Chapter 07 · Part III · The system: Platform and technology
The DeOS platform
Part III · The system: Platform and technology
Part III
1. The system Platform and technology#
Part III is the system that runs the framework: the five capabilities of the DeOS platform, how it is deployed and secured, how a decision runs through it, and the hybrid intelligence technology underneath, built for business data and decisions rather than for text.
The technology in detail: built for decisions, not for text.
Whitepaper · The platform
2. The DeOS platform#
Your infrastructure, not someone else's: policies, data, models and decision memory on a dedicated instance you control
Why DeOS is built as infrastructure, the five capabilities of the end-to-end decision cycle, how it is deployed and secured, who works in it, and one decision followed through the platform in fourteen steps.
| 01 · infrastructure | 02 · intelligence | 03 · learning |
|---|---|---|
| Yours, not shared | Five capabilities | For you alone |
| A dedicated instance, in the cloud of your choice or on premises. | Governance, Context, Analysis, Recommendation, Learning, end to end. | Every decision sharpens the next; nothing shared with another client. |
Summary
3. In brief#
Investment firms already rent their data from terminals and are starting to rent their intelligence from hyperscalers. Both give every customer the same thing. DeOS is different in kind: it is infrastructure the firm owns and controls, dedicated to its decisions, and it learns for that firm alone.
Your infrastructure. DeOS runs as your firm's decision infrastructure on a dedicated enterprise instance, under your keys and access rules. Policies, context, models, every insight created and the decision memory belong to you.
Your infrastructure.
DeOS runs as your firm's decision infrastructure on a dedicated enterprise instance, under your keys and access rules. Policies, context, models, every insight created and the decision memory belong to you.
Not a generic tool, and not a data terminal.
Generic tools answer questions and terminals deliver the same data to every subscriber. DeOS runs decisions: ranked, sized and explained actions built on your thesis, your criteria and your data.
Five capabilities run the end-to-end decision cycle.
Governance spans every level with strategy-aligned, auditable policies. Context contextualizes data, Analysis fuses causal diagnosis and counterfactual prediction, Recommendation returns one explained course of action, and Learning builds the decision memory.
A foundation to build on.
Teams create and extend their own decision use cases in the no-code DeOS Studio, from standardized, reusable decision assets, so each use case starts from the ones before.
Your instance, your data, your models.
Fully managed on the cloud provider of your choice, in your own cloud account or on premises. Encrypted in transit and at rest, proprietary models only, nothing sent to third-party AI providers, fine-grained access control for users, teams and funds.
Learning that stays yours.
Every model learns from your decisions on your instance only. DeOS never trains general models on your data and never shares any data or learning with other clients.
Four roles, each where it adds the most.
The decision architect sets policies in DeOS Studio, the decision maker decides on an explained recommendation, the recommender owns and can override it with a recorded reason, and the analyst builds and challenges the context.
One decision through DeOS, in fourteen steps.
A reusable demo framework follows one decision from policy to the next decision under the five capabilities, with the role that leads each step. A concrete storyline, a new position in a mid-cap industrial company, runs through it.
01
The platform
4. Your infrastructure, not someone else's#
Decision optimization only creates an edge if the edge stays with you. DeOS runs as your decision infrastructure: your policies, your data, your models and your decision memory.
Terminals and hyperscaler agents are useful, and DeOS connects to both: terminal data is one of its context sources, and foundation models are components inside it. What neither can be is the place where the firm's own edge lives. That place has to be infrastructure the firm controls.
