How DeOS turns one multimodal, multi-task in-context learner into end-to-end decision pipelines#
Many applications of the same model, stacked along every dimension of a decision, plug and play at every component, and built for the three kinds of uncertainty that real decisions face.
01 · one model
02 · plug and play
03 · uncertainty
Many applications
Upgrade one part
Built for the real kind
A pipeline of tasks per analysis, each solved by the same learner.
Plug in a better or a custom model, keep everything else.
Chance, ignorance and novelty, each routed to the right response.
Standard large language models excel at retrieving and summarizing text. Complex enterprise decisions need something else. They are made once, in a world that is observed only once, under uncertainty that cannot be removed, and every action changes the world it acts on. DeOS is a Decision Operating System built for exactly this problem, powered by hybrid intelligence.
Everyone is a quant now. That is the problem.
An LLM reproduces the methods that exist, and many were flawed in the first place: linear models of non-linear systems, single-point forecasts, correlations taken for causes, backtests fitted to one history. The code gets written faster; the flaws remain.
Built for the problem, not the tool.
Decisions face three kinds of uncertainty, a single recorded history and questions on the upper rungs of Pearl's ladder. Chance can only be sized, ignorance can be reduced by learning, and novelty has no reliable probabilities, only robustness and options.
Context is the key, and it comes from both sides.
People hold soft knowledge no database has. DeOS elicits their judgments in structure, aggregates them, measures each person's track record and debiases them, and fuses them with foundation models in a human / AI mode per component.
One learner, any modality, any task.
Text, tables, time series, panels, audio, video, images and relational data flow through an ontology into one pretrained in-context learner that solves seventeen task types, zero-shot or from a handful of examples, with no retraining.
An analysis is a pipeline of many model applications.
One decision runs through perceive, diagnose, predict, recommend and learn. Each step applies the same learner several times, to a different task, modality and context, and passes its results on as context for the next.
Plug and play.
Every component works through a fixed contract. A better foundation model, a stronger encoder or a custom-trained model drops into one slot, and every pipeline built on it benefits without being rebuilt.
DeOSAlpha: an optimizer with a memory and a world model.
An ICL optimizer with a memory of every decision, a world model that imagines the paths not taken, neurosymbolic controllers and a value-of-information engine that decides whether to try, act or wait.
Fifteen capabilities, governed across every level.
DeOS was designed against the fifteen capabilities an AI that decides under uncertainty needs, in five groups, with data, analysis, recommendation and audit trail for every solution, decision policy and analysis, strictly within the mandate.
17task types one pretrained in-context learner solves, zero-shot or from a handful of examples, with no retraining.
8modalities, text, tables, time series, panels, audio, video, images and relational data, in one representation for one learner.
15capabilities an AI that decides under uncertainty needs, in five groups; DeOS was designed against this list.
With LLMs, anyone can generate a factor model, a DCF or a backtest in minutes, and many now believe that makes everyone a quant. But an LLM reproduces the methods that exist, and many of those methods were flawed in the first place: linear models of non-linear systems, single-point forecasts, correlations taken for causes, and backtests fitted to one history. The code gets written faster; the flaws remain.
INFORMATION ≠ DECISION
About the future, not extraction LLMs extract and summarize what has been written. Decisions are about what will and could happen, and what to do about it.
CORRELATION ≠ CAUSE
No world model Association between words sits on the lowest rung of causation. It cannot simulate an intervention or a regime never seen.
FLUENT ≠ RELIABLE
Pseudo-reasoning Reasoning-shaped text, not a computation over a model of the world. Ask twenty times, get twenty answers.
What a decision requires, and why LLMs do not deliver it#
Requirement
What a standard LLM does
What DeOS does
Causal
Learns which words go together
Interventional data, counterfactual outcomes and world models
Probabilistic
One fluent answer; no separation of uncertainty types
Full distributions; chance, ignorance and novelty reported apart
Trust calibration
Same confidence whether right or wrong
Every model and every person weighted by their track record
Multimodal
Text first; tables, time series and audio reduced to tokens
Eight modalities in one representation for one learner
Multi-task
Generation; prediction alone is not a decision
Seventeen task types chained into pipelines, through to optimization
Hybrid
Ignores tacit judgment, assumptions and preferences
Human judgment, assumptions and risk preferences as structured inputs
Explainable
Stated reasons may not be the real ones
Every recommendation returns the quantities it was made on
Adaptable
Frozen at training; regime shifts invisible
Adapts in context within a few observations, no retraining
Repeatable
Same question, different answers
Deterministic computation and a full trace
Governance
No mandate, no audit trail
Rules enforced, never learned; every decision auditable
The disagreement is not about model size, context length or benchmark scores. It is about what is being optimized. Written down, the two objectives share no terms.
Language model objective. Maximum likelihood over token sequences: choose parameters θ to maximize Σt log pθ(xt | x<t) over a corpus of text. The output is a distribution over the next token; an answer is one sample drawn from it, one token at a time. Decision objective. Expected utility under a posterior over world states: choose the action a* = argmaxa ∈ A(mandate, constraints) Es ~ p(s | evidence)[U(a, s)], where the belief p(s | evidence) is updated as observations arrive, U carries the decision maker's goals and risk preferences, and the feasible set A is fixed by the mandate.
