State of the Union

MARS has crossed from "a funding-and-M&A pipeline" into a general SEC-and-public-records entity graph. The deal core is production-grade today; the newer wealth, people, and event layers are large but carry the resolution debt you'd expect from name-only sources. The work left is not ingest — it's resolution and enrichment: draining ~1.7M unresolved persons, executing ~92K queued merges, and thickening the ~117K operating companies. Signals' pipeline is now live and backfilling the event layer that was empty a month ago.

The core is production-grade

Deals — the original mission and the surface real customers touch — are mature: integrity-tiered, deduped by content hash + range window, swept historically, and gated at write. 95–98% of deals sit in the consumer-usable verified/vouched tiers.

The warn tier (extraction anomalies, unit-scale errors, huge-IPO escapes) is filtered from every consumer surface by contract, so the residual garbage never reaches a user. This is the layer to point a paying customer at today.

The entity graph

Private vs public — the mission lens

MARS exists to cover private companies — the Crunchbase / PitchBook game — so the split that matters isn't "has SEC data," it's traded vs not. And a CIK is not a public-company marker: hundreds of thousands of private companies carry one because they filed a Form D private placement. On the real axis — ticker — the platform is overwhelmingly private:

So "77% carry a CIK" is a strength on private data, not a sign we're just reprocessing public filings — most of those CIKs sit on private Form D issuers, giving MARS SEC-grade identity resolution (strong-ID dedup, cross-linking, exact CIK/CRD joins) on companies CB and PB only hold as free-text names. Of the private set: ~728K are SEC-known (Form D), ~215K are purely news/web-sourced.

Under the hood: it's two populations, not one

The 952K company rows split cleanly by entity_kind, and the "low coverage" numbers are almost entirely honest substrate, not a gap:

~117K genuine operating companies is the consumer-facing set — and 95% of them are private (only 5,537 publicly traded). The thin fields — description (61K), domain (91K) — are concentrated in the shells by design; the enrichment gap that does matter is on these operating companies, and it's tracked.

How connected are the companies?

Raw counts across all 952K look thin — but that's the shell dilution again (a Form-D issuer should have no employees or bios), so the meaningful denominator is the ~102K operating companies:

Deals and people are the strong dimensions on operating companies (62% / 54%). Bios (43%) and investors are the thinner ones — bios because the ADV/web crawl skews toward RIA/finance firms, investors because we only capture named backers and many rounds don't disclose. Those two are the enrichment levers with the most headroom.

Persons: the resolution frontier AMBER

3.1M persons is the headline growth story and the headline debt. 1.72M (55%) are flagged entity_resolution_pending — but that debt is not evenly spread. It's concentrated entirely in the two name-only mint sources, while every strong-ID-anchored source is clean:

Read "pending" as "minted, awaiting dedup" — not "unminted." Every one of these rows is a real person with a real name in the graph. entity_resolution_pending=TRUE is the resolution status: a name-only mint (e.g. a property owner off a deed) can't be auto-confirmed as new-vs-duplicate against the rest of the graph without the resolver pass. The mint worked; the dedup is what's outstanding.

The lesson is clean: where a strong ID exists (Form 4 CIK, ADV CRD, SEC), resolution is essentially solved. Where the source is a bare name off a property deed, it isn't — and can't be, without the deeds→persons linker and the tier-chain resolver doing their passes. "Production-grade persons" today = the ~1.4M strong-ID subset. The name-only tail is real data, correctly held out of consumer surfaces until resolved.

The new layers GROWING

Also live and daily-fresh: adv RAUM history (399K PIT rows) + regulatory events (9.9K), lp_data fund performance/commitments, flare thermal observations (1.0M), tranche debt maturity wall (1.2M), 13F manager AUM (194K).

How much deduping is left

Three distinct queues, very different sizes and difficulty:

Every mutation is logged to identity_decisions_v2 (6.97M rows) — the whole platform is reversible. The gate has processed 1.27M submissions (1.26M promoted, 45K in review, 32K cleanly rejected). Merging and un-merging are both gated verbs, so draining these queues is safe, auditable, and undoable.

What's left to reach production grade

1. Drain person resolution (the 1.72M). Ship the deeds property-owner → persons linker, run the tier-chain resolver over the bios web/ADV tail. Strong-ID-first (CRD/CIK) where present; Qwen→Grok-web for name-only. This is the single biggest lever on graph quality.
2. Put merges on a timer. 46.5K merge_candidate rows are machine-decided and executable now; move merge_entities to an hourly cadence so the queue stays near-zero instead of accreting.
3. Thicken operating companies. Domain + description + ticker enrichment on the ~117K operating cos (the consumer-facing set) — the shells are fine as-is.
4. Cross-source dedup. Resolve SEC-only shells against news-only operating companies that are the same real firm but share no strong ID — the last structural duplication class.
5. Link wealth → people. The 268M donations / 12M property / 26M foundation-officer rows are only worth their weight once bridged to canonical persons. That bridging is the deeds/donor value unlock.
6. Finish the guards. investors_v2 same-name-different-CRD (franken-investor), consolidate the name guards, close the last ungated writers. Small, known, tracked.