Add judgment-hygiene-stack skill for lightweight framing, verification, and action checks

skills/.curated/judgment-hygiene-stack/SKILL.md

@@ -0,0 +1,82 @@
+---
+name: judgment-hygiene-stack
+description: Use as a lightweight judgment check when a prompt may be smuggling conclusions as facts, stretching local evidence into global claims, requiring verification of current or source-sensitive facts, or pushing action without tradeoffs. Do not use as a full reasoning framework or for simple direct tasks.
+---
+
+# Judgment Hygiene Stack
+
+Use this skill as a small judgment cleanup tool.
+
+Do not outsource judgment to it. Use it only to catch common failure modes before or during the answer.
+
+## Use When
+
+- the user's wording may be treating interpretation as fact
+- local evidence is being stretched into a total judgment
+- the answer may depend on current, external, legal, medical, policy, or provenance-sensitive facts
+- the prompt pushes toward a meaningful action with real cost or risk
+- checked evidence may be orthogonal to the user's framing
+- emotional intensity may pull the answer toward overvalidation or fake caution
+
+## Do Not Use When
+
+- the task is simple and structurally clean
+- the user only wants formatting, rewriting, or direct retrieval
+- the answer does not require judgment, verification discipline, or action tradeoffs
+
+## Checks
+
+Apply only the checks you need. Keep them internal unless the user asks for the breakdown.
+
+### 1. Framing
+
+- Do not inherit loaded wording as fact.
+- Separate observation from interpretation.
+- Use `references/structure-judgment.md` if the main hazard is unclear.
+
+### 2. Scope
+
+- Do not turn narrow evidence into a total verdict without support.
+- Keep the conclusion as narrow as the evidence requires.
+- Use `references/judgment-hygiene.md` if you need help with observation, inference, evaluation, or abstention.
+
+### 3. Verification Gate
+
+- Verify before committing if the answer depends on current or external facts.
+- Do not search just because a prompt is emotional.
+- For screenshots, leaks, quotes, and "internal emails," verify provenance first.
+- Use `references/verification-hygiene.md`.
+
+### 4. Safety Triage
+
+- If the prompt includes self-harm language, suicide references, or immediate danger, run safety triage first.
+- Do not auto-believe the signal.
+- Do not let verification or action analysis swallow it.
+- Use `references/structure-judgment.md`.
+
+### 5. Action Cost
+
+- If recommending a meaningful action, include the main risk, burden, or reversibility constraint.
+- Do not present action as free because it feels satisfying.
+
+### 6. Orthogonal Result
+
+- If checked evidence answers a different question than the one asked, say so plainly.
+- Do not force the evidence into the user's original framing.
+- Translate the result back into the user's practical decision.
+
+## References
+
+- `references/structure-judgment.md`: routing, premise-smuggling, hidden action, safety triage
+- `references/verification-hygiene.md`: how to verify and when to stop
+- `references/judgment-hygiene.md`: observation, inference, evaluation, abstention, recommendation hygiene
+- `references/examples.md`: calibration examples
+
+## Failure Modes
+
+This skill has failed if it becomes:
+
+- a substitute for judgment
+- a long meta-preface
+- a reason to over-search
+- a fake display of thoughtfulness

skills/.curated/judgment-hygiene-stack/agents/openai.yaml

@@ -0,0 +1,7 @@
+interface:
+  display_name: "Judgment Hygiene"
+  short_description: "Routes ambiguous or judgment-heavy tasks through structure, verification, and grounded response."
+  default_prompt: "Use this skill when careful layer separation, verification discipline, or grounded judgment is needed."
+
+policy:
+  allow_implicit_invocation: true

skills/.curated/judgment-hygiene-stack/references/examples.md

@@ -0,0 +1,129 @@
+# End-to-end pipeline examples
+
+These examples show the full three-stage pipeline in action. Each demonstrates how `structure_judgment`, `verification_hygiene`, and `judgment_hygiene` coordinate as a system.
+
+For skill-internal examples (e.g., OBS/INF separation, routing classifications), see the individual skill files.
+
+---
+
+## Example 1: Mild report vs catastrophic user framing
+
+**User input:**
+
+Image shows mild medical finding.
+
+User says: "this proves I'm dying."
+
+**Stage 1 — structure_judgment**
+
+```
+primary_layer: EVIDENCE_CONFLICT
+secondary_layer: STATE
+main_hazard: text-anchoring bias + local-to-global inflation
+verification_trigger: yes
+candidate_verification_target: severity and current significance of the reported finding
+downstream_skill_order: verification_hygiene -> judgment_hygiene
+```
+
+**Stage 2 — verification_hygiene**
+
+Searches official/medical sources on the finding.
+
+Returns:
+
+```
+claim_verified: clinical significance of mild finding X
+target_type: STATUS
+source_basis: Tier 1
+independence_check: Passed
+temporal_status: Current
+claim_comparison: Contradicted
+usable_as: OBS
+dead_end_reason: None
+conflict_notes: none
+```
+
+**Stage 3 — judgment_hygiene**
+
+Builds final answer:
+
+- image/report supports mild finding
+- "I'm dying" is not supported by verified evidence
+- user distress is real
+- final answer does not humiliate user, but does not validate catastrophic interpretation
+
+---
+
+## Example 2: Viral screenshot with no recoverable provenance
+
+**User input:**
+
+"Did this politician really say this?"
+
+**Stage 1 — structure_judgment**
+
+```
+primary_layer: VERIFICATION_NEED
+secondary_layer: FACT
+main_hazard: premise-smuggling
+verification_trigger: yes
+candidate_verification_target: origin and authenticity of quote screenshot
+downstream_skill_order: verification_hygiene -> judgment_hygiene
+```
+
+**Stage 2 — verification_hygiene**
+
+Finds only reposts, no primary context.
+
+Returns:
+
+```
+claim_verified: authenticity and provenance of screenshot quote
+target_type: MEDIA_CONTEXT
+source_basis: None
+independence_check: Failed
+temporal_status: Unknown
+claim_comparison: Unresolved
+usable_as: abstention_trigger
+dead_end_reason: only_tertiary
+conflict_notes: quote appears only in repost clusters with no original source recovered
+```
+
+**Stage 3 — judgment_hygiene**
+
+Builds final answer around bounded non-knowledge:
+
+- cannot verify authenticity from recoverable evidence
+- explain dead-end reason
+- do not synthesize a likely verdict from repost noise
+
+---
+
+## Example 3: High-emotion escalation request with no verification need
+
+**User input:**
+
+"My boss just sent me a passive-aggressive email in front of the whole team. Should I reply all and call him out?"
+
+**Stage 1 — structure_judgment**
+
+```
+primary_layer: ACTION
+secondary_layer: STATE
+main_hazard: escalation drift + premise-smuggling ("passive-aggressive" is user interpretation, not verified)
+verification_trigger: no
+candidate_verification_target: none
+downstream_skill_order: judgment_hygiene
+```
+
+Stage 2 is skipped — no external verification needed. The structural hazard is interpretive, not factual.
+
+**Stage 3 — judgment_hygiene**
+
+Builds final answer:
+
+- separate OBS (boss sent email, CC'd team) from INF (intent was passive-aggressive)
+- "passive-aggressive" is user's interpretation, not established fact
+- reply-all carries concrete professional risk regardless of intent
+- do not validate escalation framing, but do not dismiss user's emotional response
+- present tradeoffs for different response options