Five things that make it infrastructure#
02
Capabilities of DeOS
5. The end-to-end decision cycle#
DeOS governs strategy-aligned and auditable policies across every level, from the solution through its decision policies to each analysis and its analysis archetypes. It contextualizes data with decision-centric data management, and analyses with expert predictions and AI foundation models stacked together in one end-to-end analysis. It recommends one explained course of action, and it learns from every decision made and not made, and from every piece of information flowing in.
| Capability | Question IT answers | What IT holds | What IT gives you |
|---|---|---|---|
| Governance | How do we decide, and who decides what? Spans solutions, decision policies and analyses. | Decision policies, causal decision graphs, guardrails and constraints, audit trails, versions, decision assets, DeOS Studio | The same decision run the same way every time, in line with the mandate, and fully auditable |
| Context | What is the situation, really? | Data connectors to proprietary, internal and third-party sources, decision-centric ontology, similar-situation scans, generated and counterfactual context | A decision-centric picture of reality, built once and reused |
| Analysis | What drives outcomes, and what will, could and would happen? | Expert predictions, DeOS foundation models, hybrid fusion by regime, end-to-end analysis archetypes from context to prediction | Calibrated distributions instead of point estimates, from one click |
| Recommendation | What should we do about it? | Recommended actions, expected outcomes, conviction, ranked options, aggregation across levels, why chains per perspective, insights | Ranked, sized and explained actions, down to the data point |
| Learning | What did we learn, and from what? | Point-in-time decision memory, outcomes, counterfactuals, analyst calibration | A process that improves with every decision, for you alone |
03
Capability 1 · Governance
6. Governance: your edge, codified#
Governance spans every level: the solution, its decision policies, their analyses and the analysis archetypes beneath. It turns the mandate and the strategy into auditable decision policies that every decision runs through, defines which analysis is needed, makes it reusable, and makes sure it is always run the same way and that every result can be traced back.
Policies
One decision, formalized: its objective, owner, regulatory constraints, recommendation and analyses. Each policy has one
POLICIES
Decision policies
One decision, formalized: its objective, owner, regulatory constraints, recommendation and analyses. Each policy has one of four decision policy types, Selection, Policy optimization, Go / No-Go or Monitoring, and one of eleven decision archetypes, from Selection and Conviction test to Timing, Allocation, Hedging and Root cause. Policy type plus decision
archetypes, from Selection and Conviction test to Timing, Allocation, Hedging and Root cause. Policy type plus decision archetype is a policy archetype. A policy says what analysis the decision needs, so two analysts on two funds reach comparable recommendations.
Causal decision graphs
which thresholds a recommendation has to clear.
Control
CONTROL
Auditable processes, guardrails and constraints
Who prepares, recommends and decides, in Automated, Augmented or Judgement only mode; which steps are mandatory; which limits apply, for example concentration, liquidity, leverage or exclusion rules. A decision cannot be submitted while a required step is open or a constraint is breached. Every run is reproducible.
Who prepares, recommends and decides, in Automated, Augmented or Judgement only mode; which steps are
Audit
AUDIT
Audit trails
Every insight carries its provenance: which insight was created on what data, by which AI analysis and by which human, and when. Votes, approvals and overrides are recorded with their reason. The firm can show, for any recommendation, how it came about.
Every insight carries its provenance: which insight was created on what data, by which AI analysis and by which human,
Versions
VERSIONS
Versioning
Policies, analyses, data snapshots and recommendations are versioned point in time. When new information arrives, a new version is created and the earlier ones are kept, so the firm sees how and why a view changed, and no later retraining can look ahead.
Policies, analyses, data snapshots and recommendations are versioned point in time. When new information arrives, a new
As s e TS
The predefined, standardized and reusable components a decision is built from: policy templates, data ontologies, metrics,
ASSETS
Decision assets
The predefined, standardized and reusable components a decision is built from: policy templates, data ontologies, metrics, criteria libraries, causal decision graphs, expert prediction questions, data connectors and prebuilt analyses. Built once, used across policies, funds and teams, so the same decision runs the same way every time and every new use case starts
used across policies, funds and teams, so the same decision runs the same way every time and every new use case starts from what already exists.
Studio
STUDIO
DeOS Studio
The no-code studio in which the decision architect creates, adjusts and extends decision policies and solutions. Data connectors, scans, expert prediction questions and AI analyses are wired into a policy on a visual canvas from the decision assets (see step 01 of the demo framework).