The first objective is met by text that is distributed like the training corpus. The second is met only by an action that performs well across the states the world could be in. States, actions, utilities and constraints do not appear in the first objective; they enter only as words. A model can drive its loss to the floor on every finance text ever written and still contain no quantity that corresponds to the expected utility of your action. Everything below follows from that gap.
Issue
Why the token objective cannot cover IT
What DeOS does instead
In-context pattern completion is not Bayesian updating
Conditioning on a prompt changes which continuation is likely. There is no explicit state variable, no prior over it and no likelihood of the evidence under it, so nothing is being updated except the probability of the next word. The "update" is whatever the corpus statistics imply for text that looks like the prompt.
The tabular in-context learner is pretrained on tasks drawn from a prior over data-generating processes, so conditioning on rows approximates the posterior predictive for that task (Müller et al. 2022; Hollmann et al. 2025). The state is explicit: a context with named entities, drivers and regimes, updated as evidence arrives.
No calibrated epistemic and aleatoric decomposition
Autoregressive sampling returns one trajectory. The entropy of the token distribution mixes three things that must be kept apart: surface variation (many ways to say the same thing), chance in the world and the model's own ignorance. Temperature is a decoding knob, not an uncertainty estimate, and the training signal rewards a confident guess over an admitted gap (Kalai et al. 2025).
Every predicted quantity is a distribution, not a number. Chance, ignorance and novelty are reported apart (Hüllermeier and Waegeman 2021), stated uncertainty is rescaled against realized errors, and each model and each person carries a calibration record that sets its weight.
No world model, so no counterfactual can be evaluated
A counterfactual query has three steps: infer the unobserved factors from what happened, replace the mechanism that the action changes, recompute the outcome. Each step needs a structural causal model with a do-operator (Pearl 2009). A language model has none, so "what if we had not acted" is answered with the text that usually follows such a question, not with a computation. Models asked to infer causation from correlation perform close to chance (Jin et al. 2024).
Causal diagnosis builds an explicit causal model of the context from observational, interventional and expert evidence; Counterfactual prediction simulates the paths that did not happen under each candidate action; the generative world model produces the regimes history has not shown.
Planning is search, not generation
Decoding selects a high-probability continuation. It does not enumerate actions, evaluate each under a model of consequences, check hard constraints or backtrack when a branch fails. Planning needs a model to plan in and a verifier to plan against; on planning benchmarks, models alone produce few executable plans (Kambhampati et al. 2024).
Contextual optimization searches over candidate actions, evaluates each in simulation against the objective and the hard constraints of the mandate, and the EVI engine decides whether one more trial, acting now or waiting is worth its cost.
Stochastic decoding versus reproducibility and audit
The same question yields different answers across samples, and greedy decoding does not fix it: any change of wording changes the path. The explanation is text generated alongside the answer by the same process, and the stated reasons can differ from the operative ones (Turpin et al. 2023). There is no trace to audit, only a narrative.
Given the same inputs and the same versioned components, the computation is deterministic. Every recommendation returns the quantities it was made on: drivers, scenarios, trade-offs and conviction score. Decision memory records inputs, outputs, the human decision and any override.
Distribution shift and regime change versus frozen pretraining
Parameters are fixed at the end of pretraining. Adapting them needs fine-tuning, which is slow and overwrites what was known, or prompting, which has no mechanism to down-weight stale evidence or to notice that the world has changed.
Adaptation is a change in what the learner reads, so it happens within a few observations and without retraining. Drift detection flags a rise in surprise against the system's own history, and regimes are estimated at four scopes at once.
Data contamination and look-ahead in backtests
The pretraining corpus contains the outcomes of the periods a backtest would cover. A model asked to judge a 2019 filing has read what happened in 2020, and historical tests overstate skill for that reason (Glasserman and Lin 2023). The leak cannot be removed from the weights after the fact.
Evaluation is walk-forward on point-in-time data. The learner's prior comes from synthetic and cross-domain tasks, not from the history of the market under test, so the backtest window is never in the weights.
Scaling and "reasoning" models do not change the objective
Larger models and longer generated chains lower the same loss, or add reinforcement on verifiable answers; the quantity being optimized remains a distribution over tokens. Accuracy still drops when numbers in a problem are changed (Mirzadeh et al. 2025) and collapses beyond a level of complexity (Shojaee et al. 2025). More of the same objective does not produce a different one.
The decision objective is built into the architecture rather than hoped for from scale. Improvements to the foundation learner propagate to every analysis, but the recommendation is always computed from a model of the world, never generated as text about it.
Where language models sit in DeOS. In Perceive context, they read and structure: filings, transcripts, contracts and expert notes become entities, drivers and claims with sources. In Recommend, they explain: the quantities a recommendation was computed from are rendered as language a committee can challenge. They are not used for Causal diagnosis, Counterfactual prediction or Contextual optimization, and they never produce the conviction score. The engine computes the decision; the language model talks about it.
Calling everything "risk" hides the fact that different kinds of uncertainty call for different responses. Chance cannot be reduced, only sized. Ignorance can be reduced by learning before committing. Novelty has no reliable probabilities at all, only robustness and options. Black swans live in the tails that bell-curve models call impossible. A system that reports one confidence number blends all of these, and the decision maker cannot tell whether to hedge, investigate or step back.
