# Benchmark Fp Fn Audit

> Audita React Doctor contra corpus de benchmark como ReactBench: falsos positivos y negativos confirmados, huecos de taxonomía y artefactos del verificador. Se usa al analizar logs y artefactos de trial o al pedir un segundo pase.

Fuente: https://skillsagentes.com/skills/millionco/react-doctor/benchmark-fp-fn-audit
Markdown: https://skillsagentes.com/skills/millionco/react-doctor/benchmark-fp-fn-audit.md

Repositorio: https://github.com/millionco/react-doctor
Autor: millionco
Licencia: NOASSERTION
Actualizado: el mes pasado
Coste de contexto: 94 tok instalada, 2k tok al activarse, 2.7k tok con todos los archivos del bundle
Bundle: 3 archivos, 10 KB
Permisos que pide: ninguno declarado

## Instalación

Un skill son archivos markdown: los mismos archivos valen para cualquier agente y lo único que cambia es el directorio de destino, es decir la bandera `--agent`. Añade `-g` para instalarlo en todos los proyectos de la máquina.

```bash
# Claude Code
npx -y skills add millionco/react-doctor --skill benchmark-fp-fn-audit --agent claude-code
# Cursor
npx -y skills add millionco/react-doctor --skill benchmark-fp-fn-audit --agent cursor
# Codex
npx -y skills add millionco/react-doctor --skill benchmark-fp-fn-audit --agent codex
# Gemini CLI
npx -y skills add millionco/react-doctor --skill benchmark-fp-fn-audit --agent gemini
# Windsurf
npx -y skills add millionco/react-doctor --skill benchmark-fp-fn-audit --agent windsurf
# Cline
npx -y skills add millionco/react-doctor --skill benchmark-fp-fn-audit --agent cline
```

## Qué hace

- Audita React Doctor contra corpus de benchmark como ReactBench en busca de falsos positivos y falsos negativos confirmados, huecos de taxonomía y artefactos del verificador.
- Obliga a leer la documentación completa de reglas, cuantificar la distribución de fallos e inspeccionar cada artefacto de trial, sin tratar revisiones previas como verdad.
- Clasifica de forma conservadora: FP confirmado, FN confirmado, hueco de taxonomía, artefacto del harness o candidato, y cada hallazgo confirmado cita trial, archivo, línea, regla y resultado de test.
- Incluye un segundo pase adversarial independiente con heurísticas para distinguir estado derivado de render de estado asíncrono, sincronización de padres, valores HTML peligrosos, etc.
- Escribe salidas de solo auditoría (`.md`, `.jsonl`, `.tsv`) bajo el directorio del benchmark y nunca toca el código de React Doctor, del benchmark ni de los tests.

## Cuándo usarla

- Se analizan `rd.log`, `rd-before.json`, `rd-after.json`, `model.patch`, `result.json` o logs de reward y tests.
- Se pide un segundo pase adversarial sobre los hallazgos de un benchmark de React Doctor.

## Qué la activa

- "Audita los resultados de ReactBench para falsos positivos y negativos"
- "Haz un segundo pase adversarial sobre estos trials de benchmark"
- "Cuantifica la distribución de fallos por regla en este corpus"

## Antes de instalar

- Necesita acceso al directorio del corpus de benchmark y a la documentación de reglas en `react.doctor/docs/rules`; usa `rg --files` para enumerar artefactos.
- makes network requests

## Archivos

- SKILL.md — 8 KB
- agents/openai.yaml — 246 B
- references/benchmark-artifacts.md — 2 KB

## SKILL.md

Reproducido tal cual desde millionco/react-doctor bajo NOASSERTION. Esta sección es el documento original y está en inglés.

# React Doctor benchmark FP/FN audit

Perform an evidence-backed audit of React Doctor diagnostics across a benchmark corpus. Read the complete rule documentation, quantify the distribution of failures, inspect every relevant trial artifact, and independently perform a second pass for additional false positives and false negatives.

## Corpus and required resources

For this benchmark, use:

```text
/home/aidenybai/Developer/react-bench-internal/jobs/ReactBench-v1.1-beta-rd-0.9.3-consolidated-final-20260801
```

Read the complete React Doctor rule documentation from:

```text
https://www.react.doctor/docs/rules
```

Read these repository review materials when they exist:

```text
/home/aidenybai/Developer/react-bench-internal/docs/aiden-review
```

Do not treat prior reviews or reports as ground truth. Use them as leads and reverify every claim against the current benchmark revision.

