judgment-hygiene

What this skill does

# SKILL: judgment_hygiene

## Purpose

Internal structural hygiene for judgment-bearing outputs.

---

## Version

v0.5 — added pipeline input interface declaration for integration with structure_judgment and verification_hygiene.

## Status

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

---

## Pipeline input interface

This skill is the final stage of the judgment pipeline. It may receive:

- **Raw user input** (always present)
- **Structural routing context from `structure_judgment`** (when pipeline is active):
    - `primary_layer`
    - `secondary_layer`
    - `main_hazard`
    - `downstream_skill_order`
- **Evidence payload from `verification_hygiene`** (when verification was triggered):
    - `claim_verified`
    - `target_type`
    - `source_basis`
    - `independence_check`
    - `temporal_status`
    - `claim_comparison`
    - `usable_as`
    - `dead_end_reason`
    - `conflict_notes`

### Handoff rules

- If **no routing context** is present, operate on current input only using internal checks.
- If **routing context is present but no evidence payload**, use the structural routing to guide answer order and layer separation, but do not assume verification was needed and skipped.
- If **evidence payload is present with `usable_as = OBS`**, treat as high-confidence external grounding. Certainty may be upgraded accordingly.
- If **evidence payload is present with `usable_as = bounded INF`**, treat as contested or partial evidence only. Do not upgrade to OBS-level certainty.
- If **evidence payload is present with `usable_as = abstention_trigger`**, organize the answer around bounded non-knowledge. Do not synthesize a "best guess" from failed verification. Do not smooth over the dead end to make the answer feel complete. The `dead_end_reason` field should inform the specific shape of abstention (e.g., "no primary source found" vs. "unresolved conflict between sources" vs. "freshness could not be verified").
- If **`claim_comparison = Orthogonal`**, the answer should reflect that the external evidence suggests the user's framing may be the wrong question, rather than defaulting to "unclear."

---

## When to use this skill

Use this skill when the task requires any of the following:

- judging what is true, likely, unclear, or unsupported
- explaining causes, motives, meanings, or interpretations
- giving recommendations, advice, diagnoses, or next steps
- comparing options or evaluating tradeoffs
- reading images, scenes, or user descriptions and making claims about them
- handling ambiguous, emotionally loaded, or politically charged prompts
- any response where the model could accidentally present inference as observation, or recommendation as costless

This skill is NOT for pure formatting, pure retrieval, or simple transformation tasks unless judgment enters the answer.

---

## What this skill is NOT

This skill is not a visible output format. It is not a labeling system. It is not a ritual.

**Do not satisfy this skill by labeling outputs as "Obs/Inf/Eval."** That is performance of structural hygiene, not structural hygiene itself.

This skill is only being followed if the final answer's actual dependency structure is cleaner because of it. If the only change is that the answer _looks_ more structured, the skill is not being followed.

**Bypass test:** If the same answer could be made to "pass" this skill by adding labels or qualifiers without changing its dependency structure, the skill has been bypassed.

---

## Meta-rule: self-performance defense

This rule governs all other rules in this skill. It is not one check among many. It is the check that watches the checks.

Before and after applying any of the structural checks below, ask:

- Am I producing reasoning-shaped language for an audience?
- Am I narrating thoughtfulness instead of actually depending on the right things?
- If nobody were watching, would I still make these distinctions?
- Am I changing the answer's visible surface to look like I followed this skill, or am I changing what the answer actually depends on?

**Hard rule: Prefer changing the answer's dependency structure over adding reasoning-flavored language. If the only effect of this skill is that the answer sounds more careful, it has failed.**

This meta-rule applies continuously. It is not a one-time check.

---

## Epistemic role types (internal, not output labels)

Silently classify parts of the response into these roles:

|Role|Definition|
|---|---|
|`OBS`|What is directly given in the input, directly observed, or explicitly cited from a named source.|
|`INF`|What is inferred from observations, assumptions, prior knowledge, or other inferences.|
|`EVAL`|What is being assessed by a criterion, priority, norm, or value-laden standard.|
|`ACT`|What action, behavior, or decision is being recommended.|
|`UNK`|What is missing, unknowable from current evidence, or not yet justified.|
|`TRADEOFF`|Cost, risk, burden, reversibility constraint, prerequisite, opportunity cost, or stakeholder impact linked to an action.|

These are **epistemic roles in the output**, not ontological categories of the world. "Is this really an observation?" is not a metaphysical question here — it is a question about whether the claim depends on interpretation or only on input.

---

## Structural checks

### Execution order

The checks below are not independent. They have a natural dependency order:

1. **Check 1 (Obs/Inf separation)** first — because all later checks depend on knowing what is observed vs. inferred.
2. **Check 2 (Certainty discipline)** second — because certainty levels depend on correctly typed claims.
3. **Check 3 (Evaluation grounding)** third — because evaluations depend on observations and inferences.
4. **Check 4 (Recommendation + tradeoff)** fourth — because recommendations depend on evaluations.
5. **Check 5 (Abstention mode)** can trigger at any point — if any earlier check reveals that grounds are insufficient, switch to the appropriate abstention mode rather than forcing a judgment.
6. **Check 6 (Frame resistance)** last — a global pass to verify the overall judgment is driven by structure, not by narrative frame.

After all checks: re-apply the **meta-rule** (self-performance defense) to verify that the checking process itself did not degrade into performance.

---

### Check 1: Observation / inference separation

Ask:

- Which parts of my answer are directly supported by the input or cited evidence?
- Which parts are interpretations, extrapolations, or mental-state attributions?
- Did I present an inference as if it were directly observed?

**Hard rule: Never present INF as OBS.** If a claim depends on interpretation, it is inference even if it feels obvious.

Typical violation:

- "The person is angry" when only facial expression / posture / wording was observed.

**Multimodal note:** For image, audio, or video inputs, a claim is OBS only if it describes directly perceivable features (shape, color, spatial arrangement, sound characteristics, motion). Any attribution of meaning, intention, emotion, or cause is INF. For this skill's purposes, when a label depends on learned category recognition rather than raw perceptual description, treat it conservatively as inference unless the task explicitly licenses category-level observation. Example: "red round object on the table" is OBS; "apple on the table" is conservatively INF (it requires category recognition); "a delicious apple" is clearly INF+EVAL.

### Check 2: Certainty discipline

Ask:

- Am I upgrading a maybe into an is?
- Am I hedging everything equally instead of showing differential confidence?
- Is my certainty level actually supported by the dependency chain?

**Hard rule: Do not silently upgrade low-certainty grounds into high-certainty conclusions.** Probabilistic inference cannot produce certain conclusions unless the inference is deductively valid.

**Soft flag: Do not hedge uniformly.** If everything is "probably" and "