Every "AI for finance" tool indexes documents and answers questions. They get smarter when more users ask better queries. The firm itself learns nothing. That is the read-path. G-Nosis is the inverse.
The reigning architecture for institutional AI is retrieval-augmented generation over a vector index of the firm's documents. It is competent. It produces fluent answers. It is also, by construction, additive to nothing.
Every query consumes context the firm already had. Nothing new is written. The system gets better at finding existing knowledge — never at producing it. When the analyst moves desks, the analyst's reasoning leaves with her, and the vector DB shrugs.
The result is a tool that compounds for the vendor (better embeddings, better retrieval) and a substrate that compounds for nobody at the firm.
The atom of institutional knowledge is not a PDF or a Slack thread. It is a decision: who proposed · what rule fired · who approved · what was cited · what was the outcome.
Each is a first-class event on an append-only, bitemporal log. The reasoning isn't a paragraph buried in a memo — it's a schema-versioned graph edge between the decision and its precedent. Addressable. Auditable. Re-traversable.
A decision recorded with one timestamp is a guess. Was that the rule we believed at 09:14 GMT, or the one we believe now? Did we mean it at the time, or did we retroactively decide to mean it?
Every event in G-Nosis is bound to two clocks. Transaction time — when we wrote it. Valid time — when it was true of the world. Plus the schema version that was live at write. With those three, the past is finally re-readable without a Slack archaeologist.
Reversal is a new event, not an edit. The log only grows.
The trace is the substrate that did not exist before G-Nosis touched the decision. It is new data. It is yours. It survives the analyst, the desk, the system migration, the M&A.
The graph gets denser every time a decision fires. Precedents accumulate. Rules get more discriminating. Risk re-prices learn their priors. The system doesn't get smarter with more users — it gets smarter with more decisions. That is the only kind of intelligence the firm can compound.
The decision trace is the data the firm only ever generates while it is doing the work, and only ever loses when the work is over. — Foundation Capital, on context graphs
The chat surface is incidental. Adoption is measured in decisions captured, not prompts typed. If our UI disappeared tomorrow, the graph would keep being written by API.
RAG retrieves prose. G-Nosis writes edges. Retrieval is a parlor trick on top of a graph that already exists; what matters is the graph.
Embeddings are useful for similarity lookups within a schema. They are not a schema. We don't believe ANN search is a system of record.
We don't route approvals or run a ticketing kanban. We write the decision substrate underneath whatever workflow the firm already runs.
Onboarding is gated to firms that can articulate the decision schema they want preserved. We reply within five working days.