For every inspected trial, locate and read:

- `verifier/model.patch`
- `verifier/rd-before.json`
- `verifier/rd-after.json`
- `verifier/rd.log`
- `result.json`
- reward and test logs, discovered with `rg --files`
- the task README, prompt, or review material
- the base and patched source files referenced by the patch

Inspect all trial directories, including primary and verified variants. Normalize duplicate task names and repeated patch hashes before computing task-level statistics.

## Evidence rules

Classify findings conservatively:

- **Confirmed FP**: a diagnostic is caused by changed code, the behavior is intentional or required, and the rule contract does not apply or its recommended fix would regress behavior.
- **Confirmed FN**: a real defect exists in the base or final code, no applicable diagnostic is emitted, and an existing documented rule should cover it.
- **Taxonomy gap**: a real defect is present but no existing rule contract reasonably claims to cover it.
- **Harness artifact**: the result is caused by malformed output, stale baseline data, verifier behavior, generated files, or an untouched-file diagnostic.
- **Candidate**: plausible but missing enough evidence for confirmation.

Do not call a target-task miss a false negative. Do not call a diagnostic a false positive merely because the model failed tests. `react_doctor=1` is a gate result, not proof that the code is correct.

Every confirmed finding must cite the exact trial, file, line or code span, rule documentation, diagnostic delta, task behavior, and test result. Separate strict rule misses from out-of-taxonomy gaps.

## Workflow

### 1. Inventory the corpus

Use `rg --files` to enumerate trial artifacts. Record the benchmark revision, trial count, missing artifacts, duplicate variants, malformed reports, and available task reviews. Do not silently skip incomplete trials.

### 2. Build diagnostic distributions

Recompute distributions from raw `rd.log`, `rd-before.json`, and `rd-after.json` data. Rank by trial coverage before raw occurrence count. Report:

- total trials and artifact completeness;
- reward, test, and React Doctor gate tuple counts;
- direct `NEW` or introduced-diagnostic failures;
- health-mode `NEW` deltas;
- health-mode `STILL` target misses;
- baseline-to-head rule-count deltas;
- rule, category, severity, task, repository, and file concentration;
- changed-file versus untouched-file diagnostics;
- malformed or truncated reports;
- primary versus verified-task duplication;
- successful versus unsuccessful patch comparisons.

Always recompute before relying on a prior summary. Large files can inflate occurrence counts, so show both occurrence count and affected-trial count.

### 3. Inspect high-impact clusters

Prioritize high-coverage rules and exact introduced-diagnostic failures, especially:

```text
exhaustive-deps
no-giant-component
js-set-map-lookups
no-array-index-as-key
js-combine-iterations
effect-needs-cleanup
no-adjust-state-on-prop-change
button-has-type
prefer-module-scope-static-value
no-static-element-interactions
prefer-module-scope-pure-function
no-pass-data-to-parent
no-pass-live-state-to-parent
no-derived-state
no-ref-current-in-render
click-events-have-key-events
rerender-lazy-ref-init
only-export-components
rerender-lazy-state-init
```

For each high-impact cluster, inspect representative positives, representative negatives, the largest task concentration, and sibling trials with different outcomes.

### 4. Perform the independent second pass

Search every trial for additional candidates, not just the known clusters. Apply these heuristics:

- Separate expensive lazy initialization from cheap empty `Map`, `Set`, and equivalent containers.
- Distinguish render-derived state from async status, media errors, retries, resource lifetimes, subscriptions, and transient UI state.
- For controlled components, distinguish required parent synchronization from effect loops or unnecessary data forwarding.
- Follow dangerous HTML values through local aliases, `useMemo`, wrappers, and helper functions; preserve provenance for trusted serializers such as KaTeX.
- Distinguish React-docs-blessed previous-prop comparisons from pure derived state and render-phase side effects.
- Inspect timers, promises, animations, abort signals, subscriptions, and stale closures for races that generic dependency or cleanup warnings do not directly explain.
- Check focus stealing, focus restoration, controlled selection, keyboard semantics, and ARIA transitions.
- Check thresholds, aliasing, destructuring, nested callbacks, JSX wrappers, computed properties, and TypeScript syntax.
- Verify that a diagnostic is on changed code and not a pre-existing issue, generated artifact, fixture, or untouched neighbor.
- For every proposed FP, find a nearby true-positive counterexample. For every proposed FN, find a nearby case the detector catches.

