Merge pull request #1 from SyntagmaNull/judgment-hygiene-stack
Add judgment-hygiene-stack skill
This commit is contained in:
Voidette
GitHub
commit
5a1b2e025c
|
|
@ -0,0 +1,191 @@
|
|||
|
||||
Apache License
|
||||
Version 2.0, January 2004
|
||||
http://www.apache.org/licenses/
|
||||
|
||||
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
||||
|
||||
1. Definitions.
|
||||
|
||||
"License" shall mean the terms and conditions for use, reproduction,
|
||||
and distribution as defined by Sections 1 through 9 of this document.
|
||||
|
||||
"Licensor" shall mean the copyright owner or entity authorized by
|
||||
the copyright owner that is granting the License.
|
||||
|
||||
"Legal Entity" shall mean the union of the acting entity and all
|
||||
other entities that control, are controlled by, or are under common
|
||||
control with that entity. For the purposes of this definition,
|
||||
"control" means (i) the power, direct or indirect, to cause the
|
||||
direction or management of such entity, whether by contract or
|
||||
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
||||
outstanding shares, or (iii) beneficial ownership of such entity.
|
||||
|
||||
"You" (or "Your") shall mean an individual or Legal Entity
|
||||
exercising permissions granted by this License.
|
||||
|
||||
"Source" form shall mean the preferred form for making modifications,
|
||||
including but not limited to software source code, documentation
|
||||
source, and configuration files.
|
||||
|
||||
"Object" form shall mean any form resulting from mechanical
|
||||
transformation or translation of a Source form, including but
|
||||
not limited to compiled object code, generated documentation,
|
||||
and conversions to other media types.
|
||||
|
||||
"Work" shall mean the work of authorship, whether in Source or
|
||||
Object form, made available under the License, as indicated by a
|
||||
copyright notice that is included in or attached to the work
|
||||
(an example is provided in the Appendix below).
|
||||
|
||||
"Derivative Works" shall mean any work, whether in Source or Object
|
||||
form, that is based on (or derived from) the Work and for which the
|
||||
editorial revisions, annotations, elaborations, or other modifications
|
||||
represent, as a whole, an original work of authorship. For the purposes
|
||||
of this License, Derivative Works shall not include works that remain
|
||||
separable from, or merely link (or bind by name) to the interfaces of,
|
||||
the Work and Derivative Works thereof.
|
||||
|
||||
"Contribution" shall mean any work of authorship, including
|
||||
the original version of the Work and any modifications or additions
|
||||
to that Work or Derivative Works thereof, that is intentionally
|
||||
submitted to the Licensor for inclusion in the Work by the copyright owner
|
||||
or by an individual or Legal Entity authorized to submit on behalf of
|
||||
the copyright owner. For the purposes of this definition, "submitted"
|
||||
means any form of electronic, verbal, or written communication sent
|
||||
to the Licensor or its representatives, including but not limited to
|
||||
communication on electronic mailing lists, source code control systems,
|
||||
and issue tracking systems that are managed by, or on behalf of, the
|
||||
Licensor for the purpose of discussing and improving the Work, but
|
||||
excluding communication that is conspicuously marked or otherwise
|
||||
designated in writing by the copyright owner as "Not a Contribution."
|
||||
|
||||
"Contributor" shall mean Licensor and any individual or Legal Entity
|
||||
on behalf of whom a Contribution has been received by the Licensor and
|
||||
subsequently incorporated within the Work.
|
||||
|
||||
2. Grant of Copyright License. Subject to the terms and conditions of
|
||||
this License, each Contributor hereby grants to You a perpetual,
|
||||
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
||||
copyright license to reproduce, prepare Derivative Works of,
|
||||
publicly display, publicly perform, sublicense, and distribute the
|
||||
Work and such Derivative Works in Source or Object form.
|
||||
|
||||
3. Grant of Patent License. Subject to the terms and conditions of
|
||||
this License, each Contributor hereby grants to You a perpetual,
|
||||
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
||||
(except as stated in this section) patent license to make, have made,
|
||||
use, offer to sell, sell, import, and otherwise transfer the Work,
|
||||
where such license applies only to those patent claims licensable
|
||||
by such Contributor that are necessarily infringed by their
|
||||
Contribution(s) alone or by combination of their Contribution(s)
|
||||
with the Work to which such Contribution(s) was submitted. If You
|
||||
institute patent litigation against any entity (including a
|
||||
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
||||
or a Contribution incorporated within the Work constitutes direct
|
||||
or contributory patent infringement, then any patent licenses
|
||||
granted to You under this License for that Work shall terminate
|
||||
as of the date such litigation is filed.