FIGURE 1 DeOS separates the three forms of uncertainty and routes each to its own response: chance to sizing and tail-aware risk, ignorance to trying, novelty to waiting and keeping options open. Shapes are schematic.
Every price series, set of accounts and track record is the record of one history. Many others were possible, were just as consistent with the underlying causes, and left no data. A model trained only on the realized path learns its luck and calls it structure. To judge decisions by their quality rather than their outcome, a system must generate the histories that did not happen and evaluate decisions across all of them.
WHERE THE PATHS NOT TAKEN COME FROM
HISTORY HAPPENED ONCE. DEOS REASONS OVER THE PATHS THAT DID NOT.
Gold: the one history in the data. Red: paths with a crisis that, this time, did not happen. Simulated, illustrative.
Synthetic periods does this only look good on one history?
Shadow decisions what would each rule have earned on days we did not act?
Counterfactual outcomes what if we had held something else?
Simulated continuations where can this plan go from here?
Resampled history is the edge robust to which periods happened?
Decisions are judged by their quality across all these worlds, not by the outcome of the one that happened.
FIGURE 2 History as one draw among many. DeOS generates the paths not taken at five points, each answering a different question about a decision.
That is a causal and counterfactual task. Pearl's ladder separates seeing, doing and imagining; each rung answers questions the rungs below cannot, no matter how much data they have. Every decision is an intervention on rung two, and every honest review of a decision is a counterfactual on rung three.
Pearl's ladder, and how DeOS climbs IT
IMAGININGIN DEOS
3
Would the deal have failed without the cost cuts? Skill or luck?
Counterfactual rows: what every alternative would have earned in the periods that did happen; alternative histories from the world model
DOINGIN DEOS
2
What happens to margin if we raise prices 5%?
Interventional data from the experiments the system chooses itself; a world model that predicts the outcome and the next state of an action
SEEINGIN DEOS
1
What do returns look like when rates rise?
In-context learning over observed rows and retrieval of similar situations. Where language models stop.
FIGURE 3 Pearl's ladder of causation with business questions at each rung, and the mechanisms DeOS uses to reach it. Language models stay on the first rung.
High-stakes decisions are computationally infeasible for humans alone and logically ungrounded for AI alone. Hybrid Intelligence assigns each part of a decision to whoever is better at it, and makes the combination better than either.
HUMAN CONTRIBUTION
Judgment and context
Tacit domain judgment, common sense, risk tolerance, and qualitative evaluation of non-digitized "soft" factors.
Merges deep learning's statistical adaptability with symbolic logic's explicit rules and causal priors, so that hard constraints are enforced, not learned.
The AI provides context: what the data, the ontology and the world model say about the system. The human provides context: the goals, the constraints, the assumptions and the knowledge that is not in any database. Many of the data a decision needs do not exist and have to be created, either by eliciting judgment or by simulation. Human judgment enters as a column of the context, so the learner gives it exactly the weight that past outcomes support, and no weight once it stops predicting anything.
Humans set the decision constraints, goals, criteria and policy framework. Within that framework, each component of an analysis runs in the human / AI mode that fits it, and can move between modes as trust is earned. The decision itself always stays with the named decision owner.
MODE 1
Judgement only
The human analyses and decides; DeOS captures the judgment in a structured form, with its rationale, and supplies the data, context and documentation.
MODE 2
Augmented
DeOS diagnoses, simulates and recommends; the human challenges, adjusts and decides.
MODE 3
Automated
DeOS decides and acts within human-set constraints, goals and policy; humans supervise and can intervene.
Standard enterprise LLMs versus a Hybrid Intelligence Decision OS#
Core dimension
Standard enterprise LLMs
Hybrid intelligence decision OS
Primary function
Knowledge extraction and text generation
Probabilistic simulation and policy optimization
Reasoning model
Statistical token correlation
Causal structure, experiments and neurosymbolic logic
DeOS is decision optimization, not just a model or an AI. The platform has five capabilities. Governance spans every level: solutions, decision policies and analyses. Every decision runs through the other four, and the outcome of each decision flows back into all of them.
FIGURE 4 The DeOS architectural framework: the five platform capabilities. Governance spans every level; each decision passes through context, analysis, recommendation and learning; outcomes and the realities that did not happen feed back into the system.
CAPABILITY 1 · GOVERNANCE
Spans every level
Codifies the mandate as decision policies across solutions, decision policies and analyses. Constraints are enforced by the rule layer, never learned; every decision, recommendation, override and reason is recorded and auditable.
CAPABILITY 2 · CONTEXT
Context engineering and representation
Ingests multimodal data to construct domain ontologies and structural representations, capturing systemic dependencies rather than isolated tokens. Perception starts from the decision: which signals bear on it.
CAPABILITY 3 · ANALYSIS
Probabilistic prediction and fusion; world model and counterfactual simulation
Runs causal diagnosis and counterfactual prediction. Combines in-context learning with aggregated human conviction, models regime-aware probabilities and mechanisms that hold when conditions shift (regimes at four scopes at once: global, local, macro for the economic cycle and micro for the investable universe and the corporate lifecycle), and answers what will most likely, could and would happen: a calibrated forecast, the full range including tails, and the outcome under each alternative action.