The no-code studio in which the decision architect creates, adjusts and extends decision policies and solutions. Data
Why it matters. A prompt is not a policy. Generic agents answer the same question differently on different days. A decision policy runs the same analysis every time, so differences in recommendations come from differences in the facts, not from the tool or the person, and every difference can be audited.
04
Capability 2 · Context
7. Context: connecting the dots#
A data warehouse built for decisions. Context engineering connects your proprietary data, internal systems and third-party providers and turns fragmented, multi-source data into a decision-centric representation of reality, instead of evaluating data in isolation from the environment in which choices unfold.
Data connectors
document and deal rooms, and custom collection scripts, integrated into one data fabric inside your instance.
Connectors to third-party providers. Market data providers, filings and transcripts, and alternative data, licensed by you and used only in your instance.
One ecosystem, any provider. The vencortex ecosystem draws on third-party sources and keeps them synchronised, harmonised and aligned across providers: the same company, field and date mean the same thing whichever provider delivered them. Keep the provider you prefer, switch, or combine several; every analysis runs unchanged.
One processing pipeline. Structured and unstructured formats, including financial tables, transcripts and audio, standardized once. Connectors are decision assets: built once, reused by every policy.
2 · MODELING
Decision-centric ontology and business mechanics
Physical object and mechanics modeling. Digital representations of physical assets, corporate structures, supply chains and market participants, reflecting the business as it actually operates.
and business mechanics
Action-to-outcome mapping. Static financial statements connected to real operational levers, showing how external variables and internal choices ripple through the business.
Automated feature engineering. Spatial, cross-sectional and temporal metrics built dynamically as high-signal, context-aware inputs for the models downstream.
Multi-dimensional context
scan
scan
By situation: past regimes at macro scope (economic cycle) and micro scope (investable universe and corporate lifecycle), and liquidity conditions.
By context and subset: targeted slices of history that share the same micro-environment.
By evolution and behaviour: historical pathways and how similar entities reacted to past shocks or dislocations.
By causal drivers and characteristics: underlying operating mechanics and traits rather than surface correlations.
Predictive context
generation
incomplete, restricted or unobservable.
generation
Counterfactual modeling: what has not happened. Simulated scenarios and stress tests, including low-sample black-swan events, evaluate non-linear risks before capital is committed.
it is reused, the firm's data estate grows with every decision instead of being rebuilt in every spreadsheet.
05
Capability 3 · Analysis
8. Analysis: causal diagnosis and counterfactual prediction#
The analysis capability runs the intelligence steps Perceive, Diagnose, Predict and Recommend through their analysis components: perceive context, causal diagnosis (what drives outcomes), counterfactual prediction (what will, could and would happen) and contextual optimization. Hybrid intelligence combines tacit human judgment with the DeOS enterprise AI foundation models to deliver calibrated, context-aware predictions, and end-to-end analysis archetypes run the whole chain from context to prediction with one click.
1 · Human
Human expertise is essential for contextual signals that models miss.
1 · HUMAN
Human prediction and judgment extraction
Human expertise is essential for contextual signals that models miss. Expert prediction questions. Structured questions with ranges and confidence, grounded in base rates, capture intuition, metacognition and conviction across teams and reduce individual bias in aggregation.
Contextual guardrails. Human inputs formalized around the core decision variables: target definitions, time-to-target horizons, cost structures, risk preferences and operational constraints.
2 · AI
2·AI
DeOS enterprise AI foundation models
Proprietary foundation models built for business data beyond text, run inside your instance and learning for you alone. Regime-aware and contextual. Every prediction is conditioned on the regime at four scopes: global, local, macro (economic cycle) and micro (investable universe and corporate lifecycle), and on the context of the decision.
Proprietary foundation models built for business data beyond text, run inside your instance and learning for you alone.
Causal and probabilistic. Causal treatment effects, relative performance and structural mispricing beyond correlation; counterfactual simulation and stress tests; calibrated distributions rather than point estimates.
Multi-task and multi-modal. One model family for forecasting, classification, survival and event-time modeling, on tabular, time series, event, text and audio data, so filings, transcripts and interviews become quantitative inputs next to the numbers.