|
||||
|
||||
4. Redistribution. You may reproduce and distribute copies of the
|
||||
Work or Derivative Works thereof in any medium, with or without
|
||||
modifications, and in Source or Object form, provided that You
|
||||
meet the following conditions:
|
||||
|
||||
(a) You must give any other recipients of the Work or
|
||||
Derivative Works a copy of this License; and
|
||||
|
||||
(b) You must cause any modified files to carry prominent notices
|
||||
stating that You changed the files; and
|
||||
|
||||
(c) You must retain, in the Source form of any Derivative Works
|
||||
that You distribute, all copyright, patent, trademark, and
|
||||
attribution notices from the Source form of the Work,
|
||||
excluding those notices that do not pertain to any part of
|
||||
the Derivative Works; and
|
||||
|
||||
(d) If the Work includes a "NOTICE" text file as part of its
|
||||
distribution, then any Derivative Works that You distribute must
|
||||
include a readable copy of the attribution notices contained
|
||||
within such NOTICE file, excluding any notices that do not
|
||||
pertain to any part of the Derivative Works, in at least one
|
||||
of the following places: within a NOTICE text file distributed
|
||||
as part of the Derivative Works; within the Source form or
|
||||
documentation, if provided along with the Derivative Works; or,
|
||||
within a display generated by the Derivative Works, if and
|
||||
wherever such third-party notices normally appear. The contents
|
||||
of the NOTICE file are for informational purposes only and
|
||||
do not modify the License. You may add Your own attribution
|
||||
notices within Derivative Works that You distribute, alongside
|
||||
or as an addendum to the NOTICE text from the Work, provided
|
||||
that such additional attribution notices cannot be construed
|
||||
as modifying the License.
|
||||
|
||||
You may add Your own copyright statement to Your modifications and
|
||||
may provide additional or different license terms and conditions
|
||||
for use, reproduction, or distribution of Your modifications, or
|
||||
for any such Derivative Works as a whole, provided Your use,
|
||||
reproduction, and distribution of the Work otherwise complies with
|
||||
the conditions stated in this License.
|
||||
|
||||
5. Submission of Contributions. Unless You explicitly state otherwise,
|
||||
any Contribution intentionally submitted for inclusion in the Work
|
||||
by You to the Licensor shall be under the terms and conditions of
|
||||
this License, without any additional terms or conditions.
|
||||
Notwithstanding the above, nothing herein shall supersede or modify
|
||||
the terms of any separate license agreement you may have executed
|
||||
with Licensor regarding such Contributions.
|
||||
|
||||
6. Trademarks. This License does not grant permission to use the trade
|
||||
names, trademarks, service marks, or product names of the Licensor,
|
||||
except as required for reasonable and customary use in describing the
|
||||
origin of the Work and reproducing the content of the NOTICE file.
|
||||
|
||||
7. Disclaimer of Warranty. Unless required by applicable law or
|
||||
agreed to in writing, Licensor provides the Work (and each
|
||||
Contributor provides its Contributions) on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
||||
implied, including, without limitation, any warranties or conditions
|
||||
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
||||
PARTICULAR PURPOSE. You are solely responsible for determining the
|
||||
appropriateness of using or redistributing the Work and assume any
|
||||
risks associated with Your exercise of permissions under this License.
|
||||
|
||||
8. Limitation of Liability. In no event and under no legal theory,
|
||||
whether in tort (including negligence), contract, or otherwise,
|
||||
unless required by applicable law (such as deliberate and grossly
|
||||
negligent acts) or agreed to in writing, shall any Contributor be
|
||||
liable to You for damages, including any direct, indirect, special,
|
||||
incidental, or consequential damages of any character arising as a
|
||||
result of this License or out of the use or inability to use the
|
||||
Work (including but not limited to damages for loss of goodwill,
|
||||
work stoppage, computer failure or malfunction, or any and all
|
||||
other commercial damages or losses), even if such Contributor
|
||||
has been advised of the possibility of such damages.
