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 impactbrussels/AINativeOS --skill eval-and-safety-harnessgit clone --depth 1 https://github.com/impactbrussels/AINativeOSWrote 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/impactbrussels/ainativeos/eval-and-safety-harness)<a href="https://agentmods.dev/skills/impactbrussels/ainativeos/eval-and-safety-harness"><img src="https://agentmods.dev/badge/skills/impactbrussels/ainativeos/eval-and-safety-harness/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/impactbrussels/ainativeos/eval-and-safety-harness"><img src="https://agentmods.dev/badge/skills/impactbrussels/ainativeos/eval-and-safety-harness.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 | $0.00176 | $0.01458 |
| Opus 5 | $0.00088 | $0.00729 |
| Sonnet 5 | $0.00035 | $0.00292 |
| Haiku 4.5 | $0.00018 | $0.00146 |
Grade A, and why
eval-and-safety-harness 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 10d 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 — 69 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Eval and Safety Harness
A demo is not evidence. You picked the input, the moment, the clean case, and the model performed because you stacked the deck. In hard mode that demo is a liability with a clean interface: the wrong answer does not stay on the screen, it walks out into a person. Trust here is engineered before you earn the right to act on a human, never assumed after. This skill builds the harness that turns generation into something a tired clinician can sign their name under.
The method
Take one real feature, the one where being wrong costs the most, and build these five layers around
it. Full method, how to design the eval set, the release gate, a worked harness, anti-patterns and a
copyable template: references/eval-harness-method.md.
-
Write the safety rubric. Turn "is it safe" into pass/fail checks a machine can enforce: every claim cites a real, resolvable source; every number traces to an input; uncertain claims are flagged uncertain; confidence below a threshold refuses rather than guesses. Map each check to a pillar of the six-pillar trust substrate. If you cannot write five sharp binary checks, your safety bar is still a vibe.
-
Pair a generator with a separate judge. One agent drafts. A second, running fresh with its own clean context and the written rubric, scores that draft before anything reaches a human. An author cannot proofread their own typo, and a model defending the context that produced its answer is the same blind spot amplified. Everything turns on the rubric: "is this good?" passes everything; the sharp check catches the hallucinated citation every run.
-
Set the human-on-the-loop boundary. Draw the line: what agents execute unattended, and what a human must review or approve. Put human-in-the-loop on the irreversible (a dose, a large trade, a product released to market). Put human-on-the-loop on everything else, where the human reviews the exceptions queue, not every output. Too low and you have built an expensive autocomplete; too high and a bad outcome reaches a person before anyone notices.
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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.
- 10d ago First seen · 69 lines · 176 tokens per session scan A f43d73c50b3f
eval-and-safety-harness is a skill published in the GitHub repository impactbrussels/AINativeOS (1 stars, last pushed 2mo ago), licensed Apache-2.0. It adds 176 tokens to every session and 1,458 once invoked, about $0.0009 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
imaging-data-commons
Query and download public cancer imaging data from NCI Imaging Data Commons using idc-index. Use for accessing large-scale radiology (CT, MR, PET) and pathology datasets for AI training or research. No authentication required. Query by metadata, visualize in browser, check licenses.
LQF_Machine_Learning_Expert_Guide
LQF Machine Learning Expert Guide - Routed skill for ML/Statistical Modeling with Critical Discussion Mode. Triggers on: machine learning, modeling, prediction, training, classification, regression, clustering, deep learning, neural network, model evaluation, feature engineering, hyperparameter tuning, overfitting…
scientific-data-preprocessing
⚠️ CRITICAL USER EXPERIENCE-BASED SKILL - ALWAYS CONSULT BEFORE DATA PREPROCESSING ⚠️ Prevents catastrophic errors (88.9% error rate in V1.0 case study) through multi-level feature analysis, data leakage detection, and semantic validation. MANDATORY for: data preprocessing, feature engineering, standardization…
embedding-strategies
Select and optimize embedding models for semantic search and RAG applications. Use when choosing embedding models, implementing chunking strategies, or optimizing embedding quality for specific domains.
perplexity-search
Perform AI-powered web searches with real-time information using Perplexity models via LiteLLM and OpenRouter. This skill should be used when conducting web searches for current information, finding recent scientific literature, getting grounded answers with source citations, or accessing information beyond the model…
similarity-search-patterns
Implement efficient similarity search with vector databases. Use when building semantic search, implementing nearest neighbor queries, or optimizing retrieval performance.