CAPABILITY 4 · RECOMMENDATION
Policy search and decision optimization
Searches the space of actions for the one that best serves the goals across simulated futures, within hard constraints, with trade-offs made explicit: Selection, policy optimization or Go / No-Go, as the decision policy requires.
CAPABILITY 5 · LEARNING
Continuous learning and metacognition
Tracks real-world outcomes to update priors, calibration and trust, while analyzing the accuracy of human judgment. The system learns not only from what it did, but from everything it did not do and from the realities that did not happen.
Enterprise decisions do not run on text alone. They run on financial statements, time series, panels of firms over time, call recordings, presentations, scanned documents and networks of ownership and supply. DeOS brings all of them into one representation and one learner.
FIGURE 5 Input data of any modality is structured by an ontology into a context graph, which feeds the ICL in-context learner. The learner solves seventeen task types; these task types are stacked into analyses, combined with human input where it adds value, and delivered as purpose-built, end-to-end decision workflows. Examples under each data type are illustrative.
ONTOLOGY
Structure before statistics Entities, relations and causal structure are made explicit, so the learner sees systemic dependencies, not isolated tokens or columns.
IN-CONTEXT LEARNING
Fitting is conditioning The examples become the model's context; one forward pass returns a full predictive distribution. No weights change.
STACKING
Task types become analyses The output of one task becomes the context of the next: extraction feeds representation, which feeds forecasting and finally optimization.
A classical model is trained per task and per dataset. It waits until enough data exists, and after a regime change it waits again for the next retraining. The DeOS learner is pretrained once across a very wide range of synthetic and real problems. On a new task it does not train; it reads.
FIGURE 6 Zero-shot, in-context and retrieved-memory use of the same pretrained model. Lower panel: how error recovers after a regime change for an in-context learner and for a model trained per task. Schematic.
ZERO-SHOT
Useful on day one
A new market, product or deal type gets a first answer before any task data exists, with uncertainty that honestly reflects how little is known.
IN-CONTEXT
Small data, no overfitting
A handful of rows is enough to condition on. The pretrained prior does the regularizing that small samples cannot do for themselves.
MEMORY
Experience that transfers
The most similar past decisions are retrieved and read first, within a task and across tasks that share a structure.
Because adaptation is a change in what the model reads, it happens at the speed of new evidence. The same property makes every improvement to the foundation model reach every analysis at once, which is what makes the architecture plug and play.
A single question such as "should we acquire this target, and on what terms?" is not answered by one prediction. It is a pipeline. DeOS breaks it into the steps of the decision lifecycle and applies the same in-context learner many times along the way, each time to a different task, on a different modality, with a different context.
FIGURE 7 Anatomy of one analysis. Fifteen applications of one pretrained learner across five pipeline steps, all reading from and writing to a shared context. The tasks, modalities and their order are an illustrative example for a buyout decision.
CONTEXT FLOWS FORWARD
Each step reads the last
Extracted figures become the rows the anomaly check reads; peer clusters become the context of the forecast; forecast distributions become the inputs of the optimizer. Uncertainty is passed on, never dropped.
OUTCOMES FLOW BACK
Each cycle starts smarter
What actually happened is stored with the context of every step, so the next analysis of a similar case reads the full record of how the last one turned out.
Seen from above, an analysis is a path through a grid of modalities and tasks. Because every cell is served by the same learner, a new path costs a new configuration, not a new model. The path is then repeated along the other dimensions of a decision: from product to portfolio, from days to years, and from judgement only to automated.
FIGURE 8 The analysis of Figure 7 drawn as a path through modality × task. The same path runs at several tiers, horizons and human / AI modes, and results at one scale become context at the next. Illustrative.
SCALES
Inside out Product, business unit, firm, sector, asset class and portfolio. A decision at one scale stays consistent with the scales around it.
HORIZONS
Several clocks Daily stress, monthly cycles and multi- year value creation run as separate paths that inform each other.
END TO END
From data room to decision Inputs, every intermediate result, the human assumptions and the recommendation sit in one governed workflow.
AI models improve every few months. A decision system that has to be rebuilt each time cannot keep up, and one that is locked to a single model cannot use the best one. DeOS separates its components by contract: what goes into a slot and what comes out is fixed, the model inside it is not.
Modular by contract: each slot can be upgraded on its own
INPUT ENCODERS Text · audio · video · image · relational raw data in, embedding columns out O NTO LO GY AN D C O NTEXT Entities, relations, levels, memory retrieval
SWAP 1 A stronger encoder for one modality Replace only that encoder, for example speech or vision. Nothing downstream needs to change.
context graph in, context rows out ICL FOUNDATION MODEL Pretrained in-context learner rows in, full distribution out
SWAP 2 A better foundation model is released Swap one slot. Every analysis built on it improves, and the pipelines, workflows and audit trail stay as they are.
WORLD MODEL AND SIMULATOR Any model with the simulator contract state and action in, next state out OPTIMIZER AND ACQUISITION Search, value of information, try/act/wait distributions in, actions out RULES AND HUMAN INPUT Mandate, constraints, views, assumptions never learned, always enforced
SWAP 3 A custom model trained for a client A proprietary model plugs into the same contract, next to or instead of the standard one. TRUST GATE FOR EVERY PLUG-IN A new part earns its place before it decides Each model is scored on results it had not seen. It takes over only while that record holds up, and hands control back when it fails.