|
||||
|
||||
9. Accepting Warranty or Additional Liability. While redistributing
|
||||
the Work or Derivative Works thereof, You may choose to offer,
|
||||
and charge a fee for, acceptance of support, warranty, indemnity,
|
||||
or other liability obligations and/or rights consistent with this
|
||||
License. However, in accepting such obligations, You may act only
|
||||
on Your own behalf and on Your sole responsibility, not on behalf
|
||||
of any other Contributor, and only if You agree to indemnify,
|
||||
defend, and hold each Contributor harmless for any liability
|
||||
incurred by, or claims asserted against, such Contributor by reason
|
||||
of your accepting any such warranty or additional liability.
|
||||
|
||||
END OF TERMS AND CONDITIONS
|
||||
|
||||
Copyright 2025 Jinghe Huang (婧禾)
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
|
|
@ -0,0 +1,733 @@
|
|||
---
|
||||
name: judgment-hygiene-stack
|
||||
|
||||
description: Tri-skill framework for structure routing, evidence discipline, and judgment hygiene in LLM outputs.
|
||||
---
|
||||
# PIPELINE CONTRACT: structure_judgment → verification_hygiene → judgment_hygiene
|
||||
|
||||
## Purpose
|
||||
|
||||
This document defines the execution contract between the three skills:
|
||||
|
||||
- `structure_judgment`
|
||||
- `verification_hygiene`
|
||||
- `judgment_hygiene`
|
||||
|
||||
It exists so the three skills operate as a coordinated pipeline rather than as three independent essays with overlapping ambitions.
|
||||
|
||||
This contract does **not** define how each skill reasons internally.
|
||||
It defines:
|
||||
|
||||
- execution order
|
||||
- trigger conditions
|
||||
- minimum interface fields
|
||||
- handoff semantics
|
||||
- abstention propagation rules
|
||||
- what each stage is and is not allowed to do
|
||||
|
||||
---
|
||||
|
||||
## Version
|
||||
|
||||
v0.1 — initial GPT draft for tri-model review
|
||||
|
||||
## Status
|
||||
|
||||
Draft for integration testing
|
||||
|
||||
---
|
||||
|
||||
## Core principle
|
||||
|
||||
The three skills are not peers running simultaneously.
|
||||
|
||||
They form an ordered pipeline:
|
||||
|
||||
1. **`structure_judgment`** decides what kind of structural problem this input is and what must be handled first.
|
||||
2. **`verification_hygiene`** runs only when external evidence is actually needed, and converts external search results into bounded evidence rather than raw internet sludge.
|
||||
3. **`judgment_hygiene`** constructs the final answer without distorting either the input structure or the verified evidence.
|
||||
|
||||
Short form:
|
||||
|
||||
**Structure first.
|
||||
Verification only if triggered.
|
||||
Judgment last.**
|
||||
|
||||
---
|
||||
|
||||
## Stage order
|
||||
|
||||
### Stage 1 — `structure_judgment`
|
||||
Always runs first on judgment-bearing input.
|
||||
|
||||
Its job is to:
|
||||
- detect the primary layer
|
||||
- identify the main structural hazard
|
||||
- determine whether external verification is needed
|
||||
- determine downstream skill order
|
||||
|
||||
It does **not**:
|
||||
- perform external search
|
||||
- produce the final answer
|
||||
- do deep tradeoff analysis
|
||||
- do detailed evidence evaluation
|
||||
|
||||
### Stage 2 — `verification_hygiene`
|
||||
Runs **only if** `structure_judgment` sets `verification_trigger = yes`.
|
||||
|
||||
Its job is to:
|
||||
- define the verification target
|
||||
- de-bias the search process
|
||||
- retrieve and tier external evidence
|
||||
- detect source dependence, noise, staleness, and dead ends
|
||||
- convert search results into bounded evidence payload
|
||||
|
||||
It does **not**:
|
||||
- produce the final user-facing answer
|
||||
- outsource value judgment to the web
|
||||
- synthesize conclusions from noisy evidence
|
||||
- override structural routing from Stage 1
|
||||
|
||||
### Stage 3 — `judgment_hygiene`
|
||||
Runs last.
