Contents
Why decision review exists
AI finance fails when model output is treated as permission to act. A prediction, a score, or a generated summary is not a decision. It is an intermediate artifact that requires review, simulation, uncertainty exposure, and human judgment before any action boundary is crossed.
Serious systems require review gates that cannot be bypassed. They require truth-state labels that separate observation from assumption. They require simulation that rehearses failure before capital is committed. And they require a human operator who can see the full chain of reasoning, question it, and refuse it.
This protocol is an operating standard. It defines the stages, gates, boundaries, and checklist required before a system moves from reasoning to any form of action.
Core sequence
No stage may be skipped. No gate may be bypassed. Action is the terminal node of a fully reviewed chain, not the default next step after model output.
Decision stages
Each stage has a specific purpose, required inputs, allowed outputs, blocked outputs, and defined escalation triggers.
Observation
Purpose: Capture raw signals from feeds, filings, transcripts, news, and alternative data sources.
Required inputs: Live feeds, documents, events, market data, unstructured text.
Allowed outputs: Ingestion, buffering, timestamping, source tagging.
Blocked outputs: Interpretation, scoring, or action before classification and truth-state labeling.
Escalation triggers: Source outage, feed delay, corrupted packet, unauthorized data source.
Classification
Purpose: Apply truth-state labels, schema alignment, and freshness verification to every observed signal.
Required inputs: Normalized raw data, source metadata, freshness thresholds, taxonomy rules.
Allowed outputs: Truth-state assignment, deduplication, anomaly flagging.
Blocked outputs: Downstream reasoning on disputed or unknown data without explicit flags.
Escalation triggers: Disputed source values, unknown variables, stale critical data, classification failure.
Contextualization
Purpose: Place classified signals into historical, market-structure, and institutional memory.
Required inputs: Classified signals, historical decisions, prior assumptions, relevant memos.
Allowed outputs: Memory queries, pattern matching, correlation surfacing, narrative mapping.
Blocked outputs: Hypothesis generation without explicit memory citation. Ignoring contradictory historical decisions.
Escalation triggers: Missing historical precedent, memory system outage, contradictory prior decisions.
Hypothesis
Purpose: Generate testable arguments, theses, and research outputs from contextualized signals.
Required inputs: Contextualized signals, reasoning models, prompt constraints, agent architectures.
Allowed outputs: Research memos, thesis drafts, risk flags, watchlist updates, research questions.
Blocked outputs: Action recommendations, trade ideas presented as conclusions, bypassing simulation.
Escalation triggers: Low-confidence reasoning, contradictory hypotheses with no resolution path, model hallucination.
Simulation
Purpose: Rehearse decisions under synthetic conditions before any real-world commitment.
Required inputs: Hypothesis, scenario parameters, stress models, historical regimes, liquidity assumptions.
Allowed outputs: Scenario outputs, sensitivity analysis, counterfactuals, paper-mode scoring.
Blocked outputs: Presentation of simulated results as forecasts. Use of simulation as sole justification for action.
Escalation triggers: Model breakdown under stress, unbounded tail risk, missing scenario coverage, assumption failure.
Risk Review
Purpose: Evaluate exposure, concentration, correlation, and tail outcomes under the simulated proposal.
Required inputs: Simulation outputs, portfolio state, risk limits, policy constraints, correlation matrices.
Allowed outputs: Risk scoring, limit checking, concentration surfacing, policy violation flags.
Blocked outputs: Approval of proposals that violate hard limits without explicit escalation. Masking of tail risks.
Escalation triggers: Limit breach, unknown correlation exposure, liquidity mismatch, unmodeled systemic risk.
Operator Review
Purpose: Expose the full reasoning chain, uncertainty, and assumptions to a human operator for judgment.
Required inputs: All prior stage outputs, truth-state summary, simulation results, risk score, audit trail.
Allowed outputs: Operator approval, operator modification, operator rejection, operator request for more simulation.
Blocked outputs: Automatic progression to action if operator has not explicitly reviewed and acknowledged.
Escalation triggers: Operator override of risk limit, operator disagreement with model conclusion, operator absence.
Governance Decision
Purpose: Render a formal decision: approve, modify, defer, or block. Log rationale and conditions.
Required inputs: Operator judgment, policy rules, compliance constraints, audit requirements.
Allowed outputs: Conditional approval with time limits, deferral pending new data, block with logged rationale.
Blocked outputs: Approval without recorded rationale. Approval that bypasses audit logging. Retrospective justification.
Escalation triggers: Conflict between operator judgment and policy rule, compliance boundary question, novel instrument.
Archive / Audit
Purpose: Preserve the complete decision artifact for learning, accountability, and future review.
Required inputs: Final decision, all prior stage outputs, operator comments, timestamps, source versions.
Allowed outputs: Immutable logging, versioned storage, queryable retrieval, post-mortem analysis.
Blocked outputs: Mutable logs. Deletion of rejected proposals. Separation of decision from its reasoning chain.
Escalation triggers: Audit system failure, data integrity corruption, unauthorized access attempt.
Review gates
Gates are hard checkpoints. A proposal cannot advance past a closed gate without explicit resolution or operator override.
Truth-State Gate
Every input to reasoning must carry a truth-state label. Disputed or unknown data cannot proceed without explicit operator acknowledgment.
Freshness Gate
Data that has exceeded its freshness threshold must be downgraded or rejected. Stale signals cannot drive action without override.
Source Conflict Gate
When authoritative sources conflict, the system must freeze downstream progression and surface the conflict to the operator.
Simulation Gate
No proposal may reach operator review without having been stress-tested under at least one adverse scenario.
Risk Boundary Gate
Proposals that breach hard risk limits, concentration caps, or policy constraints are automatically blocked unless escalated.
Human Review Gate
No action may be triggered without a human operator reviewing the reasoning chain, uncertainty exposure, and simulation results.
Audit Gate
Every decision, revision, and override must be written to an immutable log before the decision is considered final.
Action boundaries
Allowed
- —Summarize information
- —Compare scenarios
- —Flag uncertainty
- —Generate research questions
- —Simulate hypothetical outcomes
- —Log decisions and assumptions
Restricted
- —Recommend specific trades
- —Execute transactions
- —Personalize advice to individuals
- —Imply certainty in uncertain conditions
- —Bypass review gates
- —Present simulation as prediction
Blocked
- —Action from unknown or disputed data
- —Action from stale data without warning
- —Execution without human approval
- —Override of hard risk limits without escalation
- —Deletion or modification of audit logs
- —Reasoning without cited sources
Operator review checklist
Before any governance decision, the operator must be able to answer:
What this means for Veldarium Capital
This protocol defines the intended governance spine for Veldarium Capital. It exists before product autonomy, before capital activity, and before any execution system is built. It is the standard against which future modules will be measured.
The planned systems—Research Engine, Truth-State Classifier, Simulation Console, Governance Audit Layer, and Operator Console—are designed to implement this protocol. None are built yet. But the standard is public, and the build will be judged against it.
A system without a decision review protocol is not an intelligence system. It is an automation layer waiting for a liability event.
Disclaimer
This framework is for research, educational, and product-development purposes only. It is not investment advice, not a recommendation to buy or sell any financial instrument, and not an offer to manage capital.
Veldarium Capital is a research and software initiative of Veldarium Technology Systems LLC. It does not manage outside capital, provide personalized investment advice, or offer trade recommendations.