For forty years, whoever reached the fact first won. That race is over.
Information used to be scarce, and access was the whole business — a terminal on the desk, a licensed feed, an analyst paid to read the filing you couldn't. Every incumbent in this market is a monument to that scarcity. It no longer exists.
Then · 1985–2023
Information was scarce and expensive. The moat was access: reach the fact first, charge for the privilege. Depth, licensing, and a sales force defended it.
Now
Any model reads any filing, summarizes any market, in seconds, for cents. Access is no longer defensible. The question flips from "can you get the information" to "can you trust what the machine did with it."
The free machine fails hardest exactly where the decisions are expensive.
These aren't edge cases you prompt around. They are structural — and they are precisely why a naive "just ask the AI" competitor is not a competitor at all.
The frozen cutoff
A model's world ends on a training date; markets don't. Ask about this quarter and it answers, confidently, from last year — with no signal that it's stale.
No point-in-time truth
Models reason with hindsight baked in. Every backtest silently leaks the future. "What did we actually know on this date?" is a question a raw model cannot answer honestly.
Satisficing — the deep one
A model stops digging when the answer looks complete, not when the data is exhausted. It reads three reserve tables, finds a plausible number, and quits — missing the other nine. You can't prompt it away; it's how the thing works.
Not a data vendor. Not a model. The deterministic layer between them.
When synthesis is free, value moves to the three things a model can't do for itself: reach data that will never be in a training set, know what was knowable on a specific date, and extract exhaustively and deterministically rather than plausibly. We do that work — the aggregation, the joins, the point-in-time gating, the exhaustive extraction — so the model is left to do the one thing it is genuinely good at: reason.
Sources
Filings & private feeds Public disclosure + private funding, M&A, people — much of it never public.Kaleidoscope · MARS v2
Extract · structure · point-in-time · verify The deterministic substrate. Exhaustive, cited, stamped with when each fact became knowable.Any model
Reasons over it The non-deterministic step — and only that step — is left to the LLM.This is where the next year goes.
Each frontier is a place a foundation model structurally cannot follow — and each already has a working proof, not a slide.
Point-in-time as ground truth
Every fact stamped with the moment it became knowable, so a query can be answered as-of any past date — leak-free. Already our house standard; becoming an honest-backtest layer no incumbent offers at a self-serve price.
Live · point-in-time gating across the catalogPrivate-market intelligence
The data a model will never memorize because it was never public: private funding rounds, M&A, and a person-level wealth-and-affiliation graph. All of it entity-resolved into MARS v2 — ~250K canonical companies, ~100K investors, ~1.4M persons, ~116K canonical deals across funding and M&A — refreshed hourly, anchored on strong identifiers (CIK, ticker, CRD, LinkedIn), and never released to a training set. Eight verticals now write direct to the substrate. A frontier that gets more valuable the better public models become.
Live · MARS v2 substrate · hourly delta canonicalizationExtraction that refuses to satisfice
We fight the model's laziness by doing the digging deterministically and exhaustively — reading every table, not the first plausible one. Every extracted fact is scored on a live integrity tier (verified / vouched / warn); anything the machinery flags as suspect is quarantined and never surfaces to a consumer query. "Won't quit early" — and "won't ship garbage" — become filters you can name, not promises you have to trust.
Proven · same model, same filing → +50% more verified factsWe get stronger as the models get better.
Most AI products are a bet against the model — scaffolding the next release makes redundant. Kaleidoscope is the opposite bet.
A better reasoner sitting on top of exhaustively-extracted, point-in-time, private-inclusive facts is strictly better than a better reasoner sitting on its own frozen memory. Every frontier-model improvement makes our layer more valuable, not less. We don't compete with the frontier — we compound with it.
The thesis is already shipping.
Live today
- MARS v2 — the deterministic substrate beneath everything below. ~250K canonical companies, ~1.4M persons, hourly delta canonicalization, eight verticals writing direct.
- Nine vertical intelligence products on a self-serve, connect-and-go protocol — no terminal, no seat license.
- ~116 tools spanning SEC filings, ownership, credit, biotech, private markets, macro and more.
- A reports site regenerated nightly where every number is provably traced to a filing — a marketing asset and a continuous proof-of-correctness in one.
Shipping next
- Earnings-call transcripts — the one real coverage gap, and a large one to close.
- The prospect-research product on the private wealth-and-affiliation graph.
- An extraction-quality roadmap that turns "won't satisfice" into a measured guarantee, domain by domain.