calibration-guard

calibration-guard is a skill for Claude Code from ByteStack-Labs/claude-plugins. It costs 208 tokens per session (1,810 once invoked), scanned A, original, MIT.

A diagnostic for measuring whether an AI model, classifier, agent, or automated judge's confidence matches its correctness. It compares this relationship on evaluation data and on production-like data after the inputs change.

In plain words
What is it for?
Use it to measure calibration, reliability diagrams, expected calibration error, high-confidence error rates, and rules for catching confident-wrong predictions.
Why use it?
A confidently wrong result can pass unnoticed because nothing flags it. The diagnostic quantifies that risk and shows how it changes outside the test setting.

Skill for Claude Code

Written for Claude Code: shipped in a Claude Code plugin.

Part of the agent-reliability plugin — 4 skills shipped together

Good fit Use it to measure calibration, reliability diagrams, expected calibration error, high-confidence error rates, and rules for catching confident-wrong predictions.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/bytestack-labs/claude-plugins/calibration-guard
Install

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.

Any agent
npx skills add ByteStack-Labs/claude-plugins --skill calibration-guard
Clone the repo
git clone --depth 1 https://github.com/ByteStack-Labs/claude-plugins

Made for: Claude Code.

Or install agent-reliability, the plugin that ships this one along with the rest of its 4 skills.

Wrote 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.

agentmods badge for calibration-guard

README.md
[![agentmods](https://agentmods.dev/badge/skills/bytestack-labs/claude-plugins/calibration-guard/github.svg)](https://agentmods.dev/skills/bytestack-labs/claude-plugins/calibration-guard)
Your own site
<a href="https://agentmods.dev/skills/bytestack-labs/claude-plugins/calibration-guard"><img src="https://agentmods.dev/badge/skills/bytestack-labs/claude-plugins/calibration-guard/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.

agentmods 80×15 button for calibration-guard

Your own site · 80×15
<a href="https://agentmods.dev/skills/bytestack-labs/claude-plugins/calibration-guard"><img src="https://agentmods.dev/badge/skills/bytestack-labs/claude-plugins/calibration-guard.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 208 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,810 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5.1 $0.00208 $0.01810
Opus 5 $0.00104 $0.00905
Sonnet 5 $0.00042 $0.00362
Haiku 4.5 $0.00021 $0.00181

Measured 8d ago against content hash bc78a99405ff, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-08, from the pricing page.

Security

Grade A, and why

calibration-guard 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 8d 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.

agent-reliability/skills/calibration-guard/SKILL.md · 153 lines

How it starts

The opening of the file, as written. The whole thing — 153 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Calibration Guard

A system that fails loudly is recoverable: something downstream sees the error and reacts. A system that is confidently wrong is dangerous, because nothing flags it. The prediction ships, the agent acts, the report reassures, and the failure stays silent until it is expensive.

This skill measures one thing precisely: whether the system's confidence tracks its correctness, and how that relationship breaks when the input distribution moves from evaluation to production. It does not improve calibration. It proves the state of it, with numbers anyone can rerun, and hands back a rule that catches the confident-wrong cases before they ship.

How to run it

Run calibration-guard on this system. Pick a confidence signal and a correctness signal, measure calibration on the eval distribution and again on production-realistic inputs, and report the reliability diagram, ECE, and the high-confidence error rate. Show how the gap between confidence on correct and confidence on incorrect predictions changes under the shift. Then give me a selective-prediction rule that catches the confident-wrong cases, with its precision and its coverage cost. Verify every number with runnable code.

Core commitments

  • Verified numbers only. Every figure traces to a measurement produced by runnable code. No remembered or approximate values.
  • Measure the coupling, not the accuracy. Accuracy is the number that hid this failure. The object of study is the relationship between confidence and correctness, not either alone.
  • Calibration is a property of a distribution. Always measure it on both the evaluation distribution and the production distribution. A single measurement tells you nothing about what happens when inputs move.
  • One calibration number hides per-slice miscalibration. Decompose by slice before concluding.
  • Reproducibility is part of the deliverable. Seed everything. Report the compute envelope.

The procedure

Step 0 - Define the confidence and correctness signals

Read the full file on GitHub · 153 lines

Changes

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.

  1. 8d ago First seen · 153 lines · 208 tokens per session scan A bc78a99405ff

Subscribe to this mod's changes

calibration-guard is a skill published in the GitHub repository ByteStack-Labs/claude-plugins (2 stars, last pushed 2mo ago), licensed MIT. It adds 208 tokens to every session and 1,810 once invoked, about $0.0010 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.