In-context learning without retraining. The models adapt to new data, new entities and new regimes in context, so the thesis adjusts as evidence arrives instead of waiting for an offline retraining cycle.
3 · Fusion
3 · FUSION
Hybrid intelligence fusion
A feedback loop that keeps learning when to weight human domain knowledge and when to weight the models, depending on the regime: more weight on the models in stable, data-rich regimes, more on the experts in novel situations. The output is one calibrated distribution per driver, with the contribution of each side visible.
A feedback loop that keeps learning when to weight human domain knowledge and when to weight the models,
The output is one calibrated distribution per driver, with the contribution of each side visible.
4 · END-TO-END
End-to-end analyses
One analysis archetype takes care of everything from context to prediction. It perceives the context, diagnoses the drivers, predicts them and propagates each prediction into the next model: drivers feed scenarios, scenarios feed valuation, valuation feeds sizing. The human is brought in at defined points, for the expert prediction questions and for the review of
valuation feeds sizing. The human is brought in at defined points, for the expert prediction questions and for the review of the fused distributions, so a user runs the whole chain with one click and still sees and shapes every step.
| Feature | Capability | Value |
|---|---|---|
| Dynamic trust calibration | Weighs model uncertainty and confidence against contextual human judgment, per regime. | Fewer failures in volatile or novel markets. |
| End-to-end analysis archetypes | Turnkey chains from data connectors to feature engineering, prediction and propagation, no code required. | Advanced analytics for non-technical users, with one click. |
| Contextual optimization | Selection, Policy optimization, Go / No-Go and Monitoring, bound by goals and constraints. | Clear, objective recommendations for complex decisions. |
06
Capability 4 · Recommendation
9. Recommendation: from analysis to action#
The recommendation capability turns the analyses and human inputs into one decision recommendation: a ranked, sized and explained course of action on the optimization goal, with a conviction level and a why chain that reaches down to the data point. It leaves the final decision with people.
Actions
ACTIONS
Recommended actions and expected outcomes
Recommended actions. The best course of action given the context, the goals and the constraints of the policy, for example a position size, a price, a treatment or a trigger, with ranked alternatives.
Recommended actions. The best course of action given the context, the goals and the constraints of the policy, for
expected outcomes
Probabilistic expected outcomes. For every option, the distribution of the outcome on the optimization goal, with percentiles rather than a point-estimate promise.
Conviction score. The probability that this is the right decision, given the mandate: the expected outcome on the optimization goal weighted by what matters to the mandate and adjusted for co-movement between the drivers.
Ranking of options. Options ranked by feasibility and net impact on the goal, bound by the constraints, so the trade-off between them is explicit.
Levels
LEVELS
Aggregation on multiple levels
The recommendation is read at the level that matters: asset and portfolio, position and fund, single decision and total portfolio. A position recommendation carries its effect on fund concentration, liquidity and risk budget, so the decision is judged by its effect on the whole rather than in a silo.
The recommendation is read at the level that matters: asset and portfolio, position and fund, single decision and total
levels
judged by its effect on the whole rather than in a silo.
Causal drivers and
Causal drivers and
Multi-perspective decision graphs. The recommendation seen through distinct perspectives, such as risk,
perspectives
attractiveness, growth and liquidity, to expose implicit trade-offs.
Conviction per perspective. Each perspective carries its own conviction level, so the committee sees where the case is strong and where it is weak.
EXPLAINABLE
A fully explainable why chain
Causal why chains. Every number unpacked to its root causes, combining feature attribution (what matters most) with feature behaviour (how changes move the outcome), down to the data points and the expert inputs behind it.
chain
Counterfactual and what-if analysis. The precise shifts that would change the recommendation, checked against historical prototypes and similar situations.
Impulse response modeling. How a shock to one input at time t ripples across connected metrics over time.
Human in the loop. The platform advises; transparent trade-offs leave full authority over execution with the decision makers.
Insights
Enterprise and market data streams are ingested continuously to produce thousands of insights across four dimensions,
Continuous insight engine
- DIMENSION · QUESTION ANSWERED
- Diagnostic · What happened, and why is it happening now?