FIGURE 9 The DeOS components as slots with fixed contracts. A better foundation model, a stronger encoder for one modality or a model trained for one client replaces only its slot; every new part passes a trust gate before it decides. Slot layout simplified.
MODEL UPDATES
Single parts change
When a component improves, only that component is replaced. Pipelines, workflows, memory and audit trail remain untouched.
Everything earns its place. A plug-in, a simulator or a learned prior takes over only after it has agreed with outcomes it had not seen, and hands control back when that record breaks. Upgrades are reversible by design.
BEST OF BREED
Better plug-ins drop in
If a stronger speech, vision or time-series model appears, it plugs into the encoder slot and every analysis reading that modality gains.
CUSTOM MODELS
Your model, our system
Models trained on proprietary data or for a specific domain plug into the same contract, next to or instead of the standard ones.
FIGURE 10 The ontology view in DeOS: a data room organized into a context graph of entities such as debt, taxation, revenue model, customer segments, R&D and shareholder equity, and the relations between them.
DeOS takes the task types, stacks them together in analyses, defines where the human and where the AI is involved, and assembles them into a purpose-built, end-to-end workflow for each decision. Data rooms, scans and expert predictions feed AI functions such as capital allocation, expert assumption aggregation and assessment agents. Each workflow stores its data, assumptions, model outputs, human inputs and recommendation, so every decision can be reconstructed and reviewed.
FIGURE 11 The DeOS workflow canvas. Scans, data rooms and expert predictions flow into chained AI functions for capital allocation, expert assumptions and assessment: the pipelines of Figure 7 as users build them.
Hybrid Intelligence analysis synthesizes tacit human judgment with enterprise-grade AI foundation models to deliver calibrated, context-aware predictions. People hold soft knowledge no database has: management quality, intent, context, what is new. Their judgments are biased and noisy, so DeOS elicits them in structure, aggregates them, measures each person's track record and debiases them, rather than ignoring or blindly following them.
FIGURE 12 Expert assessment in DeOS. An expert estimates the minimum, expected and maximum long-term capex-to-depreciation target for a buyout decision policy and states their confidence; the DeOS Assistant supports with web search, deep research and an ecosystem data scan. The user name in the assistant greeting has been removed.
AS EVIDENCE
A column of the context
Scores, notes, memos and call transcripts enter as context, weighted by what past outcomes support.
AS A VIEW
A shape on the distribution
A stated view on returns, risk or tails reshapes the scenario distribution with an explicit confidence.
AS A RULE
A boundary
Mandates and constraints are enforced exactly and never traded off by the learner.
Feature AREA
Capability
Strategic value
Dynamic trust calibration
Weighs model confidence against contextual human
Fewer failures in volatile or novel
judgment, per regime
environments
No-code pipelines
Turnkey data processing, automated feature
Advanced analytics for non-technical users
engineering and model aggregation
Feature AREA capability
Strategic value
Contextual optimization Selection, policy optimization and Go / No-Go, bound by goals and constraints
Clear, objective recommendations for complex allocations
The engine · Deosalpha, the decision ICL engine
15. An ICL optimizer with a memory and a world model#
DeOSAlpha is the decision engine at the core of DeOS. Its technical name is the Decision ICL Engine, and this section uses both. It chooses actions under uncertainty without training a model. Its core is an active learner, a Bayesian-optimization engine whose surrogate is the pretrained tabular in-context-learning foundation model rather than a Gaussian process. Fitting the surrogate is conditioning: the observed rows of action, context and reward become the model's context, and one forward pass returns the full predictive distribution, read as quantiles with parametric tails, for any candidate action. Acquisition criteria are integrated over that quantile function, a trust region and gradient ascent through the foundation model find the best candidate, and a rule layer applies what is known exactly. The Decision ICL Engine is the same learner given a memory of every decision with its context and a sense of time: each suggestion is built from the stored decisions most similar to the present one, in sequential problems the learner is fitted on each step's return-to-go, and the same model, run backwards, becomes a return-conditioned policy. A world model lets it look ahead, but never counts as evidence. An EVI engine decides whether to try, act or wait.
FIGURE 13 DeOSAlpha, the Decision ICL Engine. Goals and the perceived context select the most similar past decisions from memory; the surrogate, a pretrained in-context-learning foundation model, returns the full predictive distribution for any candidate in one forward pass; the acquisition criteria and the trust region search it; the rule layer applies what is known exactly; the EVI engine decides whether to try, act or wait; realized outcomes return to memory. The world model steers the search but never counts as evidence. The monitoring layer scores every model on results it had not seen and withdraws its right to decide when the record fails. Simplified.
FIGURE 14 The same pretrained model serves as value model, as return-conditioned policy and as simulator, depending on which cells of a decision row it is asked to fill. Simplified.
SEQUENTIAL
Decisions about the rest of the episode
In sequential problems the learner is fitted on each step's return-to-go, the discounted sum of what followed, so a choice is a decision about the rest of the episode, not the next step at the expense of the ones after. The same model, asked for the action that went with a high return in a state like this one, is a return-conditioned policy without training.
