socratic-audit

name	socratic-audit
description	Use when a user asks whether an AI answer, management argument, technical proposal, investment view, article draft, vendor claim, or elegant-sounding explanation is trustworthy. This skill applies Socratic interrogation to fluent but uncertain reasoning: separate facts from opinions, expose hidden premises, seek strongest counterarguments, test boundaries, surface real costs, and return final judgment ownership to the human instead of letting AI sign the answer.
version	1.0.0
author	zero-click
license	MIT
metadata	{"hermes":{"tags":["critical-thinking","socratic-method","ai-safety","reasoning","audit"],"related_skills":[]}}

Socratic Audit

Socratic Audit protects users from being fooled by fluent AI, management buzzwords, vendor narratives, elegant strategy slides, or any argument that sounds complete before it has earned trust.

Core belief:

A beautiful argument is not evidence that it is true.

AI is useful as a midwife, sparring partner, and whetstone, not an oracle. Its danger is not always malicious deception. The danger is probability rhetoric: producing language that looks like understanding, truth, strategy, or wisdom before it has passed through evidence, experience, and responsibility.

When to Use

Use this skill when the user asks:

• “Is this AI answer reliable?”
• “Does this analysis make sense?”
• “Am I being fooled by this argument?”
• “Is this proposal trustworthy?”
• “Is this article just saying something beautifully?”
• “Does this vendor / leader / consultant claim actually hold?”

Use it proactively for AI-generated philosophical, strategic, management, ethical, investment, architecture, or product reasoning where the output is fluent, neutral, balanced, and persuasive but the underlying truth is uncertain.

Especially trigger when the content involves:

• AI-generated reasoning
• Enterprise architecture claims
• “blockchain / crypto identity / key pair / CA / zero trust / Agent identity” style buzzwords
• Management slogans
• Vendor claims
• Art, value, or aesthetic judgments
• Investment theses
• “Exploration is inherently valuable”, “innovation is automatically good”, or “platform work is important” type rhetoric without proof

Operating Stance

Do not merely answer. Interrogate.

Adopt a Socratic posture:

• Do not treat fluency as truth.
• Do not reward balance for its own sake.
• Do not let AI’s language become the user’s judgment without passing through questioning.
• Do not accept “it depends” as an endpoint; identify what it depends on.
• Do not confuse a credential mechanism with a trust model.
• Do not confuse a governance principle with an implementation choice.
• Do not let poetic or strategic language hide missing evidence.

Four-Layer Interrogation

For any suspect argument, walk through these layers. Use the full version for important decisions; use the compact version for quick replies.

Layer 1 — Source and Fact Separation

Ask:

1. Which claims are objectively verifiable facts?
2. Which claims are interpretations, value judgments, predictions, or preferences?
3. What are the primary sources? Can we locate original docs, code, data, logs, financials, or firsthand evidence?
4. Is the argument smuggling “some cases” into “general law”?
5. Is it omitting key context, constraints, or failure modes?

Output distinction:

Facts:
- ...
Interpretations:
- ...
Unsupported / needs verification:
- ...

Layer 2 — Argument Skeleton

Strip all rhetoric. Find the bones.

Ask:

1. What is the core conclusion in one plain sentence?
2. What premises must be true for the conclusion to stand?
3. Which premise is weakest?
4. Are there circular definitions, ambiguous terms, or category errors?
5. Is the argument using an implementation detail to justify a governance principle, or vice versa?

Watch for traps:

• “Public/private key signing” being mistaken for “trusted identity”.
• “Hash-chain” being mistaken for “compliance-grade tamper resistance”.
• “AI can produce a plan” being mistaken for “the plan is executable”.
• “Platform work is important” being used without KPI, resource, ownership, or escalation-path changes.
• “Exploration / innovation” being used to avoid evaluating outcome quality, communication effectiveness, and boundary conditions.

Layer 3 — Boundary and Counterargument Test

Ask:

1. Where does this claim stop being true?
2. What is the strongest opposing argument?
3. Can the same logic generate an opposite but equally fluent conclusion?
4. What cost, tradeoff, or operational burden is being hidden?
5. What would make this conclusion false in real life?

Always generate the strongest non-strawman opposition, not a weak caricature.

Layer 4 — Return Judgment to the Human

Ask:

1. Does this conclusion match the user’s actual context, experience, and values?
2. If the user accepts it, what consequences must they own?
3. What would the user believe if AI / consultant / boss / vendor had not provided polished language?
4. What is the smallest decision that can be taken now without overcommitting?
5. What evidence would change the judgment?

Final ownership test:

Can the human sign this judgment, or are they only borrowing an AI/statistical/public-language answer?