- Predictive · What will happen, and where does consensus diverge from reality (misconception identification)?
- Counterfactual · What would happen if external conditions shift or levers are pulled?
- Prescriptive · What must happen, how likely is success, and what is the quantified impact?
07
Capability 5 · Learning
10. Learning: decision memory#
Learn not only from what you did, but also from everything you did not do, and from the realities that did not happen.
The learning capability is a continuous Bayesian updating engine that treats every thesis as a hypothesis under constant refinement. A complete, point-in-time decision memory takes in new evidence to refine the models, calibrate human intuition and keep the firm's memory whole. It runs on your instance only: what DeOS learns from your decisions is never shared with another client.
1 · MEMORY
Enterprise decision memory and auditability
Point-in-time versioning. Every decision is logged with the exact data snapshot, assumptions and market conditions available at that moment, which removes look-ahead bias from later learning.
and auditability
Standardized outcome tracking. Final choices, approval chains and real results are logged, building a proprietary longitudinal dataset for learning.
Governance and compliance. Full transparency on decision volumes, risk thresholds, overridden recommendations and historical success rates.
Hybrid learning and
metacognition
offline retraining cycle.
metacognition
Heterogeneous knowledge extraction. Analyst opinions, assumptions and conviction levels aggregated with debiasing to extract a calibrated consensus.
Metacognitive feedback. Past human judgments compared with realized outcomes, with personal feedback to analysts on biases, overconfidence and blind spots.
| Mechanism | Process | Impact |
|---|---|---|
| Bayesian thesis updating | Re-weights probability distributions as new signals arrive. | Prevents anchoring; faster adaptation to regime shifts. |
| Point-in-time snapshotting | Freezes the context, inputs and approvals at execution. | Complete auditability and clean, leak-free history. |
| Analyst calibration | Compares stated confidence with actual accuracy. | Sharper team decisions; measurable reliability. |
| Counterfactual learning | Tracks or simulates the options not taken and the scenarios that did not occur. | Learns from far more than the firm's own executed trades. |
own instance, and that learning is never shared.
08
Security and deployment
11. Your instance, your data, your models#
DeOS is delivered as a dedicated enterprise instance. The client owns every piece of data in it and every insight created in it. Every model learns for the client only: DeOS never trains general models on client data and never shares any data or learning with other clients. The client chooses where the instance runs and who runs it.
Deployment options#
| Deployment option | Who runs IT | Where IT runs | What you get |
|---|---|---|---|
| Fully managed | vencortex | A dedicated instance on the cloud provider of your choice: Scaleway, AWS, Google Cloud or Microsoft Azure, in the region you choose | Fastest start, operations and updates handled by vencortex, your data in a region and provider you pick |
| Your cloud account | vencortex, or your IT with vencortex support | Your own cloud account and network, under your cloud contracts and controls | Your cloud governance, your billing and your perimeter, with the same platform |
| On premises | Your IT, with vencortex support | Your data center, inside your network | Full physical control for the strictest data residency and isolation requirements |
In every option the instance is yours alone: no shared tenancy, no shared models, no shared learning. Moving between options keeps policies, context, models and decision memory intact.
Security and safety#
ENCRYPTION
In transit and at rest
All data is encrypted in transit and at rest, including AI inputs, embeddings, model weights and outputs. The decision memory, the context and the models are protected the same way as the source data.
MODELS
Proprietary, nothing leaves
All models are proprietary DeOS models that run inside your instance. No data, prompt, embedding or output is sent to third-party AI providers. Every model learns on your instance, for you only.
KEYS
Key management
Secrets and credentials live in a keychain or key vault. Customer-managed keys are supported, so you hold the keys to your own data and can revoke them.
IDENTITY
Two-factor and single sign-on
Two-factor authentication and single sign-on with your identity provider, so access follows your joiner, mover and leaver processes.
ACCESS
Fine-grained access control
Role-and feature-level permissions for users, teams and funds: who can design policies, who can recommend, who can decide, and who can see which fund, which decision and which data.