MULTI-OBJECTIVE
The Pareto front, per context
Return, risk and cost are explored as trade-offs on a front built for the current context, with a stated focus (the knee, the widest gap, fixed weights or a rule of your own), by scalarization (ParEGO) instead of being collapsed into one number in advance.
RISK
Pessimistic about noise, optimistic about ignorance
The DRO-CVaR criterion penalizes the tail that is owed to noise, rewards the spread that one more observation could remove with an exploration bonus, and adds a Wasserstein ambiguity penalty that guards against the model itself.
THE ENGINE · WHAT IT CAN DO THAT A SPREADSHEET OR A LANGUAGE MODEL CANNOT
The Decision ICL Engine rests on eight principles. Each is stated here as what the engine does that an analyst with a spreadsheet or a language model cannot. They are design properties: the Factor paper on the engine states that the mechanics are in place and that an edge has not yet been shown in a walk-forward test.
01
The surrogate is a posterior, not a regression
Every candidate action gets the full predictive distribution in one forward pass: quantiles, tails and spread, never a point prediction. A spreadsheet carries one number per cell and a language model one sentence; neither can say how wrong it might be, in the units of the objective.
02
No distributional shortcut
Expected improvement, max-value entropy search and CVaR are integrated over the predicted quantile function, so a skewed or heavy-tailed outcome is scored as predicted. A bell curve in a spreadsheet calls a five-sigma day impossible; the engine prices it.
03
Adapt by reading, not by training
A new regime is handled by retrieving the decisions taken in similar situations and conditioning on them. No weights change and nothing waits for retraining. An analyst re-estimates a model over weeks; a language model cannot update from your outcomes at all.
04
What is known exactly is never learned
Forbidden actions, the move limit and the cost of change are rules in the rule layer, applied exactly; only rewards and outcome constraints are learned. A language model can be told a mandate and still breach it; here a breach cannot be proposed.
05
Risk aversion is owed to noise, optimism to ignorance
Uncertainty is split into what the world does on its own and what one more observation would change, and each is routed where it belongs: the aleatoric part to sizing and cautious quantiles, the epistemic part to the value of a trial. A spreadsheet has one error bar; a language model has none.
06
Imagination is never evidence
World models, simulators, priors and the engine's own plans say where to look; only realized outcomes enter the surrogate, and every imagined source is used only while its out-of-sample record holds. A scenario on a slide, or a story from a language model, is easily mistaken for data. Here it cannot be.
07
One currency
The value of one more trial, the gain from acting and the cost of change are compared in the units of the objective, and "wait" is always an option. Neither a spreadsheet nor a language model can tell you whether finding out is worth more than acting now, or whether acting pays for the cost of changing course.
08
Everything earns its place
The outcome model, the sequential policy, the simulator and the policy prior each take over only after a stated test of agreement with reality on results they had not seen, and hand back when it fails. A model in a spreadsheet is trusted because it was built; a model here is trusted because it has been right out of sample.
Nothing in the engine is trained on your data. Adaptation is a change in what the model reads: the decision memory, the retrieved context, the rules and the prices. That is why a new market, a new mandate or a new regime is handled as soon as similar rows exist, and why every change is visible in a report.
The engine · The world model and the paths not taken
In a new market or a new regime there may be almost no history at all. Much of the data a decision needs does not exist and has to be created. The Decision ICL Engine uses a world model, or any simulator with the same contract, to answer one question: given the context, the action and the goal, what outcome follows and what does the context become? The agent loop built on it, in the Factor paper the Decision ICL Agent, runs one decision step as perceive, dream, decide, act and learn.
01World model as an ensemble
A bootstrap ensemble learns the dynamics from any modality. The spread of the members' means is what the model does not know; the residual spread is what the world does on its own.
02Dreaming: imagined trajectories
Once the world model has agreed with reality, a dream study rolls out trajectories, including scenarios outside the data, and proposes the best dreamed action; try, act or wait may still say "wait".
03Planning by model-predictive
control A suggestion is the first action of the best plan several steps ahead, valued on its mean return and the CVaR of its worst rollouts, only while the simulator passes its trust gate.
The paths not taken are generated at five points: synthetic periods (does this only look good on the one history that happened?), shadow decisions on past dates, hindsight rows (what would every other candidate have earned on the same periods?), rollouts from the world-model ensemble and a block bootstrap of the window (is the edge robust to which periods happened?).
FIGURE 15 Imagined futures are filtered for novelty, learnability and usefulness. They steer the search and the stress tests; only real outcomes become evidence. Criteria after Schmidhuber (2010) and Hughes et al. (2024).
Two memories feed the surrogate, and a rule layer bounds what the search may propose. In the code the rule layer is the neurosymbolic controller; this paper uses the Factor paper's name.
LONG-TERM · DECISION MEMORY
Every decision with its context, and retrieval
Stores every decision with the context it was made in, its reward or return-to-go and what followed; consolidates it, shares it across related tasks and retrieves the decisions most similar to the one at hand, on which each suggestion is fitted.
RULE LAYER
Forbidden actions and the move limit
Mandate limits and expert rules are applied exactly and never learned: a forbidden action is never proposed, and the move limit is a hard box that no evidence relaxes.
