Ford classification and handoff
The intelligence flow is mapped, verified and shipped. This initiative covers the specific gaps Shawn named on the 2026-08-04 call that the shipped pipelines do not yet handle — how material gets classified on the way in, and how a company Ford learns about surfaces in the books that already own it.
Scope — cut this down rather than start from blank
| Item | Note |
|---|---|
| Private vs public company split on ingest | Called out directly; absent from the Atlas/Ford spec |
| Company-related vs LP-related classification | The spec is company/sector centric and silent on LP research, but a lot of what comes in is LP material |
| Terminal → Ford new-company profile | A deal Terminal surfaces that we know nothing about queues a Ford profile (~$1 web scrape), then shows in Terminal / DealFlow behind a toggle beside the existing record |
| Library ingestion gap | ~600 documents handed over, 297 in the library |
| IPO / S1 wiring | Ford already breaks down an S1; Terminal's IPO tab should open it |
| Positioning brief | Not the right default view of a VC firm — reserve it for Quill and website work, show a plainer firm summary instead |
| Noise filtering | Terminal's literature feed already solved suppressing mega-mentions (Apple, OpenAI). Reuse it; the intake review screen's matched / new / noise triage is the place |
Decisions already locked — do not relitigate
- Records layer, they do not merge. DealFlow stays the resident owner of a company.
- Webster links, never merges; nothing writes into a CRM book directly.
- LP Flow and DealFlow databases stay separate, deliberately.
- Terminal and Podcast ingestion are not to be touched — they publish residue to Ford.
- Every fact carries source + as-of, enforced at the database level in Ford.
- Derived flags are computed by code, never asserted by a model.
- Propose-then-approve for anything writing into a CRM book.
- Every artifact carries a visibility level; confidential material is excluded from Quill outputs and general search by construction.
- Claude proposes taxonomy terms, people approve. The taxonomy grows deliberately.
Open question
On the call Shawn suggested folding Webster into Dewey — "having two different things is too complex." The fleet map, verified the same day, keeps them separate, and the intelligence-flow brief says plainly that entity mapping is Webster's job and not Dewey's.
Recommendation: keep them separate. An entity spine and a document substrate are different concerns, both are shipped and running, and the objection reads as naming confusion rather than architecture. One sentence of confirmation and this closes.
Dependency
Webster's review queue holds 1,413 pending fuzzy name-matches. Until those are triaged, some Terminal and podcast mentions float unlinked to their DealFlow org — a data-quality ceiling on anything built here.
Out of scope
- Rewriting DealFlow's company page. Layer the records, do not merge them.
- Changing Terminal or Podcast ingestion.
- One perfect ontology. Taxonomies get smarter per document; they are not designed up front.