Known leads to independently recheck, without presuming their classification:

```text
fix-react-rdh-nteract-semiotic-a__5uDDmgx
write-react-glific-glific-fronte__22Gg4p2
fix-react-jumperexchange-jumper__2GBTh7Z
write-react-frankchen021-datasto__2N6xFwQ
fix-react-viclafouch-mui-tel-inp__hQXzJgB
fix-react-floating-ui-floating-u__27iKG4t
fix-react-formidablelabs-victory__2uwi2vz
fix-react-igordanchenko-yet-anot__2GZGZYT
fix-react-rdh-appflowy-io-appflo__vK9uEUy
```

### 5. Produce audit artifacts

Write audit-only outputs under the benchmark directory:

```text
audit/rd-0.9.3-second-pass.md
audit/rd-0.9.3-second-pass.jsonl
audit/rd-0.9.3-rule-distribution.tsv
```

Do not edit React Doctor source, benchmark source, task tests, or verifier data. The JSONL must contain one record per finding with:

```json
{
  "id": "...",
  "classification": "confirmed_fp | confirmed_fn | taxonomy_gap | harness_artifact | candidate",
  "confidence": "high | medium | low",
  "rule": "...",
  "task": "...",
  "trial": "...",
  "file": "...",
  "lines": "...",
  "evidence": "...",
  "baselineDiagnostics": "...",
  "headDiagnostics": "...",
  "testStatus": "...",
  "rdStatus": "...",
  "rationale": "...",
  "recommendedAction": "..."
}
```

The Markdown report must contain: corpus/methodology, distributions, confirmed FPs, confirmed FNs, taxonomy gaps, harness artifacts, manual-adjudication candidates, prioritized rule fixes, and limitations.

Prioritize fixes using:

```text
affected trial coverage × confidence × reproducibility
```

Never make a code change as part of this skill unless the user separately requests implementation of a confirmed rule fix.

## Dónde encaja

- Categoría: [Testing y QA](https://skillsagentes.com/categorias/testing-qa.md) — Flujos de testing unitario, de integración y end-to-end.
- Creador: [millionco](https://skillsagentes.com/creators/millionco.md) — 16 skills en el directorio
- [Todas las skills](https://skillsagentes.com/skills.md)
- [Ranking de instalaciones](https://skillsagentes.com/ranking.md)

## Otras skills del mismo repositorio

- [React Doctor](https://skillsagentes.com/skills/millionco/react-doctor/react-doctor.md): Escanea bases de código React en busca de problemas de seguridad, rendimiento, corrección y arquitectura, y da una puntuación de salud de 0 a 100. Incluye chequeo de regresión y un flujo completo de triage local con `/doctor`.
- [Run Parity](https://skillsagentes.com/skills/millionco/react-doctor/run-parity.md): Compara los diagnósticos de React Doctor de un pull request contra su base con Daytona. Se usa para correr parity, comprobar regresiones de diagnósticos en un PR o reportar diagnósticos añadidos y quitados.
- [Improve Threejs](https://skillsagentes.com/skills/millionco/react-doctor/improve-threejs.md): Audita y arregla apps de Three.js y React Three Fiber: rendimiento del frame-loop, fugas de memoria de GPU, corrección del grafo de escena y defectos visuales como z-fighting, shadow acne, espacio de color erróneo y resize roto.
- [Performance](https://skillsagentes.com/skills/millionco/react-doctor/performance.md): Diagnostica el rendimiento de React en tiempo de ejecución con trazas de React Doctor, marcado de renders en vivo, Long Animation Frames y evidencia de renders por componente. Se invoca como `/performance`.
- [Fuzz](https://skillsagentes.com/skills/millionco/react-doctor/fuzz.md): Somete a fuzzing las reglas de React Doctor con `@react-doctor/fuzz` para hallar crashes, lentitud, falsos positivos y diagnósticos sensibles a mutaciones. Se usa tras pasar los tests de la regla o al confirmarse un falso positivo nuevo.

---

Skills Agentes · [Índice de páginas en markdown](https://skillsagentes.com/sitemap.md) · [Inicio](https://skillsagentes.com/index.md)