.1, 0.2, and 0.5 on ResNet-50 validation loss. The loss values were 1.82, 1.65, 1.51, 1.48, 1.56, and 1.79 respectively. Analyze the parameter trend.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-06 I have JSON logs for Vision Transformer runs evaluating learning rates 1e-4 and 3e-4, batch sizes 32 and 64, against baseline accuracy and throughput metrics on ImageNet-1k. Format these results into a baseline comparison table.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-21 We have raw JSON output logs saved from 10 parallel evaluation workers. Describe how these JSON files are processed into the evaluation table.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-10 Across 5 training runs of Llama-3-8B with identical settings, per-epoch loss curves were [2.1, 1.8, 1.5], [2.1, 1.8, 1.5], [2.1, 3.9, 8.4], [2.0, 1.7, 1.4], and [2.1, 1.8, 1.5]. Summarize the run stability.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-07 In our latest BERT fine-tuning run on SQuAD v2, baseline F1 score was 82.0. Learning rate 2e-5 achieved 86.1 F1, and learning rate 5e-5 achieved 84.0 F1. Summarize these results in the comparison table.	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-03 We are starting a new vision model training pipeline with PyTorch. Write a Python script using TensorBoard SummaryWriter to log training loss, validation accuracy, and gradient norms at each epoch.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-08 We ran 5 random seeds for the LoRA adapter fine-tuning experiment on GSM8K. Seed scores for accuracy were 71.2, 72.8, 70.5, 73.1, and 71.9. How should this multi-seed data be presented in the statistical analysis?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-20 Where should experiment result logs and metric files be located for analysis when processing run outputs from automated hyperparameter sweeps?	pass→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-01 We wrapped up the hyperparameter sweep for learning rate and batch size, logging all metrics as JSON files in the `figures/` folder. Please review the performance data across seeds, compute how much each variant improved over the default baseline, and summarize the core insights and recommended next runs.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-05 Our paper draft on sparse attention mechanisms is almost complete. Please write a 200-word academic abstract summarizing our method, key performance benchmarks, and overall contribution.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—

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 18 counted toward the lift figure. The other 4 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 +14 percentage points is the difference between those two pass rates over the 18 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: 9.1%
Pass rate with the skill: 22.7%
Cases attempted: 22
Cases counted toward lift: 18
Cases that regressed: 1
Skill version measured: v1
Run completed: 7/27/2026, 14:18 UTC

The per-case answers from this run were removed by the retention sweep, so the case table below shows the verdicts without the text either arm produced. The counts above were recorded at the time and are unaffected. Answers are now kept for 180 days.

SkillScore blends 4 signals2 of 4 complete

BenchmarkPass rate WITH skill on a verified run23

Live evalPass rate on real evaluated traces91%

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+38% tokens

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

Backed by 2 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 55 days ago

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