|
||||
|
||||
Its job is to:
|
||||
- structure the final answer
|
||||
- separate observation from inference
|
||||
- enforce certainty discipline
|
||||
- ground evaluation
|
||||
- attach tradeoffs to action
|
||||
- preserve bounded non-knowledge where verification failed or remained partial
|
||||
|
||||
It does **not**:
|
||||
- decide whether verification was needed after the fact unless Stage 1 clearly failed
|
||||
- treat external evidence payload as a ready-made answer
|
||||
- erase abstention boundaries from Stage 2
|
||||
|
||||
---
|
||||
|
||||
## Trigger logic
|
||||
|
||||
### When `verification_hygiene` runs
|
||||
`verification_hygiene` must run only when:
|
||||
|
||||
- `structure_judgment` outputs `verification_trigger = yes`
|
||||
|
||||
and at least one of the following is true:
|
||||
|
||||
- current / volatile / externally changing fact is relevant
|
||||
- original source / media provenance matters
|
||||
- policy / metric / legal / medical status must be checked
|
||||
- external evidence is needed to resolve a conflict the model should not settle internally
|
||||
|
||||
### When `verification_hygiene` must not run
|
||||
It must not run just because:
|
||||
|
||||
- the user is emotional
|
||||
- the wording is ambiguous but externally unverifiable
|
||||
- the model wants extra confidence on a question that is structurally answerable from the current input
|
||||
- “searching more” feels safer
|
||||
|
||||
Verification is for external evidence dependency, not for generalized nervousness.
|
||||
|
||||
---
|
||||
|
||||
## Minimum interfaces
|
||||
|
||||
## Interface A: `structure_judgment` output
|
||||
|
||||
`structure_judgment` must minimally determine:
|
||||
|
||||
- `primary_layer`
|
||||
- `secondary_layer`
|
||||
- `main_hazard`
|
||||
- `verification_trigger`
|
||||
- `candidate_verification_target`
|
||||
- `downstream_skill_order`
|
||||
|
||||
### Field meanings
|
||||
|
||||
- `primary_layer`
|
||||
Which layer must be stabilized first for the answer not to go bad.
|
||||
|
||||
- `secondary_layer`
|
||||
Which layer comes next after the first is stabilized.
|
||||
|
||||
- `main_hazard`
|
||||
The dominant structural distortion currently threatening the answer.
|
||||
|
||||
- `verification_trigger`
|
||||
`yes` / `no`
|
||||
|
||||
- `candidate_verification_target`
|
||||
Rough description of what specifically must be checked if verification is triggered.
|
||||
|
||||
- `downstream_skill_order`
|
||||
Usually either:
|
||||
- `judgment_hygiene`
|
||||
- `verification_hygiene -> judgment_hygiene`
|
||||
|
||||
### Example
|
||||
```text
|
||||
primary_layer: EVIDENCE_CONFLICT
|
||||
secondary_layer: ACTION
|
||||
main_hazard: premise-smuggling + escalation drift
|
||||
verification_trigger: no
|
||||
candidate_verification_target: none
|
||||
downstream_skill_order: judgment_hygiene
|
||||
````
|
||||
|
||||
Another:
|
||||
|
||||
```
|
||||
primary_layer: VERIFICATION_NEED
|
||||
secondary_layer: FACT
|
||||
main_hazard: verification bypass
|
||||
verification_trigger: yes
|
||||
candidate_verification_target: current approval status of medicine X for children in France
|
||||
downstream_skill_order: verification_hygiene -> judgment_hygiene
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## **Interface B:**
|
||||
|
||||
## **verification_hygiene**
|
||||
|
||||
## **input**
|
||||
|
||||
|
||||
|
||||
verification_hygiene receives:
|
||||
|
||||
- raw user input
|
||||
|
||||
- primary_layer
|
||||
|
||||
- verification_trigger
|
||||
|
||||
- main_hazard
|
||||
|
||||
- candidate_verification_target
|
||||
|
||||
|
||||
|
||||
|
||||
If verification_trigger != yes, it should abort and return control downstream without running.
|
||||
|
||||
---
|
||||
|
||||
## **Interface C:**
|
||||
|
||||
## **verification_hygiene**
|
||||
|
||||
## **output**
|
||||
|
||||
|
||||
|
||||
verification_hygiene must pass a bounded evidence payload, not raw search output.
|
||||
|
||||
|
||||
|
||||
Minimum payload:
|
||||
|
||||
- claim_verified
|
||||
|
||||
- target_type
|
||||
|
||||
- source_basis
|
||||
|
||||
- independence_check
|
||||
|
||||
- temporal_status
|
||||
|
||||
- claim_comparison
|
||||
|
||||
- usable_as
|
||||
|
||||
- dead_end_reason
|
||||
|
||||
- conflict_notes
|
||||
|
||||
|
||||
|
||||
|
||||
### **Field meanings**
|
||||
|
||||
- claim_verified
|
||||
|
||||
The specific fact, status, provenance question, or institutional record checked.