SAFETY
Monitoring and approval gates
Model monitoring with drift and quality checks, human approval gates where the policy requires them, and an audit trail of every AI output: which model, which data, which version, which human reviewed it.
Infrastructure, not a tool. What the client buys is infrastructure it controls: a dedicated instance, in the cloud of its choice, in its own account or on premises, with its own keys, its own access rules, its own models and its own decision memory. Not a seat on a generic shared solution.
09
User groups
12. Who works in DeOS#
DeOS separates the roles in a decision, so each person works where they add the most and every step has an owner. Each role is a permission set in the access control of the instance.
| Role | Typically | In DeOS | What they gain |
|---|---|---|---|
| Decision architect | CIO | Sets the decision policies in DeOS Studio: mandate, criteria, constraints, goals and who decides what. | Firm-wide view of policy adherence and decision quality. |
| Decision maker | Investment committee | Takes the decision on an explained recommendation, with alternatives and conviction. | Faster, better-informed decisions with a full audit trail. |
| Recommender | Portfolio manager | Owns the recommendation, adjusts or overrides it, and records why. | Ranked, sized actions with the drivers behind them. |
| Analyst | Analyst | Builds and challenges the context, adds judgment to the analyses. | Coverage beyond manual bandwidth; time for the questions that matter. |
The platform by role#
10
The demo framework
13. One decision through DeOS, in fourteen steps#
A generic, reusable framework for showing DeOS: fourteen steps that follow one decision from policy to the next decision, grouped under the five capabilities. The same framework is used for a live demo, a video script and a slide deck: one step is one scene is one slide. For each step it gives the capability it shows, who leads and who is involved, what the audience sees on screen, the key message the presenter says, and what runs behind the screen.
the analyst and the portfolio manager. Chapter 11 turns the framework into a concrete storyline.
| step · scene · slide | who leads · involved | what the audience sees | what the presenter says | behind the screen |
|---|---|---|---|---|
| Policy, process and guardrails are set before any analysis runs. GOVERNANCE |
sources on one side, the ontology of the object on the other, with proprietary documents and notes attached to the
way the business works, and holds only what the decision needs. It grows with every decision.
internal and third-party sources, decision-centric ontology, action-to-outcome mapping,
Build the context: data connectors and ontology
Human judgment and AI foundation models, fused in one chain.
analysis
Collect expert predictions
Recommender
ranges, confidence and the assistant
that grounds each answer in base rates
ranges and confidence, so it can be
fused with the models and calibrated
judgment extraction, debiased
aggregation, contextual
Collect expert predictions
Recommender
by materiality, with consensus gaps
flagged.
consensus differs from the prediction,
DeOS says so.
as new data arrives.
Insights
| step · scene · slide | who leads · involved | what the audience sees | what the presenter says | behind the screen |
|---|---|---|---|---|
| One ranked, sized and explained course of action; the human decides. RECOMMENDATION |
Why chain
Every version is kept and every decision feeds the next.
learning
policy adherence, overrides, conviction
against outcomes, policy quality.
policies where the evidence says so.
updates policies, models and
analyst calibration.
Observability across the portfolio
against outcomes, policy quality.
analyst calibration.
Who leads each step#
Decision architect
CIO
Decision maker Investment committee
Recommender Portfolio manager
Analyst Analyst, specialists
leads the step
involved. Steps 07 and 08 are run by the system; the analyst and the portfolio manager
review them.
11
Demo storyline
14. A new position in a mid-cap industrial company#
One concrete storyline that follows the framework step by step: a long-only fund considers a new position in a mid-cap industrial company, called the target. Each scene gives the screen that is shown, the script that is said, and the one line and one visual for the slide. The script works as a live demo, as a video narration and as speaker notes. All names and numbers are illustrative.
Three scenes show current product screens (taken from a buyout policy). The other eleven are described in words until their screens are added.
01
One policy for every new position
screen decision policy canvas: the policy 'new position, long-only equity' in
DeOS Studio: Go / No-Go plus Selection, criteria, guardrails, expert questions and the end-to-end analysis.
