01Language model
Likelihood of text
Produces an articulate judgment and compresses a large context. Helpful for interpretation; not inherently resolved against the world.
Model note · 2026
Elementary Deduction · EPM-1
Forecasts should
settle.
A model for turning time-aligned evidence into event probabilities and market price distributions—not another layer of fluent text.
EPM-1 · Predictive modelOutcome-trained
From evidence to outcomeFuture state,
priced.
priced.
EPM-1 connects evidence, event probability and market response in one structured forecast.
EvidenceProbabilityPrice
Outcome-supervisedTime-boundAuditableMarket-awareContinuously revised
A different prediction object
The next world state,
not the next word.
General models are useful for reading and summarising. EPM-1 adds a second discipline: every view must have a probability, a time window and a future result it can be scored against.
02Research system
Relevant information
Finds material, organises sources and supports analysts. The result is a research surface rather than a calibrated forecast.
03EPM-1
Probability of outcomes
Learns event likelihood, timing and the associated price range. Predictions are frozen, settled and fed back into training.
Illustrative · 90 days
One forecast. Two distributions.
The system first estimates what may happen, then maps that scenario into a market-aware price range. Settlement rules and the evidence cut-off are fixed in advance.
The prediction object
One model for
event and price.
EPM-1 answers two linked questions: will an event occur within a defined window, and how could markets reprice if it does?
Event state encoder
Compresses mixed, cross-lingual evidence into a time-indexed representation.
Causal dynamics
Models how entities, policy changes and events alter the path of future states.
Probability head
Produces calibrated event likelihood, timing and scenario weights.
Pricing head
Maps each scenario into asset direction, ranges and transmission effects.
Training objectivesBrier lossLog lossCRPSCalibration
Evidence and outcomes
Evidence becomes
training signal.
The advantage is not simply finding more documents. It is preserving what was knowable at a point in time and learning from the eventual result.
Discover
Find low-indexed, cross-lingual and specialist signals.
Capture
Turn changing pages and live sources into repeatable records.
Provenance
Attach timestamps, versions, reliability and entity context.
Resolve
Link a forecast to what happened and how prices responded.
24M+source records
Continuously cleaned and time-aligned.
1.8Mstructured events
Entities, relations and settlement rules.
412Kresolved outcomes
Labels used for outcome supervision.
38Mprice-response labels
Multi-asset and multi-horizon observations.
Model to product
Distribution changes.
The model compounds.
Agent, API and workbench are delivery surfaces. The durable core is a shared forecasting model, its outcome data and a consistent evaluation system.
01Observe
Forecast agent
Monitors evidence, revises probability and explains what changed in a compact operating view.
02Integrate
Model API
Institutional systems call probabilities, price ranges and revision history directly.
03Operate
Private deployment
Dedicated model paths can combine client data, controlled evaluation and single-tenant serving.
Production is the
starting condition.
The team has already operated real-time evidence processing, multi-step reasoning, burst load and continuous revision in a live environment.
- Production users
- 3,000+
- Decision cycles
- 120M+
- Median revision
- 4.8m
- 90-day uptime
- 99.95%
A focused first conversation
Bring one decision
that matters.
We will define the event, settlement rule, horizon and price response that would make a model comparison useful.