MARS · Substrate manifest · 2026-08-05

Company manifest by entity kind

Field-coverage matrix, broken out by companies_v2.entity_kind — the taxonomy landed via tasks #336 (deterministic classifier) + #339 (Qwen-classify the unknown tail).

The previous flat coverage numbers ("96% of companies missing description") were misleading. Two-thirds of companies_v2 rows are sec_shell_unknown — SEC filing shells with no operating business, where a missing website or description isn't a gap, it's the correct absence. This report breaks coverage down by what an entity actually is.

The field-coverage matrix

Percentages are share of rows within that entity_kind with the field populated. Colors: ≥80% populated, 40-80%, 10-40%, <10%.

Entity kind Rows % total DomainDescriptionIndustriesCountryCityStateFoundedEmployeesTickerCIKDealsAtlas
Operating co (rich) 84,0159.3%74%66%41%69%28%26%2%0%6%13%71%0%
Operating co (thin — domain only) 27,1953.0%100%0%0%97%94%94%0%0%2%94%0%0%
Operating co (no domain, has deals) 7,3920.8%0%49%5%24%0%0%0%0%0%1%88%0%
Investment vehicle191,66721.2%0%0%1%14%13%13%0%0%0%87%0%0%
Foundation / nonprofit 18,4532.0%0%0%1%1%0%0%0%0%0%0%0%0%
SPAC / blank-check 4,7740.5%7%5%5%33%7%6%0%0%6%44%8%0%
Individual (person) 7280.1%0%0%15%5%0%0%0%0%0%0%0%0%
Regulatory body 1,5690.2%0%0%8%32%0%0%0%0%0%0%0%0%
SEC shell (CIK only)563,73562.3%0%0%0%23%23%23%0%0%0%100%0%0%
Unknown (residual) 5,6310.6%0%0%10%12%0%0%0%0%0%0%0%0%

Reading this table

Operating company (rich) = the target class for CB/PB-style consumer surfaces. If we want to compare to their coverage, use these rows' column percentages, not the whole-table flat number.

SEC shell (CIK only) = 63% of rows but correctly has 0% coverage on domain / description / industries / country. These are Form D filers, private-fund vehicles, etc. — they exist to file SEC forms, not to have a web presence. They contribute to the total row count but not to any "gap" a consumer would care about.

Investment vehicle = RIA firms, wealth-management LLCs, family offices, capital partners. High CRD coverage expected; low funded_year expected (they're mostly private RIAs the extractor doesn't have birth-date signal on).

Individual / regulatory body = detection classes surfaced by Qwen this session. These are things that were previously in the unknown bucket that shouldn't have been minted as companies at all. Tail work for a future cleanup pass — either delete from companies_v2 or migrate to the correct entity table.

The honest gap number a consumer would see: take operating_company row + column of interest. For a consumer targeting real operating businesses, operating_company (rich) + operating_company_thin + op_co_no_domain combined = the addressable universe. Sum those n's, then average their field percentages for the true "operating universe coverage" number.

Provenance

Task #336 shipped the deterministic classifier (SQL trigger, ~85% coverage from row-local signals). Task #339 shipped the Qwen residual — 45,499 rows Qwen-classified out of the ~50K unknown tail, methodology entity-kind-qwen-v1, every decision reversible via identity_decisions_v2. ~5,600 rows still in the residual (~9 Inferno chunks still trickling in as of report generation).