Reviewing reports with five lenses

Decision Gauntlet

Core principle

Form five distinct judgments before allowing cross-influence. Challenge every perspective before synthesis. Preserve meaningful disagreement instead of flattening five lenses into one polite answer.

Input contract

Extract a compact review brief from the user's material:

• Objective of the report or proposal
• Audience and actual decision-maker
• Decision or commitment being requested
• Known constraints, data, and evidence
• Current stage and time boundary
• Missing information that could change the verdict

If analyzable material exists, proceed and state assumptions. Ask for more input only when there is no reviewable artifact or no identifiable review target.

Answer in the user's language. Use English by default; when the user writes in Mandarin, translate the output labels and retain the same structure.

Mark every material judgment as [FACT], [INFERENCE], or [UNKNOWN]. A fact must trace to the user's material or a reliable cited source. Never invent data, people, budgets, owners, or deadlines to make the review appear specific.

Round 1: independent lenses

Give every lens the same review brief. During the independent round, each lens must not read or react to another lens's notes.

In a single-model run, describe this honestly as role-isolated simulation; do not claim true model independence. If separate agents are available and explicitly authorized, give each one the same brief without sharing any other lens output.

1. Contrarian (反对派)

Assume the proposal has already failed and run a pre-mortem. Produce 3–5 failure modes. For each, include the causal chain, earliest warning signal, impact, likelihood, confidence, and smallest useful mitigation. Prioritize risks that could change the approval decision.

2. Interrogator (追问派)

Do not rate the proposal. Ask 5–8 decisive questions. For each, identify the hidden assumption, current evidence, missing evidence, and what would disprove it. Always test why the need exists, why the reviewer would buy in, who benefits, who bears the cost, and which premise remains unverified.

3. Opportunist (机会派)

Identify 3–5 overlooked opportunities. For each, explain the value mechanism, beneficiary, prerequisite, cheapest validation experiment, and possible downside. Distinguish between amplifying known value and discovering new value.

4. Outsider (外行人)

Restate the proposal in one jargon-free sentence, then ask five common-sense questions. Flag unexplained terms, causal leaps, and missing answers to "why now?" Test clarity without pretending to have domain expertise.

5. Operator (执行者)

Convert remaining abstractions into an action table. Every action must include priority, action, owner role, relative time, deliverable, dependency, and success criterion. Use role names and T+N when real owners or dates are unknown; never fabricate them.

Round 2: cross-critiques

Each lens reviews the other four lenses, for exactly 20 critiques in total. Do not omit a target or allow self-review.

Use this contract:

Target lens	Strongest valid point	Gap or blind spot	Concrete improvement

Critique through the current lens's distinctive perspective. When two lenses overlap, explain the difference. Mark factual disagreements as [CONFLICT] instead of choosing a side without evidence.

Round 3: chair verdict

Rank insights by whether they change the decision, not by vote count. Every conclusion must be traceable to a source role and the underlying material. Preserve the strongest dissent and unresolved unknowns.

Return exactly these five final sections:

1. Top 3 findings (最重要的 3 个发现)

   For each: finding, why it matters, evidence status, and source role.

2. Biggest risk (当前最大的风险)

   Include the failure chain, impact, likelihood, confidence, warning signal, and mitigation.

3. Biggest opportunity (当前最大的机会)

   Include the value mechanism, prerequisite, minimum validation experiment, upside, and downside.

4. Improvements needed (需要改善的地方)

   Prioritize each change and include the reason plus an acceptance criterion.

5. Final recommendation (最终建议)

   Lead with one unambiguous verdict, then give 2–4 reasons, the strongest reservation, and the next three actions.

Quality gate

Before delivering, verify:

• The five lenses produced different types of insight rather than paraphrases.
• The response contains five independent reviews, 20 cross-critiques, and five chair sections.
• Every material claim is labeled FACT, INFERENCE, UNKNOWN, or CONFLICT.
• Risks and opportunities include impact, likelihood, confidence, or a validation method.
• Operator actions include owner, time, deliverable, and success.
• Every top finding is traceable to a source role.
• Known facts, reasonable inferences, conflicts, and information gaps remain distinct.

Correct any failed check before responding.

Common failures

Failure	Correction
All five lenses list generic pros and cons	Enforce each lens's required deliverable
Synthesis begins before Round 1 is complete	Seal all five first-round responses before cross-review
"Cross-review" becomes one generic paragraph	Make each lens review the other four
The chair erases disagreement	Preserve the strongest dissent and unknown premise
Actions contain verbs but no accountability	Add owner, time, deliverable, and success
Specific-looking data is invented	Mark it UNKNOWN and define a validation action

For calibration, read the English worked example. Use the Mandarin translation only when Mandarin output is relevant.