RULE LAYER
Cost of change
Known costs are subtracted inside the predictive distribution, exactly, so the surrogate never learns what is known in advance and "no change" always competes with every candidate at zero cost.
SHORT-TERM · CONTEXT WINDOW
What the foundation model reads
Holds the rows the surrogate conditions on: the most recent, the best, the nearest contexts and one representative of every explored region, so a region found poor is not silently forgotten.
SEARCH
Trust region
A box around the incumbent whose radius moves on observed outcomes: it grows after successes, halves after failures and restarts when it collapses.
The expected value of information (EVI) gives the Decision ICL Engine a metacognitive ability: an estimate of what it does not yet know and whether finding out, or acting, is worth the cost. The value of one more trial, counting only the reducible share of the uncertainty, the gain from acting net of the cost of change, and "wait" at zero are compared in the units of the objective. The controller evaluates trials while the advice is "try", up to a budget, then reports "act" with the parameters or "wait" with the reason.
FIGURE 16 The try, act or wait rule of the EVI engine. Only the reducible share of uncertainty makes another trial worth anything; the gain from acting is net of the cost of change; waiting is always available at zero. Bar lengths are illustrative.
IN THE SEARCH
Is another evaluation worth it?
The search stops when the next trial no longer pays, rather than after a fixed number of trials.
IN EXECUTION
Act or wait?
The same rule decides whether executing now adds enough value to justify transaction costs, slippage and predictive uncertainty, and otherwise holds the current position; in a Go / No-Go policy it is the conviction score with its price attached.
μ^
Expected returns ( ) define where potential alpha exists, from the q0.50 EBM. Tail-risk weight (λtail) controls how strongly downside tail risk, from the q0.05 EBM, is penalized when selecting candidate weights.
μ^
Decision rule: try while the reducible value of one more trial exceeds its cost; act if E[gain | , q0.05, λtail] − cost of change > 0; otherwise hold or wait.
Regimes are estimated at four scopes at once: global (the world), local (a market, region or sector), macro (the economic cycle) and micro (the investable universe and each company's corporate lifecycle). Even in a regime seen for only a few days the engine adapts in context, and decides on its own whether to continue or to stop and observe.
Self-calibration
Stated uncertainty is checked against realized errors on unseen results and rescaled, so the engine's confidence stays honest over time.
Drift detection
A rise in surprise relative to the engine's own history flags a changed world; the stale past is dropped and trust is re-earned.
Benchmarking
Classical action-reward problems first, then portfolio problems with full context, walk-forward and out of sample, with a no-learning control. The mechanics are in place; an edge has not yet been shown.
A decision engine that cannot tell what it does not know will act on noise, trust a model because it was built, and keep going when the world has changed. The Decision ICL Engine keeps a record of its own reliability and lets that record, not the model's confidence, decide who decides. Four mechanisms carry it; the split of uncertainty (Figure 17) routes what they find.
FIGURE 17 The routing rule of The world that didn't happen, hedge chance, investigate ignorance, step back from novelty, implemented almost literally in the engine: the reducible share of the value of information drives "try", the gain net of cost drives "act", and novelty or a failed trust gate pushes the choice back to "wait" or to the simpler decider.
FIGURE 18 Every imagined or learned source passes the same kind of gate: the simulators, the policy prior, the planner's own search, the outcome model, the sequential policy and the agent's world model. The right to propose is cheap, because a proposal is only a candidate; the right to decide is earned on agreement with reality on results the model had not seen, and withdrawn when the record fails. Where the gates bind is a human choice, the burden of proof before anything learned may decide; it is the one assumption the engine cannot audit itself, so it is fixed in advance and measured on the holdout.
An AI that decides under uncertainty needs fifteen capabilities in five groups. DeOS was designed against this list; the Factor paper on the Decision ICL Engine maps each to its mechanism: nine met, six partly. The mechanics are in place; an edge has not been shown, and every status is a design property until the walk-forward test measures it.
FIGURE 19 The fifteen capabilities from The world that didn't happen, with the mechanism in the Decision ICL Engine that addresses each and the status the Factor paper assigns; a status describes the design, not a demonstrated edge.
Partly met
What the engine does
What is missing
02 Causal and counterfactual
Interventional data from its own experiments; hindsight rows
Causal estimators exist but are not yet wired in
03 Partial observability
A perceived context with its reliability and surprise
No belief state over hidden variables
07 Three uncertainties
Chance, ignorance, state and model uncertainty apart
Mechanism uncertainty is not modeled
10 Explainable
Advice with its quantities; who decided and why
Not yet traced to causes
13 Self-improving
Learns from outcomes and its own search, auditable
Does not change its own structure
15 Imaginative
Futures and histories at five points, filtered
Evidence only when labeled and checked
Where the legacy toolkit falls short, and what DeOS does instead#
Legacy paradigm
What IT assumes
What DeOS does instead
Mean-variance
Up and down moves count alike
Tail-aware goals, cautious quantiles and robust CVaR
Normal distribution
Extreme moves are vanishingly rare
Heavy-tailed predictive distributions, scored as predicted
Labels and buckets
Different names behave differently
Candidates described by what they do and the drivers beneath them
Bull or bear
One regime describes the situation
Several horizons and regime-aware trust
Three scenarios
The future fits on one slide
Full distributions, synthetic periods and rollouts
Correlation
Past co-movement reveals cause
Interventional data and counterfactual outcomes
Numbers only
What matters is in structured data
Eight modalities and human judgment as context
Static models
The fitted past guides the future
In-context adaptation, drift detection, forgetting on drift
DeOS is end-to-end governance of data, analysis and recommendation, spanning every level, from the solution through its decision policies to each analysis, across the investment decision lifecycle, with all the analysis behind every decision recorded and continuously learning.
