Most AI in finance is a model that emits a score. Ours is a cognitive loop situated in the event stream: it observes, retrieves, reflects, decides, and scores its own estimates against what follows.
Some of what follows is running today. Some of it is the bar we are building toward. We are precise about which is which — that precision is the product.
Inference is event-driven rather than polled. When a state change occurs — a candidate signal, a decision point, a realized outcome — agents instantiate, condition on the current feature lens, retrieve from memory, and emit a judgment with an explicit probability estimate. The estimate is a commitment: it is scored against what subsequently occurs.
Live data yields hundreds of measurements per state. Agents attend to a distilled subset and commit still less to the memory stream: episodes that carry information, each written with an importance score at the time of encoding. Attention is engineered; retention is conditional. That economy is what keeps retrieval tractable as the stream grows.
Experience is stored as timestamped episodes in a memory stream and surfaced by a composite retrieval score — the Stanford generative-agents formulation, adapted to market state. In human cognition the first term is recency; here it is regime fit: an episode encoded under a state the classifier reads as congruent with the present one outranks a merely recent episode, so a session from years ago can dominate this morning's. Importance is assigned at encoding from magnitude and surprisal, so a confidently mis-estimated episode outranks an expected one; relevance is embedding similarity over structural state, not lexical overlap. Accumulated over thousands of episodes, this is engineered intuition rather than a heuristic.
At intervals triggered by accumulated importance, agents reflect: they query their own recent stream, pose the salient questions it raises, and synthesize higher-order insight — each insight written back to the stream with pointers to the episodes it was derived from. Insights carry a lifecycle: proposed, active, inverted, retired. When a high-confidence estimate is contradicted, the agent can attribute the miscalibration and invert the insight responsible, with its citation set intact. Beliefs here are inspectable objects.
An unscored judgment carries no information. Every estimate resolves: accepted decisions against the realized outcome, declined decisions against a deterministic counterfactual simulation over years of historical state. Because the counterfactual is always observed, improvement is estimated rather than asserted — reported as calibration, discrimination, and avoided-loss capture.
Following the meta-labeling principle, the secondary intelligence is not built to generate additional candidates. It is built to decline the unfavourable ones and to quantify how well it declines. Precision, calibration error and avoided-loss capture form the objective — not candidate volume.
Before an agent's estimates are consumed, it is run over years of decision history in strict chronological order — no look-ahead — accumulating episodes, reflections and insights and banking a scored record. We then require an ablation-verified learning curve: calibration must improve as the memory stream grows, and must degrade when memory is removed. If accumulated experience is not measurably improving the agent, the architecture returns to the bench.
Specialized agents evaluate identical states independently, with estimates committed before any exchange. They then conduct structured disagreement — one rebuttal each — and the exchange itself is scored. If deliberation improves group calibration it is retained; if it induces herding, the independent estimates prevail. The communication protocol is subject to the same evaluation as the agents.
Any agent can account for any decision in natural language, drawn from its own stream: the observations retrieved, the insights applied, the estimate emitted, the outcome realized, the revision that followed. Interpretability is an interface, not a log file parsed after the fact.
Agent cognition here is grounded in open, peer-reviewed work. The memory stream, retrieval function and reflection hierarchy follow the Stanford generative-agents architecture (Park et al.): an append-only stream of observations, a retrieval function scoring recency, importance and relevance, and reflection that periodically synthesizes episodes into higher-order insight which itself re-enters the stream and can be cited. In our adaptation recency is displaced by regime fit — elapsed time carries little information about a market state, while congruence of state carries most of it.
The decision architecture follows meta-labeling from the quantitative-finance literature (López de Prado): a primary process surfaces candidates; a learned secondary model estimates which merit action and at what conviction. Estimating the quality of a candidate decision is a distinct — and more tractable — problem than generating one.
We credit those foundations and extend them. The proprietary layer is how importance is assigned at encoding, how retrieval is conditioned on regime, how insight is converted into a calibrated estimate, and how each estimate is scored afterward. We claim no affiliation with or endorsement by the authors of that work.
A single capable agent is the product. The destination is larger. We are building toward continuously running simulated societies of agents — populations resident in the data around the clock, through the session and beyond it: observing, retaining, reasoning over one another's estimates, deliberating, and widening their own decision authority only as far as their scored record supports.
Experience compounding into judgment at a scale no single model reaches — an architecture that continues to adapt as the underlying models advance.
The foundation underneath it is already sound. The society being raised on it is where we believe this field is heading — new ways of dealing with data, of finding patterns, of deciding. It is early, it is proprietary, and we are precise about the difference between what runs today and what we are reaching for. We'll say more when it's ready.