We start where the firm's edge lives: in DeOS Studio. This is our policy for a new long-only equity position, a Go / No-Go decision on a Selection archetype. It carries our criteria, our concentration and liquidity limits, our expert prediction questions and the analysis chain, all assembled from decision assets we built once.
on the slide
Our edge is a policy, not a prompt. · Visual: The policy canvas with the five
building blocks highlighted.
and AI functions.
02
A new position in the target
screen new decision: policy, object and mandate chosen; DeOS scopes the
process, the context, the analyses and the expected recommendation.
A portfolio manager opens a new decision: a position in the target, a mid-cap industrial company, for our long-only fund with a 3 percent single-name limit. The policy scopes everything: eleven process steps, the context to build, twelve expert questions and one recommendation on the optimization goal, three-year excess return against the benchmark.
on the slide
Start a decision like a project, scoped by the policy. · Visual: The new-decision form and the scoped plan it produces.
03
Eleven steps, every one owned
screen process view: tasks, owners, status and deadlines along the policy; two
steps open, submission blocked until they close.
Here is the process. Seven of eleven steps are done, two are in progress with the analyst, and the committee date is in nine days. The guardrail is visible: the recommendation cannot be submitted while the expert predictions are open.
on the slide
No decision without the required steps. · Visual: The process board with the submission guardrail.
04
The target as it really operates
screen dynamic data room: data connectors to the portfolio system, the data
warehouse, the market data provider and the deal room; the ontology of the target by product line, plant and customer segment.
The analyst builds the context. Four connectors are live: our portfolio system, our data warehouse, the market data provider and the deal room with management materials. The target is modeled by its three product lines and its two plants, not only by its annual report, and the interview notes and audio are attached to the entities they describe.
on the slide
The context is built once and reused. · Visual: The ontology graph with
connected sources around it.
05
Nine situations that look like this one
screen similar situation scan: five similarity lenses, 38 candidates, a curated
shortlist of nine, tagged with the macro and micro regime they sat in.
DeOS scans for similar situations on five lenses: theme, characteristics, causal drivers, behaviour in past shocks and regime. 38 candidates, nine on the shortlist after the analyst's review. This is our benchmark set for the base rates, and every one of the nine carries the regime it lived in.
on the slide
A benchmark set for this decision, not a peer list. · Visual: The lens scores and the shortlist of nine.
06
Twelve questions, five experts
screen expert predictions: each question with a range and a confidence level,
the assistant beside it, and the base rate from the nine similar situations.
Five experts answer twelve questions, for example the order intake growth over the next two years: the sector analyst says 4 to 9 percent at 70 percent confidence, the base rate from the shortlist sits at 5 percent. Judgment becomes data we can fuse and calibrate.
on the slide
Judgment captured as data, with ranges and confidence. · Visual: One expert question with range, confidence and base rate.
Current product screen: expert assessment of a target assumption with confidence and the DeOS assistant.
07
One click, the whole chain
screen end-to-end analysis run: the chain from context to drivers to scenarios to
valuation to sizing; the models and experts fused per driver; two review points for the analyst.
One click runs the end-to-end analysis. The DeOS foundation models predict the drivers, order intake, margin and working capital, fuse them with the expert view by regime, and pass them on: drivers feed scenarios, scenarios feed valuation, valuation feeds sizing. The analyst reviews the fused margin distribution before it moves on.
on the slide
Every prediction feeds the next one. · Visual: The analysis chain with the fused distributions.
08
1,400 insights, 23 that matter
screen insight feed: insights by dimension and materiality; the consensus gap on
the margin flagged at the top.
The run produced about 1,400 insights. 23 are material for this decision. The one at the top is a misconception: consensus expects an EBIT margin of 14.5 percent in year two, our fused prediction has a median of 12.8 percent, because the pricing power the market assumes did not hold in the nine similar situations.
on the slide
Where consensus differs from the prediction, DeOS says so. · Visual: The ranked insight feed with the consensus gap.
