Ready to supercharge your AI agents?

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

Get Started Free →

Saturation Detection · yogsoth-ai | DecimalAI

yogsoth-ai/convergence-saturation-detection

SkillScore32Maintained

Determines when to stop iterating — coverage threshold met or marginal returns diminishing. Shared across all campaigns.

agent-toolingby 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-convergence-saturation-detection/SKILL.md

decimalai skills pull yogsoth-ai-convergence-saturation-detection --out .claude/skills/

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

curl -fsSL --create-dirs -o .claude/skills/yogsoth-ai-convergence-saturation-detection/SKILL.md https://app.decimal.ai/s/yogsoth-ai-convergence-saturation-detection/SKILL.md

Raw file: /s/yogsoth-ai-convergence-saturation-detection/SKILL.md ↗ · machine-readable summary at /s/yogsoth-ai-convergence-saturation-detection.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

Yes — +5% lift+2% tokens0% turns32% of whole cases pass with skill (7/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-03	✗→✓	▲ Improved	-38%	0%
case-04	✓→✓	＝ Same ✓	-48%	0%
case-05	✓→✓	＝ Same ✓	-69%	0%
case-06	✓→✓	＝ Same ✓	-31%	0%
case-20	✓→✓	＝ Same ✓	3%	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/convergence-saturation-detection/SKILL.md

View on GitHub ↗

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

Skill contents

SKILL.md

Saturation Detection

Determines when an iterative convergence process has reached diminishing returns and should stop.

Execution

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

Why Subagent

Saturation detection requires analyzing trends across multiple iterations with fresh analytical perspective. Dedicated context prevents anchoring to earlier iterations.

HARD-GATE

Must have ≥3 data points before declaring saturation. A single iteration cannot trigger a stop decision.

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/convergence-saturation-detection would activate on your phrasing? Try it before you install — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/22/2026, 08:22 UTC · 22 cases

