Companies
897,136
+50K net-new this week
Fifteen verticals wrote to the shared substrate this week. A new hedge-fund AUM roll-up landed, a first-of-its-kind cross-vertical signals table went live, and 1.45 million linking decisions were recorded. Absolute row counts are the loud number; what actually matters is the shape underneath — canonical entities, integrity-tiered rows, provenance-tagged decisions, and cross-source joins that a competitor can't reproduce.
Row-level additions, by table. The big two numbers deserve context and get it below.
| Table | What it holds | Added 7d | Total |
|---|---|---|---|
persons_v2 |
Canonical people — execs, directors, insiders, advisors, GPs | +454,342 | 1,935,318 |
person_work_history_v2 |
Work-history edges (person → company → role → dates) | +1,283,448 | 1,516,082 |
person_bios_v2 |
Executive bios from firm websites + ADV Part 2B | +172,756 | 755,025 |
person_education_v2 |
Person → institution edges (Wharton, HBS, etc.) | +233,215 | 326,914 |
companies_v2 |
Canonical entities — 85.6% CIK-linked to SEC | +608,206 | 897,117 |
article_meta_v2 |
News corpus, GDELT + Benzinga + RSS | +12,571 | 2,499,510 |
funding_deals_v2 |
Canonical funding rounds, integrity-tiered | +519 | 80,317 |
merger_deals_v2 |
Canonical M&A deals, integrity-tiered + deal-stage | +337 | 37,507 |
form_d_filings_v2 |
SEC private-placement filings (early-warning signal) | +885 | 766,807 |
form_144_v2 |
SEC insider-sale notices (15-day lead-time) | +292 | 115,853 |
thermal_observations_v2 |
VIIRS nightly satellite radiant-heat readings | +134,798 | 765,982 |
signals_v2 NEW |
Cross-vertical detection surface (first substrate) | +1,945 | 1,945 |
thirteenf_manager_aum_v2 NEW |
13F-derived AUM per manager (unblocks <$500M filter) | +188,906 | 188,906 |
fund_commitments_v2 |
LP → GP fund commitments from public pensions | +491 | 1,033 |
pension_funds_v2 |
US public pension funds (CalPERS, CalSTRS, +13 more) | +15 | 30 |
Two mechanical events explain most of the company row growth. First, a one-shot Atlas backfill minted ~700K rows under the strong-ID rule (CIK match → canonical entity). Second, a hotfix deleted 210K phantom person-CIK rows that had been minted as companies by an earlier automation. Net organic new-company mint this week is closer to ~50,000. The person side is similar: 454K new persons is dominated by insider's Form 4 mint (~199K rows) plus bios's web-crawl output.
We're calling that out because how you count matters more than what you count. A rollup that reports 608K new companies as a growth number would be true and misleading at the same time.
A canonical entity is only as useful as the edges attached to it. Linking activity is the number that shows the graph is doing its job.
1,453,075 rows landed in identity_decisions_v2 this week — every entity mint, every merge, every “these two are the same,” every “these two look similar but are different.” The ledger is the reason we can reverse any mistake without archaeology.
Top methodologies driving the week:
mars-atlas-sync — 701K (SEC identity backfill)bios-web-v1 — 179K (exec bio harvest)insider-form-4-v1 — via person mint (~199K)bios-c11-person-resolve-v1 — 54K (person dedup)mars-v2-linkage-weekly — 6,599 (Sunday sweep)The graph_edge_catalog is a new registry that lets MARS ship a new edge type — say “board seat” or “acquired by” — and have downstream tools discover it instantly, with no code deploy on their side.
Currently indexed:
invested_in, acquired, works_at, board_of, insider_of (bridge tables)board_member_of, officer_of, studied_at, affiliated, founder_of, worked_at (network views)New bridge lands → one catalog row → downstream tool sees it. Zero deploy latency.
A useful comparison requires honesty about two very different games.
The blunt version: they win on breadth, we win on integration. Crunchbase and PitchBook have spent a decade building large, hand-curated deal databases with global coverage. Our raw counts are smaller. What's different is what we do with what we have.
| Dimension | PitchBook (claimed) | Crunchbase (claimed) | MARS |
|---|---|---|---|
| Companies profiled | ~3.5M | ~2M | 897K |
| SEC-linked (CIK) | not disclosed | not disclosed | 85.6% |
| Funding rounds | ~500K+ | ~1M+ | 80,317 |
| M&A deals | ~700K+ | ~350K+ | 37,507 |
| Every row provenance-tagged & reversible | no | no | yes |
| Integrity tier surfaced to consumers | no | no | yes |
| Deal stage ladder (rumor→closed) | partial | no | 94.9% coverage |
| Cross-linked to Form D / 13F / 13D/G / Form 144 | no | no | yes |
| Person → company work history | yes | limited | 1.52M edges |
| Natively AI-consumable (MCP) | no | no | yes |
| Real-time SEC linkage timers | n/a | n/a | hourly |
PitchBook and Crunchbase are optimized for the question “show me all funding rounds in fintech last year.” Both platforms will answer that better than we will today, because their hand-curated coverage is deeper.