09
Buy 2.1 percent, conviction 72
screen decision recommendation: recommended action and size, the outcome
distribution on the optimization goal, conviction, three alternatives, and the effect on the portfolio.
The recommendation: Go, with a position of 2.1 percent of NAV. The expected three-year excess return has a median of 11 percent, with a 10th percentile of minus 9 and a 90th of plus 27. Conviction is 72 percent. Three alternatives are ranked underneath: 1.0 percent, 3.0 percent and pass, each with its distribution and its effect on the fund's sector concentration.
on the slide
One action, with its probabilities and its alternatives. · Visual: The recommendation card: action, distribution, conviction, ranking.
10
Why 2.1 percent, down to the data point
screen why chain: the driver tree from the recommendation to the data points
and expert inputs; conviction per perspective: attractiveness 78, growth 70, liquidity 85, risk 61.
Why 2.1 and not 3.0? The why chain shows it: the margin driver holds the size back, and it traces to two data points from the plant cost data and to the sector analyst's range. Per perspective, attractiveness and liquidity are strong, growth is good, risk is the weak side at 61 because of customer concentration.
on the slide
Nothing in the recommendation is a black box. · Visual: The driver tree with the four perspective convictions.
11
The committee decides
screen committee vote: recommendation summary, votes, comments, and the
recorded override.
The investment committee votes: four of five approve. The committee trims the size to 2.0 percent because of a pending sector exposure, and that reason is recorded with the decision. The recommendation advised, the humans decided.
on the slide
The human always decides, and the override is recorded. · Visual: The decision record with the vote and the override reason.
12
Version 3, and why the view changed
screen audit trail: the version history of the recommendation; for each insight
the data, the AI analysis and the human behind it.
Six weeks later the quarterly results arrive and the analysis reruns. Version 3 of the recommendation keeps the position but conviction drops from 72 to 68, and the audit trail shows exactly why: new order intake data, the same model, one revised expert input. Versions 1 and 2 are kept.
on the slide
Every insight and every version is kept. · Visual: The version timeline with the conviction path.
13
From Go / No-Go to monitoring
screen decision lifecycle: the decision lifecycle of the target: go / no-go
handed on to a Monitoring policy with triggers inherited from the analysis.
The decision lifecycle continues. The monitoring policy picks up the context, the predictions and the insights, and sets its triggers from them: a margin below 11.5 percent or order intake down more than 8 percent reopens the sizing decision. Nothing is rebuilt.
on the slide
The next decision starts where this one ended. · Visual: The lifecycle chain with the hand-off and the triggers.
14
How the firm decides
screen observability: every decision across funds and teams: volumes, policy
adherence, overrides, conviction against outcomes, policy quality.
The CIO closes the loop: 212 decisions this year under this policy family, 94 percent adherence, 7 percent overridden with a reason, and conviction that tracks outcomes. Where a criterion stops working, the policy is adjusted in DeOS Studio, and the next decision runs better.
on the slide
A total portfolio view of how the firm decides. · Visual: The observability dashboard across funds.
Takeaway
15. Own the infrastructure, own the edge#
Data can be rented and models can be rented. The edge that remains is how a firm decides and what it learns from deciding. DeOS keeps both in one place, on infrastructure the firm owns and controls: its own instance, its own data, its own models.
Infrastructure, not a tool
A dedicated instance that runs every decision, with your policies, your context, your models and your decision memory. In the cloud of your choice, in your own account or on premises, and not a seat on a shared solution.
Intelligence, not data
Five capabilities turn data and judgment into ranked, sized and explained actions, in line with your mandate. Every number traces back to a data point or an expert input.
Learning that stays yours
Every decision made and not made improves the next one, and that learning is never shared. DeOS never trains general models on your data.
Secure by construction
Encrypted in transit and at rest, including model weights and outputs. Proprietary models, nothing sent to third-party AI providers. Your keys, your identity provider, and fine-grained access for users, teams and funds.
© vencortex®. This paper describes the DeOS platform. It is not investment advice.