# Sharing Policy

This org is accessible to external collaborators — external coaches, writers, reviewers. Treat it as a semi-public space. Before committing anything here, check the lists below.

## ✅ Safe to commit

- Article prose (drafts, iterations, finals)
- Tool output values for specific hands referenced — solver action frequencies, equity, ranges, blockers per hand. These per-hand results are not IP.
- Reviewer feedback files
- Editorial decisions
- Claude transcripts where ideation happened (ad-hoc track)
- Image references — URLs, captions, alt text
- Article structure files, outlines, briefs
- Hygiene check status (the canonical numbers are already public at internals.quintace.ai/solver-qa/#hygiene)
- Verifier outcomes per claim (PASS / FAIL / AMBIG)

## ❌ Never commit

- DA pipeline prompt internals (CAI track — decomposer, FactGraph, ClaimManifest, verifier-first work)
- System prompts for the DA pipeline / quint-ai surface (ad-hoc publishing pipeline prompts in `ad-hoc/tools/generate_tier2.py` etc. are **carved out** — see below)
- Knowledge layer modules (CAI track)
- MCP server code, agent implementations, any `quint-ai` repo content
- Solver / RLServ endpoint URLs, model names (`moe_dynamic_*`), training details
- Decomposer / FactGraph / ClaimManifest schemas (Chang's verifier-first work)
- DA pipeline orchestration code, deployment scripts, GitHub Actions internals
- Partner-specific data (WPT / Gold / Global stats, partner-tagged player data, anything partner-confidential)
- Coach roster, comp, contract terms
- Unpublished competitive intel, fundraising material, investor decks
- Anything from `engineering-department/`, `hr-department/`, `company/` in the internal repos
- Raw research from `engineering-department/gameplay-ai/projects/llm-verifier-game-expansion/` — the upstream of the ad-hoc publishing pipeline. Books rendered from it ship here; the source `.md` files stay internal.

## 🟢 Carve-out — ad-hoc publishing tools (committed deliberately)

The 82 publishing tools in `ad-hoc/tools/` and their embedded prompts (in `generate_tier2.py`, `review_article.py`, `review_article_editorial.py`) are committed publicly **on purpose** (Thanh 2026-05-24, ad-hoc retirement decision):

- External collaborators studying our publishing methodology benefit from seeing the actual pipeline that produced the rendered books and articles.
- The tools alone are **inert** without (a) cross-workspace access to `engineering-department/.../llm-verifier-game-expansion/` raw research, and (b) an Anthropic API key. So shipping the source does not let an outsider rebuild books on their own.
- The system prompts in these tools are **voice + format guidance** for publishing — not the deeper DA-pipeline / verifier-first IP that stays in the CAI track.

The contract: tools-as-build-system internally; tools-as-documentation publicly. See `ad-hoc/tools/REBUILD.md` for the operator guide.

## ⚠️ Gray zone — be deliberate

- `tool_trace.json` — share the resolved data for the specific hand; do NOT include pipeline metadata, agent runtime info, or anything beyond per-hand solver/equity/range output
- `claim_manifest.json` — share per-claim PASS/FAIL outcomes if needed; do NOT share the decomposer schema itself
- Per-article methodology notes — if your notes describe HOW the pipeline was invoked (which prompts, which tools, in what order), keep those in the internal repo. The article itself is public-bound; the methodology used to make it may not be.

## When unsure

Default to "this stays in the internal repo." If you're not certain something is safe to share here, ask in `#cai_general` (internal Slack) or DM @Thanh / @Harold before pushing.

## Pre-push audit

Before opening a PR, grep your changed files for:
- Anything in the never-commit list above
- Email addresses other than your own
- API keys, tokens, secrets
- Internal URLs (e.g. `*.aceguardianrl.com`, internal staging URLs, prompt repo paths)

If anything looks borderline, don't push. Ask first.

## What if I accidentally commit something I shouldn't?

Tell Thanh immediately. The fix is:
1. Force-push a clean version (rewriting history)
2. Rotate any secrets that may have been exposed
3. Note the incident so we tighten the audit