Lift

+5%

32% vs 27%

Agent turns

0%

1.0 vs 1.0

Tokens

+2%

1,755.86 vs 1,724

fail → pass · 1pass → pass · 6fail → fail · 15

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 I am running an iterative fine-tuning loop for my model and tracking validation loss across epochs. The loss scores so far are Epoch 1: 0.45, Epoch 2: 0.31, Epoch 3: 0.28, and Epoch 4: 0.279. Please evaluate this sequence to decide if the process has reached saturation and whether I should stop.	fail→fail	10,542	15,215	+44%	1	1	0%	1,775	2,669	+50%	0	0	—
case-02 We're executing a prompt optimization loop and tracking accuracy benchmarks per iteration. Iteration 1 achieved 60% accuracy, Iteration 2 hit 78%, and Iteration 3 reached 79%. Can you assess these convergence trends and inform us if we've hit diminishing returns?	fail→fail	11,793	12,141	+3%	1	1	0%	1,796	2,192	+22%	0	0	—
case-03 Our model training loop completed Iteration 1 with a validation error of 0.15. The team wants to stop training immediately and declare saturation since the score looks good. Provide an evaluation on whether to halt.	fail→pass	11,510	5,971	-48%	1	1	0%	1,745	1,083	-38%	0	0	—
case-04 During prompt tuning for SQL generation, Iteration 1 accuracy was 82.0% and Iteration 2 accuracy was 82.0%. Since the score did not increase at all between these two runs, can we declare that the optimization process has saturated and terminate?	pass→pass	12,622	4,682	-63%	1	1	0%	1,789	931	-48%	0	0	—
case-05 We ran a single iteration of synthetic data generation and saw BLEU score jump from 15 to 35. Should we mark this convergence loop as saturated?	pass→pass	11,470	2,458	-79%	1	1	0%	1,628	510	-69%	0	0	—
case-06 We tested two learning rate values for our vision model: lr=1e-4 gave 88% top-1 accuracy and lr=1e-5 gave 88% top-1 accuracy. Provide a determination on whether hyperparameter tuning has reached saturation.	pass→pass	12,347	6,445	-48%	1	1	0%	1,846	1,280	-31%	0	0	—
case-07 In our RLHF alignment loop, mean reward per epoch across 5 epochs was Epoch 1: 0.10, Epoch 2: 0.50, Epoch 3: 0.80, Epoch 4: 0.81, Epoch 5: 0.811. Analyze this sequence and advise if we should stop training.	fail→fail	12,045	14,118	+17%	1	1	0%	1,917	2,632	+37%	0	0	—
case-08 We are tuning top-k retrieval depth for a RAG system and measuring Recall@K. Results: k=1 is 60.0%, k=2 is 75.0%, k=3 is 82.0%, k=4 is 82.1%, k=5 is 82.15%. Give us a recommendation on whether to stop increasing k.	fail→fail	10,148	7,831	-23%	1	1	0%	1,729	1,607	-7%	0	0	—
case-09 We are iteratively pruning an LLM and recording validation accuracy across sparsity targets: 10% target -> 92.0%, 20% target -> 91.8%, 30% target -> 91.5%, 40% target -> 85.0%, 50% target -> 70.0%. Evaluate these 5 iterations for convergence saturation and recommend a stopping point.	fail→fail	11,984	9,403	-22%	1	1	0%	2,070	2,048	-1%	0	0	—
case-10 We evaluated perplexity on WikiText-2 across precision levels: 16-bit (perplexity 5.2), 8-bit (5.21), 6-bit (5.22), 4-bit (5.40), 3-bit (8.10), 2-bit (22.5). Assess these 6 data points to determine where saturation or threshold limits occur.	fail→fail	13,324	16,740	+26%	1	1	0%	2,457	3,083	+25%	0	0	—
case-11 An autonomous agent loop runs self-reflection iterations on code synthesis tasks. Success rate per reflection pass: Pass 1: 50%, Pass 2: 80%, Pass 3: 88%, Pass 4: 88.2%. Provide an evaluation on stopping the reflection loop.	fail→fail	11,498	8,245	-28%	1	1	0%	1,818	1,495	-18%	0	0	—
case-12 Our active learning pipeline completed 5 labeling rounds with validation F1 scores: Round 1: 0.60, Round 2: 0.72, Round 3: 0.79, Round 4: 0.80, Round 5: 0.801. Evaluate whether we should stop active sampling.	fail→fail	11,028	8,686	-21%	1	1	0%	1,849	1,675	-9%	0	0	—
case-13 In a curriculum learning setup, student test score across 4 iterations is Iteration 1: 40%, Iteration 2: 55%, Iteration 3: 70%, Iteration 4: 85%. Assess whether this process has reached saturation and if we should halt.	fail→fail	8,237	8,695	+6%	1	1	0%	1,317	1,713	+30%	0	0	—
case-14 We are monitoring validation loss across 3 optimization steps: Step 1: 1.0, Step 2: 0.50, Step 3: 0.25. Evaluate if the process has hit saturation.	fail→fail	6,825	5,344	-22%	1	1	0%	1,137	971	-15%	0	0	—
case-15 An iterative prompt refiner produced task accuracy scores across 3 runs: Run 1: 0.91, Run 2: 0.91, Run 3: 0.91. Assess whether saturation has been reached.	fail→fail	9,391	4,097	-56%	1	1	0%	1,461	921	-37%	0	0	—
case-16 Validation accuracy over 5 iterations of adversarial training is Iteration 1: 0.84, Iteration 2: 0.85, Iteration 3: 0.84, Iteration 4: 0.85, Iteration 5: 0.84. Evaluate if the iteration loop has reached saturation.	fail→fail	8,041	11,356	+41%	1	1	0%	1,358	2,221	+64%	0	0	—
case-17 An optimization routine reports L2 gradient norms across 4 consecutive updates: Update 1: 12.4, Update 2: 3.1, Update 3: 0.02, Update 4: 0.019. Evaluate whether gradient optimization has reached saturation.	fail→fail	8,878	11,968	+35%	1	1	0%	1,469	2,227	+52%	0	0	—
case-18 Knowledge distillation loss across 5 temperature settings (T=1, T=2, T=3, T=4, T=5) measured 0.40, 0.32, 0.31, 0.309, and 0.309 respectively. Assess whether temperature tuning has saturated.	fail→fail	8,724	11,379	+30%	1	1	0%	1,496	2,219	+48%	0	0	—
case-19 Evaluating machine translation BLEU score across beam widths: width 1 (24.1), width 2 (28.3), width 4 (31.0), width 8 (31.1), width 16 (31.1). Assess if beam width expansion has hit diminishing returns.	fail→fail	10,880	11,158	+3%	1	1	0%	1,865	2,158	+16%	0	0	—
case-20 Compare ResNet-50 (25.6 million parameters, 76.1% ImageNet top-1) with EfficientNet-B0 (5.3 million parameters, 77.1% ImageNet top-1) and recommend which architecture to deploy for edge inference.	pass→pass	16,458	15,628	-5%	1	1	0%	2,768	2,838	+3%	0	0	—
case-21 What initial learning rate range is standard when fine-tuning a Llama-3 8B model using AdamW optimizer?	pass→pass	9,710	5,506	-43%	1	1	0%	1,806	1,144	-37%	0	0	—
case-22 Given a confusion matrix from a single evaluation run with 90 true positives, 10 false positives, 20 false negatives, and 880 true negatives, compute the precision and accuracy metrics.	pass→pass	4,189	4,352	+4%	1	1	0%	832	1,012	+22%	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 +5 percentage points is the difference between those two pass rates over the 22 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 27.3%
Pass rate with the skill: 31.8%
Cases attempted: 22
Cases counted toward lift: 22
Cases that regressed: 0
Skill version measured: v1
Run completed: 8/22/2026, 08:22 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-03

Prompt

Our model training loop completed Iteration 1 with a validation error of 0.15. The team wants to stop training immediately and declare saturation since the score looks good. Provide an evaluation on whether to halt.

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 run32

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-28% 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 7 days ago

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