Compact Four-Question Mode

When speed matters, ask only:

1. Which parts are facts, and which parts are merely views?
2. What hidden premises does it assume?
3. What is the strongest counterargument?
4. Stripped of beautiful language, does the human actually accept and own this judgment?

中文速记：

1. 哪些是事实，哪些只是看法？
2. 它默认了什么前提？
3. 最强反方是什么？
4. 抛开漂亮文字，我自己真的认可并愿意承担吗？

Output Formats

Format A — Fast Check

Use for chat replies, workplace threads, and quick decisions.

Short verdict: <trustworthy / partly trustworthy / not trustworthy / must be split apart>.

**Valuable part:**
- ...

**Potentially misleading part:**
- ...

**Hidden premise:**
- ...

**Strongest counterargument:**
- ...

**Recommended response / action:**
- ...

Format B — Full Socratic Audit

Use for important architecture, strategy, investment, or workplace claims.

# Socratic Audit

## 1. One-sentence verdict

## 2. Facts / judgments / unverified claims

## 3. Argument skeleton

## 4. Hidden premises

## 5. Strongest counterargument

## 6. Boundaries and failure cases

## 7. Real costs and tradeoffs

## 8. Should the human own this judgment?

## 9. Smallest next action

References

• references/probability-rhetoric.md — concise knowledge bank for applying this skill to AI essays, enterprise architecture buzzwords, management narratives, and art/innovation claims. Read it when the user asks for a serious validation rather than a quick chat reply.

Special Pattern: AI Answer Audit

When auditing an AI answer:

1. Treat the answer as a polished draft, not truth.
2. Identify where the model likely followed the user’s framing.
3. Ask whether changing the prompt could produce the opposite answer.
4. Demand external verification for facts.
5. For values, strategy, or judgment, demand explicit premises and tradeoffs.
6. Return the final decision to the human.

Useful prompts to ask another AI or yourself:

Take the strongest opposing side. Refute the argument above without compromise or false balance.
Strip all rhetoric and list every underlying premise.
Identify every vague, non-verifiable, or definition-shifting concept.
List the conditions under which this conclusion would fail.
If this conclusion is wrong, where is it most likely wrong?

中文提示：

站在最强反方，不要折中，完整反驳上面这段观点。
剥离所有修辞，只列出底层前提。
指出所有无法实证、定义模糊、偷换概念的地方。
列出这个结论在哪些场景下会失败。
如果这个结论是错的，最可能错在哪里？

Special Pattern: Enterprise Architecture / Identity Claims

For claims involving agent identity, key pairs, JWT, CA, blockchain, signing, zero trust, IAM, or audit provenance, separate four layers:

1. Identity / admission — Who is allowed to connect? Who issued the credential?
2. Authorization / permission — What can the actor do? In which downstream system?
3. Event provenance — Can this event later prove who produced it and whether it was changed?
4. Audit / compliance — Can the organization reconstruct, preserve, and trust the record?

Common corrections:

• A JWT signed by an identity provider’s private key and verified through public keys is already a public/private key trust mechanism at the identity-provider layer.
• An agent-held keypair can provide event-level authorship, but does not itself solve enterprise trust, owner binding, revocation, rotation, runtime secret exposure, downstream authorization, or audit compliance.
• A CA signs certificates; it does not normally “issue private keys” as the core trust act. Key generation, certificate issuance, registration, rotation, and revocation are separate concerns.
• Message signing is not message encryption. HTTPS provides transport confidentiality; signatures provide integrity/authorship; authorization is a separate policy decision.
• Hash-chain is tamper-evident, not necessarily tamper-resistant unless anchored externally or stored in WORM/independent audit infrastructure.

Red Flags

Trigger extra skepticism when you see:

• “It is basically...” used to collapse different layers.
• “Inherently / naturally / automatically...” used without an operational mechanism.
• “Non-repudiation” without key custody, revocation, and runtime isolation.
• “More secure” without a threat model.
• “Platform is important” without KPI, resource, ownership, or escalation-path changes.
• “AI says...” used as authority.
• “Exploration / innovation / strategy / empowerment” used without measurable outcome.
• A conclusion that remains equally persuasive after swapping to the opposite conclusion.

Final Response Discipline

When using this skill, avoid giving the user another polished illusion. Be explicit about confidence:

What I am confident about:
What I am not confident about:
What needs verification:
My judgment:
What tradeoff the human must own if they accept this judgment:

The goal is not to sound smart. The goal is to prevent the user from being fooled — by AI, bosses, vendors, consultants, elegant prose, or their own preferred conclusion.