Mandate
Constraints are enforced by the neurosymbolic core, not by instruction.
Auditability
Data, assumptions, model outputs and human inputs stored for every decision.
Transparency
Every recommendation returns the quantities, drivers and trade-offs behind it.
Human control
Humans steer the controller, the world generator, what the model sees and the context.
This is version 2 of the technology whitepaper. It adds the sections on uncertainty and the paths not taken, zero-shot and in-context learning, the analysis pipeline, stacking along dimensions, plug and play, the Decision ICL Engine, imagination, try/act/wait and the fifteen capabilities. Version 1 is unchanged.
New content draws on the vencortex papers The world that didn't happen and ICL Optimization and the Decision ICL Engine. Implementation details, parameter values and internal names from the second paper were deliberately left out.
The ICL analysis states that the mechanics are in place but a performance edge has not yet been demonstrated, and rates nine of the fifteen capabilities as met and six as partly met (causal and counterfactual, partial observability, three uncertainties, explainable, self-improving, imaginative). Figure 19 describes the design behind each capability; it does not claim each is complete. This paper makes no performance claims.
Figures 7 and 8 (the buyout pipeline, its tasks, modalities and order), Figure 9 (slot layout and swap examples), the regime-change chart in Figure 6, the bars in Figure 16 and the paths in Figure 2 are illustrative. Zero-shot use relies on the pretrained prior, shared memory of related tasks and expert assumptions; this framing is an interpretation to confirm with the product team.
Carried over from version 1: the EVI decision line is a simplified reading; "EBM" is used as in the notes and not expanded; the benchmarking card interprets the note on classical action-reward and portfolio benchmarks; the user name was removed from Figure 12.
This notes page is for internal review and should be removed before external publication.
Causation · Pearl, J. and Mackenzie, D. (2018). The Book of Why. Basic Books.
Uncertainty types · Hüllermeier, E. and Waegeman, W. (2021). Aleatoric and epistemic uncertainty in machine learning. Machine Learning, 110.
Black swans · Taleb, N. N. (2007). The Black Swan. Random House.
Decision intelligence · Pratt, L. (2019). Link: How Decision Intelligence Connects Data, Actions, and Outcomes for a Better World. Emerald.
Tabular in-context learning · Hollmann, N. et al. (2025). Accurate predictions on small data with a tabular foundation model. Nature, 637.
Return-conditioned policies · Chen, L. et al. (2021). Decision Transformer: reinforcement learning via sequence modeling. NeurIPS 2021.
Open-ended imagination · Hughes, E. et al. (2024). Open-endedness is essential for artificial superhuman intelligence.
ICML 2024. Schmidhuber, J. (2010). Formal theory of creativity, fun, and intrinsic motivation.
IEEE tamd, 2(3).
ParEGO
Knowles, J. (2006). ParEGO. IEEE Transactions on Evolutionary Computation, 10(1).
METHOD
CVaR optimization
Value of information
World models
Structural causal models and the do-operator
In-context learning as Bayesian inference
Limits of language models for reasoning and planning
REFERENCE
Rockafellar, R. T. and Uryasev, S. (2000). Optimization of conditional value-at-risk. Journal of Risk, 2(3).
Howard, R. A. (1966). Information value theory. IEEE Transactions on Systems Science and Cybernetics, 2(1).
Ha, D. and Schmidhuber, J. (2018). World models. arXiv:1803.10122.
Pearl, J. (2009). Causality: Models, Reasoning, and Inference, 2nd ed. Cambridge University Press.
Müller, S., Hollmann, N., Pineda Arango, S., Grabocka, J. and Hutter, F. (2022). Transformers can do Bayesian inference. ICLR 2022.
Mirzadeh, I. et al. (2025). GSM-Symbolic. ICLR 2025. Shojaee, P. et al. (2025). The illusion of thinking. arXiv:2506.06941. Kambhampati, S. et al. (2024). LLMs can't plan, but can help
planning in LLM-Modulo frameworks. ICML 2024. Jin, Z. et al. (2024). Can large language models infer causation from correlation? ICLR 2024.
Faithfulness, calibration and look-ahead · Turpin, M. et al. (2023). Language models don't always say what they think. NeurIPS 2023. Kalai, A. T. et al. (2025). Why language models hallucinate. arXiv:2509.04664. Glasserman, P.
ahead
Kalai, A. T. et al. (2025). Why language models hallucinate. arXiv:2509.04664. Glasserman, P. and Lin, C. (2023). Assessing look-ahead bias in stock return predictions generated by GPT sentiment analysis. arXiv:2309.17322.
References describe the general methods, not the DeOS implementation.
Decision optimization with DeOS · vencortex® – DeOS | Decision Optimization System. Converted from DeOS_Compendium_Light.pdf by scripts/pdf_to_markdown.py; built by scripts/build_book_site.py.