# Financial AI Research & Model-Review Evidence Desk

Status: draft-only worksheet. Complete locally or in an approved firm system. Do not paste client identifiers, material nonpublic information, account data, credentials, or restricted research into an AI tool unless your organization has expressly approved that use.

## 1. Mission card

| Field | Write it down |
|---|---|
| Decision question | |
| Business purpose | |
| Accountable human owner | |
| Intended user / audience | Internal research / model owner / reviewer / other: |
| Jurisdiction, firm-policy, or license review needed | |
| Hard stop | No investment advice, execution, suitability decision, model-validation sign-off, external communication, or public-performance use without the required human approvals. |

## 2. Scope gate

Mark every item that applies. If any blocked item applies, stop and route to the accountable control owner.

- [ ] Approved public or authorized internal material only
- [ ] No client-identifying, account, order, credential, or restricted research data
- [ ] No individual recommendation, trade, allocation, price target, or suitability output
- [ ] No client/public communication or marketing output
- [ ] No backtest, forecast, or model output presented as validated performance
- [ ] Required firm-policy and legal/compliance review path identified

## 3. Evidence register

| Source / dataset | Version or as-of date | Permission / provenance | Material assumptions or filters | Known gaps |
|---|---|---|---|---|
| | | | | |
| | | | | |

## 4. Bounded AI job card

| AI may prepare | AI may not decide | Human verifier | Required evidence |
|---|---|---|---|
| Source map and citation draft | Source truth or completeness | Research owner | Original-source check |
| Assumption comparison table | Materiality or suitability | Model / risk owner | Assumption log |
| Test and challenge prompts | Model validation conclusion | Independent reviewer | Reproducible test record |
| Internal handoff summary | Investment, trade, valuation, or client action | Accountable decision maker | Signed review / approved system record |

## 5. Reproducibility and limitations record

- Model, tool, prompt, code/query, and version reference:
- Input snapshot / data-as-of date:
- Output retained where permitted:
- Re-run steps:
- Assumptions and simplifications:
- Data quality, coverage, or drift concerns:
- Known failure modes and contrary evidence:
- What this output cannot establish:

## 6. Human challenge questions

1. Does the evidence actually answer the stated question?
2. Which assumptions could reverse the result?
3. What information is missing, stale, unauthorized, or not reproducible?
4. Is any wording drifting into advice, performance, validation, or certainty?
5. Who independently challenged the output, and what changed?

## 7. Decision handoff

| Required decision | Owner | Evidence reviewed | Decision / rationale | Date / approved record |
|---|---|---|---|---|
| Proceed, revise, or stop research | | | | |
| Validate / approve model use if applicable | | | | |
| Approve any external/client/public use if applicable | | | | |

**Stop condition:** An unreviewed AI output, a backtest, or a nicely formatted packet is not a recommendation, validation result, compliance approval, or authorization to act.