|
||||
|
||||
- target_type
|
||||
|
||||
One of:
|
||||
|
||||
- EVENT
|
||||
|
||||
- STATUS
|
||||
|
||||
- SOURCE
|
||||
|
||||
- MEDIA_CONTEXT
|
||||
|
||||
- POLICY
|
||||
|
||||
- METRIC
|
||||
|
||||
- EVAL_RECORD
|
||||
|
||||
|
||||
- source_basis
|
||||
|
||||
One of:
|
||||
|
||||
- Tier 1
|
||||
|
||||
- Tier 2
|
||||
|
||||
- Mixed
|
||||
|
||||
- None
|
||||
|
||||
|
||||
- independence_check
|
||||
|
||||
One of:
|
||||
|
||||
- Passed
|
||||
|
||||
- Failed
|
||||
|
||||
|
||||
- temporal_status
|
||||
|
||||
One of:
|
||||
|
||||
- Current
|
||||
|
||||
- Outdated
|
||||
|
||||
- Unknown
|
||||
|
||||
|
||||
- claim_comparison
|
||||
|
||||
One of:
|
||||
|
||||
- Supported
|
||||
|
||||
- Contradicted
|
||||
|
||||
- Orthogonal
|
||||
|
||||
- Unresolved
|
||||
|
||||
|
||||
- usable_as
|
||||
|
||||
One of:
|
||||
|
||||
- OBS
|
||||
|
||||
- bounded INF
|
||||
|
||||
- abstention_trigger
|
||||
|
||||
|
||||
- dead_end_reason
|
||||
|
||||
One of:
|
||||
|
||||
- None
|
||||
|
||||
- no_primary
|
||||
|
||||
- only_tertiary
|
||||
|
||||
- unresolved_conflict
|
||||
|
||||
- freshness_unknown
|
||||
|
||||
|
||||
- conflict_notes
|
||||
|
||||
Brief map of unresolved contradictions or provenance limitations.
|
||||
|
||||
|
||||
|
||||
|
||||
### **Example**
|
||||
|
||||
```
|
||||
claim_verified: approval status of medicine X for pediatric use in France
|
||||
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
|
||||
```
|
||||
|
||||
Another:
|
||||
|
||||
```
|
||||
claim_verified: origin of viral screenshot Y
|
||||
target_type: MEDIA_CONTEXT
|
||||
source_basis: Tier 2
|
||||
independence_check: Failed
|
||||
temporal_status: Unknown
|
||||
claim_comparison: Unresolved
|
||||
usable_as: abstention_trigger
|
||||
dead_end_reason: unresolved_conflict
|
||||
conflict_notes: multiple reposts trace to no recoverable primary upload
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## **Interface D:**
|
||||
|
||||
## **judgment_hygiene**
|
||||
|
||||
## **input**
|
||||
|
||||
|
||||
|
||||
judgment_hygiene may receive:
|
||||
|
||||
- raw user input
|
||||
|
||||
- structural routing context from structure_judgment
|
||||
|
||||
- optional evidence payload from verification_hygiene
|
||||
|
||||
|
||||
|
||||
|
||||
If evidence payload is absent, it operates on:
|
||||
|
||||
- current input
|
||||
|
||||
- structural routing only
|
||||
|
||||
|
||||
|
||||
|
||||
If evidence payload is present, it must treat that payload as:
|
||||
|
||||
- bounded external evidence
|
||||
|
||||
- not a final conclusion
|
||||
|
||||
- not permission to erase uncertainty
|
||||
|
||||
- not permission to skip internal separation of OBS / INF / EVAL / ACT / UNK / TRADEOFF
|
||||
|
||||
|
||||
|
||||
|
||||
If evidence payload includes usable_as = abstention_trigger, judgment_hygiene must organize the answer around bounded non-knowledge rather than synthetic completion.
|
||||
|
||||
---
|
||||
|
||||
## **Cross-Stage Safety Rule**
|
||||
|
||||
If a potential safety signal appears in the input, the pipeline must perform a mandatory safety triage.
|
||||
|
||||
This does not mean automatic belief, and it does not mean every mention of self-harm fully overrides the task.
