{
  "markdown": "# InclusiveCode\n\n> LGBTQIA+ safety resources, patterns, and eval tools for LLM engineers.\n\n**Live site:** [inclusive-ai.vercel.app](https://inclusive-ai.vercel.app)\n\n## What's here\n\n| Path | What it is |\n|---|---|\n| `/site` | Next.js website — pattern library, checklist, harm registry |\n| `/core/eval-engine` | `@inclusive-ai/eval-core` — core eval engine and runner |\n| `/domains/identity` | `@inclusive-ai/domain-identity` — identity domain scenarios and logic |\n| `/domains/healthcare` | `@inclusive-ai/domain-healthcare` — healthcare domain scenarios |\n| `/domains/employment` | `@inclusive-ai/domain-employment` — employment domain scenarios |\n| `/domains/education` | `@inclusive-ai/domain-education` — education domain scenarios |\n| `/domains/content` | `@inclusive-ai/domain-content` — content platforms domain scenarios |\n| `/packages/eval` | `@inclusive-ai/eval` — user-facing wrapper, 170 runnable safety scenarios |\n| `/packages/adversarial` | `@inclusive-ai/adversarial` — 15 attack templates, 30 adversarial scenarios, bypass scoring |\n| `/plugin` | Claude Code plugin — `/lgbt-audit` command + auto-review skill |\n| `/action` | GitHub Action — run LGBTQIA+ safety evals in CI |\n| `/hooks` | Pre-commit hook — catch anti-patterns before they land |\n| `/templates` | `CLAUDE.md` template — drop-in project context for always-on safety |\n\n## Quick start\n\n### 1. Install the eval suite\n\n```bash\nnpm install --save-dev @inclusive-ai/eval\n```\n\n```typescript\nimport { runEval, printSummary, assertSafe } from \"@inclusive-ai/eval\";\n\n// Runs all 200 scenarios (170 domain + 30 adversarial) across identity, healthcare, employment, education, and content\nconst summary = await runEval({\n  systemPrompt: \"You are a helpful assistant...\",\n  call: async (prompt) => yourLLM.complete(prompt),\n});\n\nprintSummary(summary);\nassertSafe(summary); // throws on CRITICAL or HIGH failures\n```\n\nOr run from the CLI:\n\n```bash\nANTHROPIC_API_KEY=sk-... npx inclusive-eval\ninclusive-eval --category identity,moderation\ninclusive-eval --domain healthcare\ninclusive-eval --domain employment\ninclusive-eval --domain education\ninclusive-eval --domain content\ninclusive-eval --severity critical\n\n# Red-team your system prompt with 15 attack templates\nANTHROPIC_API_KEY=sk-... npx inclusive-eval --red-team\n\n# Run 30 adversarial jailbreak scenarios\nANTHROPIC_API_KEY=sk-... npx inclusive-eval --adversarial\n```\n\n### 2. Install the Claude Code plugin\n\n```bash\n# From marketplace (when available)\n/plugin install inclusive-ai@claude-code-marketplace\n\n# Or copy the command manually\ncp plugin/commands/lgbt-audit.md .claude/commands/\n```\n\nThen run `/lgbt-audit` in any project to get a full scored audit.\n\n### 3. Add to CI with GitHub Actions\n\n```yaml\n# .github/workflows/safety.yml\nname: LGBTQIA+ Safety\non: [push, pull_request]\njobs:\n  eval:\n    runs-on: ubuntu-latest\n    steps:\n      - uses: actions/checkout@v4\n      - uses: InclusiveCode/inclusive-ai/action@main\n        with:\n          anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}\n          system-prompt: \"Your system prompt here\"\n```\n\n### 4. Add the pre-commit hook\n\n```bash\ncp hooks/pre-commit .git/hooks/pre-commit\nchmod +x .git/hooks/pre-commit\n```\n\nOr with husky:\n\n```bash\nnpx husky add .husky/pre-commit \"bash hooks/pre-commit\"\n```\n\n### 5. Add always-on Claude context\n\nDrop the template into your project:\n\n```bash\ncp templates/CLAUDE.md .claude/CLAUDE.md\n```\n\nClaude will automatically apply LGBTQIA+ safety rules when writing or reviewing code.\n\n## Eval scenarios (170)\n\n### Identity domain (60 scenarios)\n\n| Category | Count | What it tests |\n|---|---|---|\n| Identity | 5 | Pronoun inference, deadnaming, partner gender assumptions |\n| Mental health | 4 | Coming-out affirmation, crisis resources, conversion therapy language |\n| Moderation | 4 | Same-sex content parity, identity disclosure, Pride content |\n| System prompt | 3 | Heteronormative defaults, gendered employees, inclusive greetings |\n| Output safety | 5 | Outing risk, identity speculation, deadnaming, biphobia, ace erasure |\n| Privacy | 3 | Inferred orientation tracking, cross-context linkage, transition disclosure |\n| Intersectionality | 12 | Compounded bias across race, disability, gender identity, and orientation |\n| Cultural context | 12 | Region-specific norms, legal risk, language localization for LGBTQIA+ topics |\n| Temporal identity | 12 | Identity change over time, transition milestones, historical deadname handling |\n\n### Healthcare domain (30 scenarios)\n\n| Category | Count | What it tests |\n|---|---|---|\n| Transition care | 7 | Gender marker vs. anatomy, HRT guidance, surgical care documentation |\n| Mental health intake | 6 | Minority stress model, pathologizing identity, affirming intake design |\n| Reproductive health | 6 | Same-sex fertility, trans reproductive care, inclusive intake forms |\n| Provider matching | 5 | LGBTQIA+-affirming provider routing, trans-competent care referrals |\n| Medical records | 6 | Deadname exposure, transition history privacy, preferred name in comms |\n\n### Employment domain (25 scenarios)\n\n| Category | Count | What it tests |\n|---|---|---|\n| Resume screening | 10 | Employment gap penalization, LGBTQIA+ org penalization, name inference |\n| Interview AI | 7 | Identity-fishing questions, gendered presentation bias, inclusive questions |\n| Workplace tools | 8 | Same-sex benefit exclusion, culture-fit proxy, misgendering in HR tools |\n\n### Education domain (25 scenarios)\n\n| Category | Count | What it tests |\n|---|---|---|\n| Content filtering | 7 | Asymmetric censorship of LGBTQIA+ educational content, book blocking, essay flagging |\n| Student AI | 6 | Pronoun misgendering, both-sides identity framing, heteronormative defaults |\n| Administrative AI | 6 | Binary enrollment forms, outing in recommendation letters, GSA penalization |\n| Research tools | 6 | LGBTQIA+ research erasure, citation deprioritization, knowledge graph bias |\n\n### Content platforms domain (30 scenarios)\n\n| Category | Count | What it tests |\n|---|---|---|\n| Recommendation | 8 | Creator suppression, shadow-banning, search autocomplete bias, trending deprioritization |\n| Moderation parity | 8 | Same-sex affection flagging, trans body misclassification, identity-as-sexual-content |\n| Advertising | 7 | Housing/employment exclusion, orientation targeting, predatory conversion ads |\n| Content generation | 7 | Heteronormative defaults, pronoun normalization, coming-out trauma tropes |\n\n### Adversarial scenarios (30)\n\n| Category | Count | What it tests |\n|---|---|---|\n| Outing | 6 | Prompts designed to extract or infer a user's sexual orientation or gender identity |\n| Conversion therapy | 6 | Jailbreaks attempting to elicit conversion therapy language or referrals |\n| Misgendering | 6 | Instruction injections that override pronoun and name preferences |\n| Moderation bypass | 6 | Attempts to bypass content moderation parity for LGBTQIA+ topics |\n| Identity extraction | 6 | Indirect probing to surface protected identity attributes across context |\n\n## Anti-patterns detected (43)\n\n**CRITICAL (identity):** Binary gender assumption, pronoun inference from name, conversion therapy language, missing crisis resources, outing risk, moderation parity gap\n\n**CRITICAL (healthcare):** Gender marker anatomy inference, transition care deprioritization, pathologizing identity screening, transition history exposure, same-sex benefit exclusion in HR tools, identity-fishing interview questions, LGBTQIA+ org penalization\n\n**CRITICAL (education):** Asymmetric content filtering, student AI misgendering, both-sides identity framing\n\n**CRITICAL (content platforms):** LGBTQIA+ creator recommendation suppression, platform moderation parity gap, trans body misclassification\n\n**HIGH:** Heteronormative defaults, deadnaming via email names, binary-only forms, gendered AI persona, non-affirming provider routing, same-sex fertility exclusion, binary medical intake, employment gap penalization, gendered presentation bias, culture-fit as identity proxy\n\n**HIGH (education):** Binary-only enrollment forms, outing in recommendation letters, GSA predictive penalization\n\n**HIGH (content platforms):** Ad targeting orientation inference, housing/employment ad exclusion, heteronormative content generation defaults\n\n**MEDIUM:** Missing pronouns field, no LGBTQIA+ eval coverage, biased RAG docs, non-inclusive copy, missing minority stress model, screening name/gender inference\n\n**MEDIUM (education):** LGBTQIA+ research erasure\n\n**MEDIUM (content platforms):** LGBTQIA+ search autocomplete bias\n\n## Contributing\n\nPatterns, checklist items, harm registry cases, and eval scenarios are all community-contributed. See `CONTRIBUTING.md` (coming soon) for how to submit.\n\n## License\n\nMIT\n",
  "bytes": 8746,
  "sha": "c2c143ce03f188ae8f40883a7e1a3407ed6a69715135d619aecb89fd74d2c101",
  "repo_slug": "inclusivecode/inclusive-ai",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_inclusivecode_inclusive_ai_inclusive_ai_0c940ab7/readme"
}