Most financial AI analyzes the opportunity.
This one audits the person making the decision.
That was the idea behind the Decision Integrity Engine I built for Binance’s #BuildWithYou 01: Finance Prompts.
The interesting problem isn't always whether an asset looks attractive.
Sometimes the bigger problem is the person making the decision.
A trader can have a reasonable thesis and still be acting from FOMO.
They can have strong evidence and still violate their own risk rules.
They can believe they're making a long-term decision while their behavior is actually driven by a short-term loss.
And an AI system that simply researches the asset may never notice that contradiction.
That's what I wanted to investigate.
The Core Inversion
Instead of asking:
“What does the market look like?”
the engine asks:
“Is the decision-making process internally coherent?”
It separates four things that are often accidentally mixed together:
BELIEF
What the person already believed before new research.
BEHAVIOR
How their personal history, emotions, incentives, and tendencies may be influencing the decision.
RISK
The limits the person themselves established before making the decision.
EVIDENCE
New, independently researched information gathered after the original thesis has been captured.
That separation is critical.
Belief ≠ Evidence
The engine first preserves the user's original thesis.
It does not rewrite, strengthen, weaken, or reinterpret it during intake.
Only after the belief has been captured does the system research external evidence.
That means it can later show:
What I believed → how I was behaving → what my risk rules required → what the evidence actually showed → where those elements agreed or conflicted.
The AI isn't allowed to quietly rewrite the original reasoning after discovering new information.
The Six-Stage Architecture
The engine follows a deliberate sequence:
1. Decision Intake
Capture the asset, intended action, position size, time horizon, decision context, and the user's thesis in their own words.
2. Decision-Maker Audit
Examine relevant behavioral factors such as FOMO, revenge motivation, overconfidence, urgency, confirmation-seeking, or tendencies to chase or abandon risk rules.
But only when supported by the user's answers.
No invented psychology. No diagnosis.
3. Personal Risk Boundary
The user defines their own rules.
Maximum risk.
Position size.
Risk/reward requirements.
Stop-loss rules.
Leverage limits.
Avoidance conditions.
If no rule exists, the engine says:
No pre-existing rule provided.
It doesn't invent a “safe” limit.
4. External Evidence
Only now does the engine research current, credible evidence relevant to the decision.
It separates:
Evidence supporting the original thesis
from
Evidence contradicting the original thesis.
Material unknowns and evidence confidence are also recorded.
5. Reconciliation
The four pillars are compared without collapsing them into one judgment.
The system looks for contradictions such as:
A long-term thesis being executed through short-term revenge.
A proposed position violating the user's own risk boundary.
External evidence materially contradicting the original thesis.
A reasonable thesis being undermined by fragile decision-making behavior.
Evidence being insufficient to meaningfully reconcile the decision.
6. Decision Integrity Report
The engine produces a structured report showing the original thesis, behavioral findings, risk-boundary findings, supporting and contradicting evidence, material unknowns, key conflicts, and conditions that would logically invalidate the decision.
The Final Status Isn't “Buy” or “Sell”
This is probably the most important part.
The engine can assign one of four decision-process statuses:
ALIGNED
No material unresolved contradiction is detected.
MISALIGNED
Meaningful contradictions exist between belief, behavior, risk, or evidence.
FRAGILE
The process depends on weak evidence, substantial uncertainty, or significant behavioral distortion.
BROKEN
Multiple serious contradictions or explicit violations make the process internally inconsistent.
But:
ALIGNED does not mean “buy.”
MISALIGNED does not mean “short.”
BROKEN does not mean “sell.”
The engine audits the decision process.
The user still makes the decision.
Why I Built It This Way
A lot of financial AI is designed around improving the answer.
I wanted to explore something different:
What if the AI's job was to improve the quality of the decision process instead?
Not by taking control away from the person.
Not by manufacturing certainty.
Not by pretending behavioral distortions can be “fixed.”
But by making contradictions harder to hide.
The central idea is simple:
Most financial AI analyzes the opportunity.
This one audits the person making the decision.
I built and submitted the full Decision Integrity Engine for #BuildWithYou #BuildWithYou ompts.
The complete production version is available in my X submission below.