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 pktikkani/agent-skills --skill eval-disciplinegit clone --depth 1 https://github.com/pktikkani/agent-skillsWrote 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/pktikkani/agent-skills/eval-discipline)<a href="https://agentmods.dev/skills/pktikkani/agent-skills/eval-discipline"><img src="https://agentmods.dev/badge/skills/pktikkani/agent-skills/eval-discipline/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/pktikkani/agent-skills/eval-discipline"><img src="https://agentmods.dev/badge/skills/pktikkani/agent-skills/eval-discipline.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 | $0.00066 | $0.00910 |
| Opus 5 | $0.00033 | $0.00455 |
| Sonnet 5 | $0.00013 | $0.00182 |
| Haiku 4.5 | $0.00007 | $0.00091 |
Grade A, and why
eval-discipline 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 2d 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 — 71 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Eval discipline
How to structure the evals themselves. This complements the evals-skills pack (hamel's) — that pack covers the workflows (error-analysis, judge validation, synthetic data); this skill governs the shape of the evaluators and the report.
Directives
- The KPI list comes from
PRODUCT.mdwhen it exists. Its Failure modes table is the evaluator list — one Evaluator per row,nameequal to the row's post-ship signal name so a production drop maps straight to the eval to rerun. Thresholds come from Pre-ship evidence. A column with no row in PRODUCT.md is ornamental until the human adds the row. - KPI-based, never generic. Every evaluator traces to a real failure mode (found via error analysis) or a business KPI. Ten meaningful columns beat thirty ornamental ones. If you can't name the failure mode or KPI a column defends, don't add it.
- One small Evaluator per KPI — never a mega-scorer. Each is a tiny class with a
uniform minimal interface (
name,evaluate(trace, ctx) -> Score). The harness composes them; adding, deleting, or re-tuning one never touches another. - Report the vector, never a blend. Every KPI is its own column in the report; regressions are tracked per column and gated with independent thresholds. NEVER collapse into one blended quality score for decisions — a mean hides the regression that matters.
- Shared preprocessing computed once. Parse/normalize the trace one time into a
context object (
ctx) passed to all evaluators. Never re-parse the trace N times. - Right tool per KPI. Use a deterministic code-check wherever the KPI is checkable (schema, latency, exact rules, presence/format). Reserve an LLM judge for KPIs where interpretation is unavoidable (tone, faithfulness, relevance).
- Every judge is calibrated before it's trusted. An LLM judge earns its column only after being validated against human labels (TPR/TNR) — see the evals-skills validate-evaluator skill for the how. An uncalibrated judge is a guess with a number on it.
- Deletion and tightening are surgery-free. Removing a dead metric drops one class and one column. Tightening one gate changes one threshold. If either forces you to re-tune others, the evaluators are entangled — split them.
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.
- 2d ago Changed · +4 lines 807c07f4658f
- 9d ago First seen · 67 lines · 66 tokens per session scan A c0e0113cdfde
eval-discipline is a skill published in the GitHub repository pktikkani/agent-skills (2 stars, last pushed 2d ago), licensed MIT. It adds 66 tokens to every session and 910 once invoked, about $0.0003 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
cli-eval
Create and run evaluation suites, watch live benchmark progress, view scorecards, compare model performance, and integrate eval runs with CI workflows from the CLI.
model-merging
Merge multiple fine-tuned models using mergekit to combine capabilities without retraining. Use when creating specialized models by blending domain-specific expertise (math + coding + chat), improving performance beyond single models, or experimenting rapidly with model variants. Covers SLERP, TIES-Merging, DARE, Task…
darwinian-evolver
Evolve prompts/regex/SQL/code with Imbue's evolution loop.
validate
Validate Semantica pipelines, extraction quality, graph schemas, and ontology consistency. Returns structured error/warning checklists. Uses PipelineValidator, PipelineBuilder.validatepipeline(), GraphValidator, and OntologyValidator. Sub-commands: pipeline, step, dependencies, extraction, graph, ontology, performance.
launching-evals
Run, monitor, analyze, and debug LLM evaluations via nemo-evaluator-launcher. Covers running evaluations, checking status and live progress, debugging failed runs, exporting artifacts and logs, and analyzing results. ALWAYS triggers on mentions of running evaluations, checking progress, debugging failed evals…
nemo-automodel-recipe-development
Create and modify NeMo AutoModel training and evaluation recipes, including YAML structure, builders, and execution flow.