EV

evidence-quality-check

A validation tool to mark claims as verified, inferred, or assumed based on actual source data.

Install

mkdir -p .claude/skills/evidence-quality-check && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/15316" && unzip -o skill.zip -d .claude/skills/evidence-quality-check && rm skill.zip

Installs to .claude/skills/evidence-quality-check

Activation

This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.

**Призначення**: перевірити що твердження у відповіді мають реальні докази, не вигадані. **Коли викликати**: при будь-якому RECON / DISCOVERY / review / ADR.
157 chars · catalog descriptionno explicit “when” trigger
Intermediate

Key capabilities

  • Verify claims in responses have real evidence
  • Audit evidence quality in RECON, DISCOVERY, review, or ADR processes
  • Identify fabricated line numbers in code references
  • Detect 'should be' statements disguised as facts
  • Flag phantom imports not verified by grep

How it works

The skill defines canonical markers like [VERIFIED path:line] and [INFERRED] to tag claims. It also outlines forbidden anti-patterns such as fabricated line numbers and 'should be' statements.

Inputs & outputs

You give it
Agent's response with claims about code or system
You get back
Audited response with evidence markers and an evidence quality summary

When to use evidence-quality-check

  • Validating technical documentation
  • Reviewing analysis findings
  • Auditing ADR claims

About this skill

Evidence Quality Check Skill

Призначення: перевірити що твердження у відповіді мають реальні докази, не вигадані. Коли викликати: при будь-якому RECON / DISCOVERY / review / ADR.

Канонічні маркери (SSOT)

МаркерЗначенняПриклад
[VERIFIED path:line]Бачив код, перевірив[VERIFIED core/smc/engine.py:42]
[VERIFIED terminal]Запустив, побачив output[VERIFIED terminal: pytest passed]
[INFERRED]Логічний висновок з даних[INFERRED from naming pattern]
[ASSUMED — verify: <cmd>]Гіпотеза, треба перевірити[ASSUMED — verify: grep "FINAL_SOURCES"]
[UNKNOWN — risk: H/M/L]Сліпа зона[UNKNOWN — risk: M (legacy code path)]
[path:?]Шлях відомий, line — НЕ перевірений[runtime/store/uds.py:?]

Forbidden anti-patterns

A1 — Fabricated line numbers

core/smc/engine.py:42 коли ти не читав файл ✅ core/smc/engine.py:? або [INFERRED]

A2 — "Should be" disguised as fact

❌ "Function X handles error correctly" ✅ "Function X appears to handle error correctly [INFERRED — not VERIFIED with terminal run]"

A3 — Phantom imports

❌ "Module Y imports Z" коли ти не grep'ав ✅ "Module Y likely imports Z [ASSUMED — verify: grep 'import Z' Y.py]"

A4 — Hand-wave conclusions

❌ "Tests cover this case" ✅ "Tests cover this case [VERIFIED tests/test_X.py:N test_case_Y]"

A5 — Stale memory

❌ Цитуєш файл по пам'яті як "I remember it says..." ✅ read_file зараз → [VERIFIED path:line]

Audit checklist

Перед здачею тверджень:

  1. Кожне твердження про код має маркер
  2. [VERIFIED path:line] — line був реально прочитаний цим запитом? (не з summary, не з пам'яті)
  3. [INFERRED] — ясно що це не factual, а логічний висновок
  4. [UNKNOWN] чесно прописані з risk level
  5. Жодного fabricated line number

Червоні прапори у власній відповіді:

  • "Це працює" без [VERIFIED terminal]
  • "У файлі X є функція Y" без [VERIFIED X:line]
  • "Я перевірив" без showing how
  • "Я впевнений" — це слово майже завжди = no evidence

Output format

При повідомленні висновків після RECON:

## FACTS
- [VERIFIED core/smc/engine.py:128] SmcEngine.on_bar() returns None on stale data
- [VERIFIED terminal: pytest tests/test_smc_e1.py] all 12 tests pass
- [INFERRED] config drift у smc.max_compute_ms (default 50ms у code, 30ms у docs)
- [ASSUMED — verify: grep "DERIVE_CHAIN"] DERIVE_CHAIN визначений тільки у core/derive.py

## UNKNOWN
- [UNKNOWN — risk: M] чи D1 anchor offset актуальний для weekend gap (не тестовано)

## EVIDENCE QUALITY: HIGH (3 VERIFIED, 1 INFERRED, 0 fabricated)

Правило використання

  • Будь-який RECON / DISCOVERY = обов'язкове маркування
  • R_REJECTOR: окремо аудитує evidence quality (CA4)
  • R_PATCH_MASTER: проводить evidence-check у RECON gate
  • R_BUG_HUNTER: ніяких висновків без VERIFIED маркерів

When not to use it

  • When the goal is to generate new content without requiring evidence
  • When the task does not involve reviewing or auditing existing claims

Limitations

  • Requires explicit marking of every statement about code
  • Relies on the agent to apply markers correctly
  • Does not generate evidence, only verifies its presence

How it compares

This skill provides a structured, auditable tagging system for claims, unlike a manual review that relies on subjective assessment.

Compared to similar skills

evidence-quality-check side by side with the closest alternatives in the catalog.

SkillInstallsUpdatedSafetyDifficulty
evidence-quality-check (this skill)04moNo flagsIntermediate
citation-management237moReviewBeginner
microsoft-docs105moReviewBeginner
peer-review117moReviewAdvanced

Try saying

Example prompts that trigger this skill in your AI assistant.

You might also like

citation-management

davila7

Comprehensive citation management for academic research. Search Google Scholar and PubMed for papers, extract accurate metadata, validate citations, and generate properly formatted BibTeX entries. This skill should be used when you need to find papers, verify citation information, convert DOIs to BibTeX, or ensure reference accuracy in scientific writing.

2395

microsoft-docs

github

Query official Microsoft documentation to understand concepts, find tutorials, and learn how services work. Use for Azure, .NET, Microsoft 365, Windows, Power Platform, and all Microsoft technologies. Get accurate, current information from learn.microsoft.com and other official Microsoft websites—architecture overviews, quickstarts, configuration guides, limits, and best practices.

1059

peer-review

davila7

Systematic peer review toolkit. Evaluate methodology, statistics, design, reproducibility, ethics, figure integrity, reporting standards, for manuscript and grant review across disciplines.

1146

notion-research-documentation

openai

Research across Notion and synthesize into structured documentation; use when gathering info from multiple Notion sources to produce briefs, comparisons, or reports with citations.

551

context7-auto-research

davila7

Automatically fetch latest library/framework documentation for Claude Code via Context7 API

550

fact-check

leonardomso

Verify technical accuracy of JavaScript concept pages by checking code examples, MDN/ECMAScript compliance, and external resources to prevent misinformation

746

Search skills

Search the agent skills registry