Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add topprismdata/cultivating-ml-agent --skill learned-value-beats-heuristic-augmentationgit clone --depth 1 https://github.com/topprismdata/cultivating-ml-agentWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/topprismdata/cultivating-ml-agent/learned-value-beats-heuristic-augmentation)<a href="https://agentmods.dev/skills/topprismdata/cultivating-ml-agent/learned-value-beats-heuristic-augmentation"><img src="https://agentmods.dev/badge/skills/topprismdata/cultivating-ml-agent/learned-value-beats-heuristic-augmentation/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/topprismdata/cultivating-ml-agent/learned-value-beats-heuristic-augmentation"><img src="https://agentmods.dev/badge/skills/topprismdata/cultivating-ml-agent/learned-value-beats-heuristic-augmentation.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00144 | $0.01657 |
| Opus 5 | $0.00072 | $0.00829 |
| Sonnet 5 | $0.00029 | $0.00331 |
| Haiku 4.5 | $0.00014 | $0.00166 |
Grade A, and why
learned-value-beats-heuristic-augmentation scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 11d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 125 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Learned Value Function + Search Beats Heuristic Augmentation
When a strong hand-crafted heuristic resists every naive ML augmentation (search, behavioral cloning, type-classification), the failure isn't data volume or feature richness — it's the absence of a learned value function. The breakthrough comes from combining (a) a learned value network as a leaf evaluator with (b) sufficient search depth. Neither alone is enough.
This skill captures the hard-won methodology from a multi-week PTCG (Pokémon TCG) AI Battle project where a rank-304 rule-based agent (nursrijan_adv) defeated every naive ML upgrade — until a ReBeL-style value network + 2-ply search finally beat it.
The Anti-Pattern: Strong Baselines Resist Crude Augmentation
A strong heuristic encodes deep domain knowledge (matchup awareness, sequencing, timing). Naive ML augmentation disrupts its internal coherence:
- Search with a hand-crafted eval overrides good heuristic picks with worse ones (the eval is cruder than the heuristic it's trying to improve).
- Behavioral cloning on winners picks the wrong move ~45% of the time, and errors compound across a game. 10× more data doesn't fix this — 437K rows still lost 98%.
- Type-guided policy (predict action TYPE, let heuristic pick within type) breaks the heuristic's coherent sequencing — 6% win rate, worse than random.
Common failure signal: every augmentation approach lands in the same 5-45% range regardless of data volume or feature richness. If you see this, you're not fighting a data problem — you're fighting the coherence problem.
The Breakthrough: V_net + Search Depth, Together
The combination that finally wins:
- Train a value network V(state) → P(win). This is a classification problem (who's-ahead), not an imitation problem. Use existing game logs with outcome labels. Targets: >70% accuracy, well-calibrated win-probabilities.
- Use V_net as the leaf evaluator inside search (MCTS / minimax / forward-rollout). The search provides depth (sees the opponent's response); V_net provides the learned judgment at the leaf.
- Override the heuristic only when V_net is confident (margin ≥ 0.03), so the heuristic's coherence is preserved on uncertain decisions.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 11d ago First seen · 125 lines · 144 tokens per session scan A a7f96f94ee07
learned-value-beats-heuristic-augmentation is a skill published in the GitHub repository topprismdata/cultivating-ml-agent (5 stars, last pushed 14d ago), licensed MIT. It adds 144 tokens to every session and 1,657 once invoked, about $0.0007 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
huggingface-hub
Hugging Face Hub CLI (hf) — search, download, and upload models and datasets, manage repos, query datasets with SQL, deploy inference endpoints, manage Spaces and buckets.
tensorboard
Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit.
mlflow
Track ML experiments, manage model registry with versioning, deploy models to production, and reproduce experiments with MLflow - framework-agnostic ML lifecycle platform.
datachain-knowledge
Use whenever datasets, cloud storage buckets, or data pipelines are mentioned — creating, saving, querying, listing, exploring, deleting, or processing data in S3, GCS, Azure Blob, or local storage. Also use when running any script that may create datasets as a side effect. Maintains a knowledge base at dc-knowledge/…
install-openviking-memory
Install and configure the OpenViking long-term memory plugin for OpenClaw via natural conversation. Once installed, the plugin automatically captures facts from chats and recalls relevant context before each reply (auto-capture + auto-recall, cross-session). Covers prerequisites, install through OpenClaw's plugin…
pufferlib
Version-aware guidance for PufferLib reinforcement-learning environments, vectorization, policies, PuffeRL training, evaluation, and safe checkpoint review. Use when adapting Gymnasium/PettingZoo environments to published PufferLib 3.0.0 or working with the redesigned native 4.0 source line.