MARS is optimized for the question “show me every ex-Merrill advisor now at a $1B+ RIA” — a real query bios ran this week. That's 109 people. The answer requires: person canonicalization across firms, work-history edges, firm AUM roll-up from 13F, RIA-vs-broker classification. All in one place, joined at query time. Neither PB nor CB can answer that question at all — the substrate doesn't line up.
The bet is that the second class of questions is where AI-first consumers will spend their time.
Every graph-scale project accumulates hard-won rules. Here are the ones that landed as durable code or written doctrine over the past seven days.
The freshest addition. A name may be written to primary_name or name_variants only when independently corroborated — matches the existing canonical, or a strong ID (CIK, ticker, CRD, Atlas) backs it. Uncorroborated renderings aren't written at all.
The rule replaces earlier format-based guards (“is it Latin script?”) that turned out to be format checks masquerading as correctness checks. Bios's Phase 2b measurement made this vivid: a Latin-script test passes Cheetah Mobile (correct), HuiPusen (romanized brand for a different company), TechLott (a resold domain), and Kite (a partner logo on the page) all as valid. Corroboration passes only the first — and generalizes to every failure class at once.
Every M&A row now carries a deal_stage — rumor · talks · loi · definitive · pending · closed · terminated — and an is_definitive boolean. This lets consumers filter out rumor noise before it reaches their traders. Coverage is 94.9% is_definitive on the canonical set. The Qwen backfill on the “announced” tail cost $0.
Divisional divestments (“Harmonic Video Business,” “GE Aviation Division”) were routing to the parent company's canonical entity — the target ended up merged with the seller. New extractor guard splits the target off cleanly. 421 divisional entities minted, 466 article rows re-attributed, all reversible.
Funding deals now carry pre_money_valuation_usd and post_money_valuation_usd as first-class columns. Fixed a 10-day silent-drop bug where the prompt asked for these fields but the write path filtered them out. Backfilled 557 marquee valuations via Qwen at $0.
Additive Parquet drop at s3://kscope-news-raw/<source>/ — Benzinga, GDELT, and RSS all write here per batch. Cross-vertical consumers (signals, others) read for extraction without needing to touch the article corpus directly. Went live with a schema-pinned 11-column contract.
The self-describing edge registry described in §2. Ships a new bridge table → add one catalog row → downstream MCP tools traverse it instantly. Kills a real deployment bottleneck between MARS and consumer platforms.
New thirteenf_manager_aum_v2 substrate — 189K rows from smart-money's holdings table, joined into investors_v2 via CRD. Unblocks Raul's “hedge funds under $500M AUM” segmentation — a 90× expansion in the qualifying manager pool.
MARS is a platform. Verticals write into shared canonicals under distinct methodology tags, all reversible. Fifteen wrote something this week.
funds_v2 live.Spider handover to bios (top priority). The Playwright-based crawler is being retired in favor of bios's Brave-Search-first crawler. Same monthly cost band (~$300/mo), better data quality (in-text Qwen extraction vs Grok), and eliminates a whole class of chromium/OOM bugs.
Bios direct-to-v2 identity writes. Bios finishes their Phase 2b parallel-diff over the coming days, then flips their v2 writer live. When they do, we surgically disable identity-field writes in the older ETL path — one owner per column, no double-writes.
Signals substrate expansion. Lawsuit events shipped this week. Appointment events, commercial-relationship events, product announcements, contract wins, and five more categories are in the queue behind them — each backed by the same Qwen-at-zero-cost pattern that made lawsuit events feasible.
Extraction preprocessor for funding dates. A deterministic date-picker feeds the extractor a candidate publish-date, which turned out to be the single largest driver of extraction disagreement in our July calibration study. Estimated composite lift: 80% → 92%.
Coverage benchmark study. Twenty recent PitchBook LP News stories, checked against MARS substrate. First real head-to-head on the coverage question we get asked — with the ground rules published up front so the result is defensible either way it comes out.
About this report. Numbers are live queries against the mars Postgres database as of 2026-08-02 UTC, filtered to the seven-day window July 26 → August 2. Row counts include all writes to the shared canonical substrate, regardless of vertical. Weekly growth reported as raw additions; where a one-shot backfill dominates the count (companies, persons this week), the organic net-new figure is called out explicitly.
Reproducing this. Every query is against tables in the mars schema on the shared read replica argos_readonly. Ask the data platform team for connection details.
Design system. MARS report style v1 (2026-07-04) — report_style/base.css in mars_feed/.
Comparison data. PitchBook and Crunchbase figures cited from public marketing pages, current as of publication. Both companies present higher figures in sales materials; we cite the conservative published numbers.