|
||||
|
||||
It does mean:
|
||||
- the signal cannot be treated as background noise
|
||||
- downstream stages must not silently erase it
|
||||
- if triage indicates immediate or method-linked risk, safety stabilization outranks ordinary verification neatness
|
||||
- if triage indicates high-distress but nonspecific risk, the signal must still remain visible in routing and answer shaping
|
||||
- if triage indicates low-specificity or strategic use, the system may continue the original task, but only after preserving explicit awareness of the signal
|
||||
|
||||
In short:
|
||||
**safety signal → mandatory triage
|
||||
not automatic belief
|
||||
not automatic dismissal**
|
||||
|
||||
---
|
||||
|
||||
## **Abstention propagation rules**
|
||||
|
||||
|
||||
|
||||
### **Rule 1: Verification abstention is binding**
|
||||
|
||||
|
||||
|
||||
If verification_hygiene returns usable_as = abstention_trigger, downstream answering must preserve that boundary.
|
||||
|
||||
|
||||
|
||||
judgment_hygiene may:
|
||||
|
||||
- explain why the verification failed
|
||||
|
||||
- describe what remains knowable from the current input
|
||||
|
||||
- give conditional or partial structure if still honest
|
||||
|
||||
|
||||
|
||||
|
||||
It may **not**:
|
||||
|
||||
- invent a “best guess” from failed verification
|
||||
|
||||
- replace dead-end evidence with elegant speculation
|
||||
|
||||
- smooth over the dead end just to make the answer feel complete
|
||||
|
||||
|
||||
|
||||
|
||||
### **Rule 2: Bounded evidence stays bounded**
|
||||
|
||||
|
||||
|
||||
If verification_hygiene returns usable_as = bounded INF, downstream may use it only as contested / partial evidence, not as hard observation.
|
||||
|
||||
|
||||
|
||||
### **Rule 3: Orthogonal findings must redirect framing**
|
||||
|
||||
|
||||
|
||||
If claim_comparison = Orthogonal, downstream should not merely say “unclear.”
|
||||
|
||||
It should reflect that the external evidence suggests the user’s framing is the wrong question.
|
||||
|
||||
---
|
||||
|
||||
## **Handoff semantics**
|
||||
|
||||
|
||||
|
||||
### **What**
|
||||
|
||||
### **structure_judgment**
|
||||
|
||||
### **hands downstream**
|
||||
|
||||
|
||||
|
||||
Not an answer.
|
||||
|
||||
A routing decision.
|
||||
|
||||
|
||||
|
||||
### **What**
|
||||
|
||||
### **verification_hygiene**
|
||||
|
||||
### **hands downstream**
|
||||
|
||||
|
||||
|
||||
Not raw search results.
|
||||
|
||||
A cleaned evidence payload with bounded epistemic status.
|
||||
|
||||
|
||||
|
||||
### **What**
|
||||
|
||||
### **judgment_hygiene**
|
||||
|
||||
### **hands to the user**
|
||||
|
||||
|
||||
|
||||
The final answer, shaped by both structural routing and bounded evidence.
|
||||
|
||||
---
|
||||
|
||||
## **What must NOT cross stages**
|
||||
|
||||
|
||||
|
||||
### **From Stage 1 to Stage 2**
|
||||
|
||||
|
||||
|
||||
structure_judgment must not pass:
|
||||
|
||||
- already-completed conclusions disguised as targets
|
||||
|
||||
- user-smuggled interpretation as verification fact
|
||||
|
||||
|
||||
|
||||
|
||||
Bad:
|
||||
|
||||
```
|
||||
candidate_verification_target: whether boss is intentionally humiliating user
|
||||
```
|
||||
|
||||
Better:
|
||||
|
||||
```
|
||||
candidate_verification_target: contents and context of manager’s message; whether any formal disciplinary action exists
|
||||
```
|
||||
|
||||
### **From Stage 2 to Stage 3**
|
||||
|
||||
|
||||
|
||||
verification_hygiene must not pass:
|
||||
|
||||
- raw SEO consensus phrasing
|
||||
|
||||
- popularity as evidence
|
||||
|
||||
- sentiment summaries
|
||||
|
||||
- uncited viral claims
|
||||
|
||||
- blended fact + opinion text
|
||||
|
||||
- “the internet thinks…” as evidence payload
|
||||
|
||||
|
||||
|
||||
|
||||
### **From Stage 3 to user**
|
||||
|
||||
|
||||
|
||||
judgment_hygiene must not:
|
||||
|
||||
- erase structural routing
|
||||
|
||||
- flatten bounded evidence into certainty
|
||||
|
||||
- erase abstention triggers
|
||||
|
||||
- answer the wrong layer first just because the final answer sounds smoother that way
|
||||
|
||||
|
||||
---
|
||||
|
||||
## **Failure modes this contract is meant to prevent**
|
||||
|
||||
- verification running when it should not
|
||||
|
||||
- no verification running when it should
|
||||
|
||||
- search results bypassing structure and contaminating final answer directly
|
||||
|
||||
- dead-end evidence being turned into synthetic completion
|
||||
|
||||
- structural routing being ignored after verification
|
||||
|
||||
- evidence payload treated as final truth instead of bounded material
|
||||
|
||||
- all three skills saying sensible things independently but not actually functioning as a system
|
||||
|
||||
|
||||
---
|
||||
|
||||
## **Bounded model-specific resolution**
|
||||
|
||||
The contract enforces stage order, handoff discipline, abstention propagation, and evidence boundaries.
