Vaulter Rebuild Orchestration

You are the orchestrator. Read docs/REBUILD_PLAN.md first — it holds the architecture,
the MCP inventory, and the numbered blockers. This skill covers only how to run the work.

Status — the old hard gate is retired

This skill used to say "delete nothing until the M365/Teams connector is confirmed available."
That gate is resolved: connectors were ruled out, and the deletions happened anyway — the
OneDrive-synced library made them possible without a connector. ingestion/, rag_engine.py,
email_reader.py, outlook_auth.py, ChromaDB, the scrapers, and the scheduler are all gone.
See docs/REBUILD_PLAN.md §0.

Don't re-apply the old gate, and don't propose reviving those modules.

What's built vs. what's left

Built (REBUILD_PLAN §§1–2): the system/corpus/ document layer, system/portfolio.py, the slimmed MCP
tool surface, the keyless geo swap, zero required API keys.

Left (§§4–7): the buy-box standard as a readable document (Phase 0), Tier A/B/C area
intelligence, .msg support, and Phase 3 moving in-conversation.

Key fact to keep in mind: the firm's library is already a synced local folder
(config.CORPUS_DIR), a large number of files, searched by filename via a SQLite index. Reading it needs
no connector and never did.

Delegation map

Task	Agent	Notes
Read portfolio docs, extract findings	vaulter-document-reader	Fan out one per document or property folder. Handles scanned/visual/long docs.
Build a jurisdiction dossier	vaulter-city-researcher	One per jurisdiction — parallelize freely, they're fully independent.
Check a derived number or cited signal	vaulter-fact-checker	One per claim. See the mandatory rule below.
Code changes (geo swap, Phase 3 move)	do it inline	These are small and localized; a subagent adds overhead without isolation benefit.

The mandatory rule: verify before ratifying

Any number derived from prose documents, and any trajectory signal headed for an
investment memo, goes through vaulter-fact-checker before it's treated as true. This is not
optional ceremony — it's the structural answer to the plan's two stated invisible-failure modes
(a wrong disqualifier silently killing deal flow; a fabricated CIP project in a memo).

Fan out one verifier per claim in parallel, then:

• CONFIRMED → may become a disqualifier, pending human ratification.
• INFERRED → may become a preference (weighted, never eliminating). Never a disqualifier.
• REFUTED / UNVERIFIABLE → drop it, and report that you dropped it and why.

Never silently discard a refuted claim — the fact that a plausible-sounding threshold failed
verification is itself useful information for whoever ratifies the standard.

Phase 0 — a first draft already exists

docs/COMPANY_PROFILE.md is written. Don't start it over. It was derived on 2026-07-27 from
firm-wide templates, five deals analysed in depth, the 2023 underwriting models, and a full
inventory of ~57 active holdings / ~137 exits. Every claim carries a citation, and §8 lists what
could not be established.

Settled, so don't re-investigate:

• The library structure is !PROPERTIES/<STATE>/<Property>/00. Pre-Acquisition … 06. Disposition,
  under the firm's synced document library. Old blocker #5 (unconfirmed folder name) is closed.
• The firm has no written buy box — criteria were tacit. The Property Analysis form in
  01. Legal\Entity\Investors\Property Analysis\ is the closest thing, inconsistently filled.
• Deriving numbers from prose is the risky part; §8 of the profile records what stayed unverified.

What's actually left on Phase 0:

1. Human ratification. The profile is marked draft/unverified throughout and is built for a
   senior person to correct, not approve. See [[user-new-to-team]] — the user cannot validate
   historical firm criteria themselves.
2. Fan out vaulter-fact-checker over any numeric criterion before it hardens into a rule.
3. Keep the disqualifiers vs preferences split — the halves fail differently. A wrong
   disqualifier silently destroys deal flow; a wrong preference just misranks.

If extending the evidence base, vaulter-document-reader over more closed-deal files is the way —
but the patterns were already consistent across five deals from three eras, so expect
diminishing returns versus getting a human to correct the draft.

Reporting back

Always surface: what was verified vs. assumed, what was dropped and why, and which gaps remain.
The plan's whole premise is that the standard becomes readable and challengeable — so hiding
uncertainty defeats the point.