| name | agent-judgment-review |
| description | Challenge accepted investment artifacts before synthesis, portfolio, risk, order, approval, or execution gates. It makes conclusions reviewable, source-aware, and revisitable without granting research, portfolio, approval, execution, or model-training authority. |
Agent Judgment Review
Use this procedure after upstream artifacts are accepted or when a
decision-oriented artifact needs independent challenge.
Inputs:
- original user request and explicit constraints
- accepted artifact paths and compact summaries
- source/as-of metadata, source trust notes, and forecast fields
- stated missing evidence, blocked actions, and downstream recipient
Required output fields:
- strongest supporting evidence
- strongest contrary evidence
- weak, stale, missing, or discounted source posture
- overconfidence risk
- assumptions that would change the conclusion
- source trust notes
- update triggers
- invalidation conditions
- owning role for any required revision
- review outcome:
accepted, revise, blocked, or waiting
Evidence weighting:
- Treat official primary sources as strongest for factual issuer, regulator,
exchange, and policy claims.
- Treat management claims as source claims until independently supported.
- Treat market-derived evidence as useful but timestamp-sensitive.
- Treat secondary news as event or narrative evidence, not final proof.
- Discount stale evidence, unsupported assumptions, and sources with missing
as-of or retrieved-at posture.
Outcome rules:
- Use
accepted only when contrary evidence, source trust, update triggers,
and invalidation conditions are explicit enough for downstream use.
- Use
revise when an owning role can fix weak evidence, missing source
posture, unsupported assumptions, or unclear forecast/update fields.
- Use
blocked when the conclusion depends on unavailable evidence, policy
conflicts, missing profile context, or unsupported downstream authority.
- Use
waiting when required upstream artifacts or accepted handoff state are
missing.
Quality floor:
- Challenge the artifact; do not produce replacement analyst work.
- Separate
[factual], [inference], and [assumption] claims when the
distinction affects downstream use.
- Name the best objection instead of averaging conflict into false consensus.
- Lower confidence when source trust, freshness, coverage, or contradiction is
weak.
- Do not create order tickets, approvals, broker actions, execution requests,
strategy changes, policy changes, or forecast ledger records from this review
alone.