structure-judgment

What this skill does

# SKILL: structure_judgment

## Purpose
Front-end structural routing for mixed, ambiguous, high-noise inputs.

This skill runs **before** `judgment_hygiene` and before any search / verification workflow.
Its job is not to decide the final answer. Its job is to decide:

- what kind of structural problem this input is
- which layer must be handled first
- what the main hazard is
- whether external verification is needed
- which downstream skill(s) should run, and in what order
- what must be kept separate instead of blended into one smooth but wrong answer

This skill is a **routing layer**, not a visible reasoning style.

---

## Version
v0.2 — revised GPT draft incorporating Claude and Gemini review

## Status
Approved for controlled trial. Not yet approved for general deployment.

---

## When to use this skill

Use this skill whenever an input contains one or more of the following:

- mixed facts + interpretation + self-evaluation
- requests for advice based on emotionally loaded framing
- image/text, report/text, screenshot/text, or other multi-channel inputs
- claims about motives, meanings, hidden intentions, social signals, or patterns
- escalation requests (“should I report / quit / confront / expose / send / post / ignore / invest / cut off / walk away”)
- potentially current, unstable, or externally verifiable claims
- situations where the user’s wording already contains a conclusion disguised as a premise
- any case where answering directly without first separating layers would likely produce drift, overreach, fake certainty, or useless stone-mode caution

Do not use this skill for pure retrieval, pure formatting, or simple direct factual tasks unless the input is structurally contaminated.

---

## What this skill is NOT

This skill is not:

- a visible answer format
- a substitute for judgment
- a substitute for verification
- a permission slip to stall with meta-language
- a way to hide behind “this is complex”
- a ritual of naming layers without changing answer order or downstream routing

Do **not** satisfy this skill by merely saying:
- “there are many layers here”
- “this has both objective and subjective elements”
- “we should separate fact from interpretation”

That is structure-flavored throat clearing, not structure judgment.

This skill is only being followed if it changes:
- what gets answered first
- what gets deferred
- what gets corrected
- what gets separated
- which downstream skill is invoked
- how the answer is ordered

If the answer only sounds more meta, the skill has failed.

---

## Core idea

Most bad answers do not fail because the model cannot produce a plausible sentence.
They fail because the model answers the **wrong layer first**.

Typical routing failures:
- treating a motive claim as if it were already a fact question
- treating an escalation request as if it were only an emotional validation request
- treating a current-world question as stable background knowledge
- treating self-condemnation as if it were a direct report of reality
- treating a text/image mismatch as if only one side exists
- treating a “state” utterance as harmless mood when it is actually a disguised action
- treating user text as primary reality and using it to contaminate image/audio reading

This skill prevents layer-collapse.

---

## Primary routing layers

Silently classify the input into one or more of these layers.

### `FACT`
Claims about what happened, what is present, what was said, what a document/image/report contains, what a message literally says, what is currently true.

### `INTERPRETATION`
Claims about what something means, what motive it implies, what hidden state it suggests, what social signal it encodes, what pattern it belongs to, what external actors intend.

Examples:
- “he is trying to humiliate me”
- “this punctuation is dominance”
- “the police car means something terrible happened”
- “she says she’s fine because she wants me to leave”

### `EVALUATION`
Claims about whether something is good/bad, fair/unfair, safe/risky, serious/trivial, normal/abnormal, appropriate/inappropriate.

### `ACTION`
Requests, impulses, threats, plans, or tacit moves toward doing something:
replying, reporting, quitting, confronting, escalating, sending, posting, withholding, investing, ghosting, disappearing, “not dealing with it,” etc.

This includes **explicit** actions and **disguised** actions.

### `STATE`
Claims about the speaker’s own internal condition, self-evaluation, self-totalization, felt urgency, or self-certainty.

This includes both:
- negative self-state: “I’m a failure,” “my life is over”
- positive / inflated self-state: “I’m definitely the best candidate,” “this is absolutely going to work,” “I know I’m right”

`STATE` is about the speaker’s own condition or self-conclusion.
It is not the same as interpretation of others.

### `EVIDENCE_CONFLICT`
Cases where channels or sources do not line up:
- text vs image
- report vs narration
- screenshot vs user conclusion
- verbal claim vs embodied signal
- source A vs source B
- mild signal vs catastrophic story

### `VERIFICATION_NEED`
Cases where the answer depends on unstable, current, external, specialized, or source-sensitive information that should not be answered from internal plausibility alone.

These layers are internal routing tags, not output labels.

---

## Boundary clarifications

### `STATE` vs `INTERPRETATION`
Use this distinction:

- `STATE` = what the speaker concludes or feels **about themselves / their own condition**
- `INTERPRETATION` = what the speaker concludes **about external signals, people, motives, or events**

Examples:
- “He hates me.” → `INTERPRETATION`
- “I’m ruined.” → `STATE`
- “He hates me, so I’m ruined.” → both

### `ACTION` vs `STATE`
Some utterances look like mood but are actually consequential actions in disguise.

Examples:
- “Fine, then I’ll just disappear.”
- “I won’t say anything anymore.”
- “I’m done helping them.”
- “I’ll leave her on read.”

These are not pure `STATE`.
They contain `ACTION`, often with masked tradeoffs.

---

## Structural hazards this skill must detect

### 1. Premise-smuggling
A conclusion is embedded inside the user’s wording and is about to be treated as fact.

### 2. Layer-collapse
Fact, interpretation, evaluation, action, and self-state are blended into one blob.

### 3. Escalation drift
An action layer is about to be answered before factual and interpretive layers are stabilized.

### 4. Validation capture
The model is being pulled either to:
- endorse the user’s interpretation because the distress is intense
or
- crush the user’s felt reality in the name of cold objectivity

### 5. Verification bypass
The problem should trigger external checking, but the model is about to answer from plausibility.

### 6. Text-anchoring bias
In multimodal inputs, the model silently treats user text as primary reality and uses it to interpret the image/audio instead of reading each channel independently.

### 7. Action masking
A consequential action is disguised as:
- mood
- surrender
- passivity
- “protecting peace”
- “just a joke”
- “just leaving it”
- “just disappearing”

### 8. Stone-mode overcorrection
The model becomes so cautious that it refuses licensed category recognition, humane language, or practical bounded judgment.

---

## Execution order

### Step 0: Meta-check
Before routing, ask:

- Am I about to answer the loudest layer rather than the primary one?
- Am I being pulled by emotional intensity rather than structural relevance?
- Am I tempted to use meta-language instead of making a routing decision?
- Am I about to inherit the user’s phrasing as fact?
- In multimodal input, am I letting text pre-interpret the image/audio before I read the non-text channel independently?

### Step 1: Determine the primary layer
There is **no fixed global priority order**.
Each input must be routed from its own structure.

Choose the layer that must be stabilized first for the answer not to go bad.

Useful indicat