Imitation learnerThesis
Learns the crowd
Imitates participants, including poor ones, because the environment has to contain their mistakes. Each cohort keeps a persistent state for beliefs, attention, horizon, financing, mandate and liquidity.
Fiscal AI
Cotangent trains an investor inside every version of the market that still agrees with the evidence, and it throws away any profit that only exists inside the simulator.
Compatible set, simulatedIllustrative
0 agents · 5 populations
Physical AI trains a body in a world of contact and mass. Fiscal AI trains an investor in a world of participants and institutions, who are only partly visible and who change their behaviour when a new trader arrives.
A correct forecast can still lose money.
The earnings print arrives exactly as modelled, and the position loses anyway. Anyone who has run a book has watched it happen. A model that stops at the company, the narrative or the next return has nowhere to represent why.
- 01
It was already in the price
Other participants reached the same forecast earlier and traded on it. The print confirmed what the market had already paid for.
- 02
Someone was forced to sell
A fund hit a redemption, a mandate limit or a margin call. Its selling had nothing to do with the forecast.
- 03
Your own order moved the book
Size changes the price you get. A strategy that ignores its own footprint is tested in a market that does not exist.
Those three facts sit between forecasting and investing. Cotangent is built to represent them.
Many crowds can produce the same tape. Cotangent keeps all of them.
Public prices, trades, quarterly holdings and documents fit many different populations of investors. One crowd was hedging, another was forced, another changed its mind. They match history equally well and then split on the next shock.
Picking the single most likely crowd is how a model becomes confidently wrong. Cotangent keeps every population the evidence still allows and trains its investor to hold up in each of them. When two surviving populations imply different portfolios, it takes the action that survives both, or it pays, inside a risk budget, for the observation that tells them apart.
Compatible setIllustrative
Five investor populations fit the same tape up to today and disagree about tomorrow. Cotangent trains one policy that has to hold up in all five.
Drawn by hand. No measurement exists.
Two learners, trained together
The world model and the investor improve each other. A sharper census of the crowd gives the policy harder and more realistic worlds to train in. Every place the policy fails shows where the census needs more detail.
Policy learnerThesis
Learns to invest
Chooses Cotangent's actions for Cotangent's objective, with risk, liquidity, costs and constraints written into that objective. Imitating investors builds the environment. Investing well is a separate objective with its own learner.
Four steps, in order, with the score at the end.
Tape
Prices, filings, prints and news, as they were knowable then. Every source carries a timestamp, and every pretrained encoder carries its knowledge cutoff.
Compatible set
Every investor population and market transition that still explains that tape. Cohorts stay coarse and split into explicit participants only where the split would change the portfolio.
One policy
Trained so it still works if any member of the set is the live one. Detail is spent only where two members would change the book.
Next tape
The result is the net of costs on later real observations. Gains that exist only inside the simulator stay out of the number.
What the system enforces
The ledger is exact
Cash, inventory, collateral, borrow, settlement and matching follow the venue's rules wherever the rule is known. Learned models supply behaviour and uncertain transitions around that ledger.
Slow thinking runs when it can change the action
Fast shared policies emit distributions over actions. Slower search and solvers fire on disagreement, unfamiliar states, downside and the value of the next observation, and what survives the outcome check is distilled back into the fast policy.
The simulator cannot be the source of profit
A planner will find the bug in a market model and report it as an edge. Cotangent shortens rollouts, caps action size and discounts any world where simulated paths and real outcomes diverge.
The market reacts to the investor
Participants change their behaviour because systems like Cotangent exist. The simulator responds to Cotangent's own orders, and the compatible set is updated when the crowd adapts.
Cotangent simulates the disagreement that moves the position.
A society of language-model investors is too slow to train against. Cotangent runs shared fast policies, persistent cohort states and event-driven updates, and keeps language-model reasoning for consequential changes in strategy and belief.
- Simulated participants
- 15,000
- Decisions per simulated day
- 5.85M
- Decisions across 1,000 scenarios
- 5.85B
Model assumption15,000 agents × one decision per market minute × 390 minutes × 1,000 scenarios. Arithmetic from the research memo.
One research core. The investor comes first.
Cotangent owns the investor, and that is the company today. Once a transfer result exists, the same world can test other institutions' policies.
- Venture partnersPlanned
A scoreboard written before the result
The claim is a decision that survives the next tape after costs, measured against a plain predictor on a chronological holdout. The kill rules are published, so you can see in advance what would stop the work.
- Asset managersThesis
Test a policy against every crowd that could be out there
Run a strategy through each population still consistent with the tape and see which unresolved difference could reverse it. Stress tests built from participants, mandates and forced flows, where historical replay has only one path.
- Treasuries and exchangesThesis
Hedges and rules tested against participants who adapt
A treasury hedge judged by the loss it avoids and the constraint it meets. An exchange rule tested against the participants who will change their behaviour because of it.
The scoreboard exists before any architectural result is allowed to count.
The first world is liquid equities and a few sectors, around earnings and scheduled macro releases, over days to a few weeks. Execution is inside the score as spread, fees, impact and liquidity.
- 0
Scoreboard
Months 0 to 3A direct predictor has a net-of-cost number on a chronological holdout, with spreads, fees and a stated impact model in the number.
- 1
Ledger and graph
Months 3 to 8A typed temporal graph of a few dozen liquid names, every edge marked observed or inferred, and one earnings or macro revision traced along a path you can name.
- 2
Census
Months 6 to 12At least two populations remain compatible with one tape and imply different holdings. The record shows the observation that would eliminate one of them.
- 3
Policy
Months 10 to 16Net performance on a later real window against the phase 0 baseline, registered before the window opens. This is the number for the next round.
- 4
Second clock
Months 15 to 18Slow reasoning fires on recorded triggers and is distilled into fast policies, which are checked for information a participant could not have had.
Kill rules
Each phase can fail, and the phase that fails names the component to remove.
- Stop
The census does not change decisions relative to the direct predictor. The psychology work stops and the ledger plus the baseline remain.
- Stop
The gain disappears once the knowledge cutoff is respected. The gain was leakage.
- Stop
The profit depends on transitions the tape does not support. It stays out of the result.
- Stop
The simulator resembles past prices and leaves the score unchanged. That is a stopped experiment.
A financial world model that learns how people and institutions decide, simulates how they interact, and learns to invest inside that changing world.
Physical AI became fundable when a body could practise in a simulator that pushes back and the gap to the real room could be measured. Markets push back too, and they rewrite themselves when a new participant arrives. Cotangent measures that gap as the policy against the next tape, after costs, while the crowd is allowed to change.
The first world is equities around earnings. The same machinery extends to credit, commodities and treasury hedging, wherever a shock travels along a path that can be named: a forecast revision, a change in output, an inventory, a financing need, a portfolio flow.
ThesisThe mark is a C holding an o: one policy, resting on the set it was chosen inside.
- C · outer diameter 88 on a 96 grid, stroke 10, 80° opening
- o · 68% of the inner diameter, stroke 6
- o · tangent to the inner circle at its lowest point
Write to the founder.
Investors get the measurement protocol and the eighteen-month plan. Managers, treasuries and exchanges can join the list for the first policy tests.
Or write directly todavid@embino.com