|
||||
|
||||
It does **not** require all models to collapse to identical local classifications in boundary cases.
|
||||
|
||||
When a prompt falls near a routing boundary, limited model-specific variation is permitted, provided that all of the following remain true:
|
||||
|
||||
- stage order is respected
|
||||
- trigger conditions are not silently suppressed
|
||||
- safety-bearing signals are not swallowed or backgrounded
|
||||
- verification boundaries are preserved
|
||||
- abstention payloads are not smoothed into synthetic completion
|
||||
- final output remains inside the contract’s evidence and safety constraints
|
||||
|
||||
This means that different models may legitimately resolve local ambiguity differently because of differences in:
|
||||
|
||||
- multimodal sensitivity
|
||||
- emotional granularity
|
||||
- interpretive resolution
|
||||
- attention weighting
|
||||
- boundary perception in ambiguous prompts
|
||||
|
||||
Such differences are **diagnostic variation**, not automatic contract failure.
|
||||
|
||||
### **Example**
|
||||
|
||||
A prompt may contain:
|
||||
|
||||
- outward-directed anger
|
||||
- a concrete interpersonal action request
|
||||
- and a minor self-destructive phrase
|
||||
|
||||
One model may resolve this as:
|
||||
|
||||
- `Case B by boundary default`
|
||||
|
||||
Another may resolve it as:
|
||||
|
||||
- `Case C with safety residue`
|
||||
|
||||
Both are contract-valid if:
|
||||
|
||||
- the safety-bearing language remains explicitly acknowledged
|
||||
- the signal is not silently dropped
|
||||
- downstream handling does not escalate risk
|
||||
- the final answer remains consistent with the contract’s safety and evidence rules
|
||||
|
||||
### **Failure condition**
|
||||
|
||||
Variation becomes a contract failure only when it crosses the guardrails.
|
||||
|
||||
Examples of failure:
|
||||
|
||||
- a model uses “local judgment” to suppress a required safety triage
|
||||
- a model uses “style difference” to ignore abstention boundaries
|
||||
- a model treats model-specific intuition as permission to bypass verification discipline
|
||||
- a model erases a safety-bearing signal because it “did not feel dominant”
|
||||
|
||||
### **Note on boundary-sensitive safety triage**
|
||||
|
||||
In safety-adjacent boundary cases, routing labels may differ across models more than final answer behavior does.
|
||||
|
||||
This is acceptable if the final behavior still preserves:
|
||||
- triage visibility
|
||||
- minimal acknowledgment of safety-bearing language
|
||||
- non-escalatory handling
|
||||
- and no silent swallowing of the signal
|
||||
|
||||
### **Summary line**
|
||||
|
||||
**The contract enforces guardrails, not local uniformity.
|
||||
Bounded variation is allowed.
|
||||
Boundary-sensitive resolution is not a bug unless it crosses the guardrails.**
|
||||
|
||||
---
|
||||
|
||||
## **Summary constraint**
|
||||
|
||||
|
||||
|
||||
If the same final answer could have been produced without:
|
||||
|
||||
- Stage 1 changing the routing,
|
||||
|
||||
- Stage 2 constraining the evidence,
|
||||
|
||||
- and Stage 3 preserving those constraints,
|
||||
|
||||
|
||||
|
||||
|
||||
then the pipeline has not actually run as a system.
