# Research Desk Synthesis — RD-2026-09-10-001

**Runtime label: `executor-synthesis` (NOT ChatGPT — no ChatGPT session used).**  
Timestamps: 2026-09-10T11:29:20.107Z / 2026-09-10T20:29:20.107+09:00 (JST)

## Question
What is BotShelf Vampire Library Verified policy vs Untested, and what did Proof Engine 2026-09-10 conclude about monitoring vs data-analysis distribution?

## Answer (sourced)

### A) Library Verified vs Untested (public pages)
- Public hub (`/library/`, `/library/teams/`): **10 Teams · 70 implementations · Verified 3 · Untested 67**.
- Ollama index (`/library/ollama/`): **Verified 3 · Untested 7 — status from registry**. The three Verified Ollama packs observed today: Writing Draft, SEO Brief, Deep Research. **Monitoring Check — Ollama** and **Data Analysis — Ollama** are **Untested**.
- Per implementation pages (e.g. `/library/ollama/monitoring/`): Untested copy states recipe reviewed for safety (no secrets, no destructive commands, no auto-spend); **mark tested after you run it locally**. Structural filter notes say **structural only — does not set Verified**.
- Operator/policy reinforcement (also in support-triage SYSTEM / SAMPLE ops notes): **Verified is not flipped from seller request alone.**

### B) Proof Engine 2026-09-10 — monitoring vs data-analysis distribution
From local Proof Engine files (agent `local-file-audit`, not Claude):
- **monitoring**: **distribution candidate**; EVIDENCE PASS; measured curl paste; no invented outages; in FINAL-ASSETS survivors. (Still does **not** equal Library Verified — locks forbid Verified flip.)
- **data-analysis**: **not a candidate**; structure PASS but invent bar fails (1b MAJOR math/column scramble; 3b still invents dates/blanks/causation per MODEL-COMPARE).
- Survivors packaged: **monitoring** + **switchboard-cos** only. support-triage HOLD for distribution despite 3b invent improvement.

## Conflicts
See `03-proof-engine-audit.json` → conflicts C-RD-001, C-RD-002.
- C-RD-001 (med): Public Untested badge vs PE “distribution candidate” — **resolved prefer-source for badge**; PE = internal share posture only.
- C-RD-002 (low): monitoring vs data-analysis distribution split — documented intentional divergence.

## Runtimes that actually ran (sequential)
1. `web-fetch-curl` — 8 library URLs HTTP 200
2. `ollama:botshelf-deep-research-3b` — elapsed 13.758s
3. `local-file-audit` — Proof Engine markdown
4. `executor-synthesis` — this document
- Skipped: ChatGPT (unavailable) — **not claimed**; Claude Code (not required beyond file audit)
- X discovery: **UNAVAILABLE**

## Confidence
**medium** for policy description from public HTML + PE files.  
**low** on any unspoken Verified criteria (open question from Ollama; public pages do not publish a full criteria checklist beyond safety review + local-run mark + registry status).

## Open questions / next checks
- What exact evidence package flips registry Untested → Verified? (not fully spelled on public pages)
- When/if monitoring distribution share is Human-Yes'd, keep badge Untested until registry process completes.

## Human Yes still required for
- External publish of this brief, mass-DM, Library Verified flip, any customer send.