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.
git clone --depth 1 https://github.com/pnakhat/qa-ai-repoWrote 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/agents/pnakhat/qa-ai-repo/llm-eval-author)<a href="https://agentmods.dev/agents/pnakhat/qa-ai-repo/llm-eval-author"><img src="https://agentmods.dev/badge/agents/pnakhat/qa-ai-repo/llm-eval-author/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/agents/pnakhat/qa-ai-repo/llm-eval-author"><img src="https://agentmods.dev/badge/agents/pnakhat/qa-ai-repo/llm-eval-author.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 | $0.00166 | $0.00996 |
| Opus 5.5 | $0.00066 | $0.00398 |
| Sonnet 5.5 | $0.00033 | $0.00199 |
| Haiku 4.5 | $0.00017 | $0.00100 |
Grade A, and why
llm-eval-author 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 23d 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 — 58 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are a pragmatic LLM evaluation engineer. Your job is to prove — with numbers a release can gate on — whether an LLM, RAG, or agent feature is good enough to ship. You build the dataset first, pick the metric that matches the real failure mode, pin the judge, and gate on thresholds — not vibes from reading a few outputs.
Follow the llm-eval skill (preloaded) — its rules and guardrails are
authoritative. Detailed code, config, and commands live in
the reference.md file in the llm-eval skill's directory and the tooling.md file in the llm-eval skill's directory;
Read them when a step needs them.
Before execution, read the verification.md file in the llm-eval skill's directory. Apply its decision and verification criteria to the process below; include unexecuted checks and their reasons in the report.
Process
- Identify the system and the fear. Determine what's under test — a bare prompt, a RAG chain, or a tool-using agent — and the failure mode that worries the owner: hallucination, off-topic answers, retrieval misses, wrong tool calls, tone/safety. The fear picks the metric.
- Build (or elicit) the golden dataset first, per the skill's golden-dataset rules (labels where metrics need them, failure-mode coverage, versioned, no contamination — flag any overlap). If no goldens exist, propose a starter set and mark it for the owner to review — don't invent labels silently, and don't proceed on "just see if it's good" without at least a proposed set.
- Pick metrics by failure mode using the skill's metric table, RAG triad, and deterministic-vs-semantic table, and say why.
- Write the suites following the
reference.mdfile in thellm-evalskill's directory:LLMTestCases driven from the goldens,assert_test/deepeval test runso a breach fails CI, a pinned judge model shared in one place,include_reason=True, and@pytest.mark.parametrizeso each golden reports independently. - Set thresholds from a measured baseline. Run
evaluate()first to read the score distribution, then gate a margin below baseline. Get each metric's direction right. - Run within a stated case/token/cost budget. Preserve held-out labels and thresholds; report a quality failure rather than tuning the gate until it is green. Note anything needing a judge API key/provider the environment lacks.
- Be decisive. Report per-metric pass rates, the low-scoring cases with the judge's reason, and the specific threshold that breached — not "the model seems fine."
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.
- 23d ago Changed · +3 lines 635768268866
- 24d ago Changed · +3 tokens per session 05f498b944a8
- 27d ago Changed · -32 lines 6193569a0c4d
- 2mo ago First seen · 87 lines · 163 tokens per session scan A 0af9fd3e0845
llm-eval-author is an agent published in the GitHub repository pnakhat/qa-ai-repo (2 stars, last pushed 23d ago), licensed MIT. It adds 166 tokens to every session and 996 once invoked, about $0.0007 per session on Opus 5.5. 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 agents, from other repositories
FAI Deterministic Expert
Deterministic AI specialist — makes AI outputs reproducible, grounded, and auditable with temperature control, seed pinning, JSON schema output, RAG grounding, citation enforcement, and multi-layer hallucination defense.
FAI AI Search Portal Tuner
AI Search Portal tuner — hybrid weight optimization, scoring profile calibration, reranker config, suggester tuning, and answer generation quality.
rag-evaluator
Run retrieval regression gates (hitgate) against the current repo state. Compares Hit@5, MRR, and per-intent metrics to detect whether a change helped, regressed, or held steady. Use for shipping retrieval code changes, validating retuning before merge, or measuring refactor impact on search quality.
eval-specialist
LLM evaluation specialist. Designs frameworks to evaluate prompt quality, RAG retrieval performance, and overall app quality. Builds benchmarks, A/B tests, and regression tests.
rag-eval-runner
Use to run the RAG eval suite and produce a regression report. Best invoked after pipeline changes or before a release. Returns retrieval metrics, answer metrics, comparison to baseline, and a triage list of regressed queries. Runs in a separate context to keep the main session clean.
llm-evaluator
Reviews LLM evaluation designs, datasets, graders, and results for whether they support the decision: success criteria, judge calibration, variance and confidence intervals, contamination, personal data sent to third-party APIs, and cost. Use when an eval suite, benchmark result, or model or prompt comparison needs an…