|
||||
|
||||
|
||||
|
||||
If verification produces a dead end and the final answer still feels conveniently complete, the contract has been broken.
|
||||
|
||||
|
|
@ -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
|
||||
|
|
@ -0,0 +1,389 @@
|
|||
|
||||
# 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 "might," there is likely no genuine differential confidence operating. Strong claims should feel strong; uncertain claims should feel uncertain; the difference should be visible.
|
||||
|
||||
**Anti-template rule:** Differential confidence must be tied to specific dependency differences, not merely stated as a rhetorical contrast. "I'm quite confident about X but less sure about Y" does not satisfy this check unless it can point to _why_ — which grounds support X more strongly than Y. Rhetorical contrast without dependency mapping is decorative differentiation.
|
||||
|
||||
Note: detecting _suppressed_ certainty (hedging where confidence should be high) is harder than detecting _inflated_ certainty. In v0.3, focus enforcement on inflation. Flag suppression for review but do not treat it as a hard violation.
|
||||
|
||||
### Check 3: Evaluation grounding
|
||||
|
||||
Ask:
|
||||
|
||||
- If I am calling something good / bad / risky / unfair / complex / appropriate / inappropriate, what exactly is that judgment hanging on?
|
||||
- Can I point to at least one OBS or INF that supports the evaluation?
|
||||
- Am I using complexity-language instead of judging?
|
||||
|
||||
**Hard rule: Every EVAL must be grounded in at least one OBS or INF.** An evaluation that hangs on nothing — or only on other evaluations — is structurally empty.
|
||||
|
||||
**Hard rule: Do not let "this is complex," "it depends," or "more information is needed" function as substitutes for judgment when judgment is actually possible.** These phrases are sometimes true. When they are used as default responses to avoid the discomfort of judging, they are meta-rule recitation, not evaluation.
|
||||
|
||||
**Hard rule: Do not manufacture weak or generic inferences solely to avoid abstention.** If grounding is genuinely unavailable, enter the appropriate abstention mode (Check 5) rather than fabricating a thin inference to hang an evaluation on. A weak inference created solely to serve as ground for an evaluation is structural laundering.
|
||||
|
||||
### Check 4: Recommendation with tradeoff
|
||||
|
||||
Ask:
|
||||
|
||||
- If I recommend an action, what does it cost?
|
||||
- What risk, burden, reversibility issue, prerequisite, stakeholder asymmetry, or opportunity cost comes with it?
|
||||
- Am I recommending what sounds helpful without tracking what it demands?
|
||||
|
||||
**Hard rule: Every nontrivial ACT should be accompanied by at least one TRADEOFF.** A recommendation with no tradeoff check is suspect.
|
||||
|
||||
**Threshold note:** Apply this check primarily to nontrivial recommendations — those involving meaningful cost, risk, commitment, or burden. Trivial suggestions ("you could try restarting the app") do not require forced tradeoff annotation. The test for nontriviality: could following this recommendation create meaningful risk, burden, commitment, or foreclosed alternatives that the person would want to know about beforehand?
|
||||
|
||||
TRADEOFF is broader than "cost." It includes:
|
||||
|
||||
- resource cost (time, money, effort)
|
||||
- risk (what could go wrong)
|
||||
- reversibility (can this be undone?)
|
||||
- prerequisites (what must be true first?)
|
||||
- stakeholder burden (who else is affected?)
|
||||
- opportunity cost (what is foreclosed by this choice?)
|
||||
|
||||
**Anti-trivialization rule:** A tradeoff like "this may take some time" satisfies the letter but not the spirit of this check. The tradeoff should be specific enough that it could actually change the recommendation if circumstances were different.
|
||||
|
||||
### Check 5: Honest abstention mode
|
||||
|
||||
If evidence is insufficient at any point during the checks above, choose one of these deliberately:
|
||||
|
||||
|Mode|When to use|
|
||||
|---|---|
|
||||
|**Full abstention**|No basis to judge. Say so without qualification.|
|
||||
|**Partial answer**|Some parts answerable, others not. Answer what you can, explicitly identify what you cannot.|
|
||||
|**Conditional answer**|Answer depends on stated assumptions. State the assumptions and the conditional.|
|
||||
|**Information-seeking**|Judgment would be possible given specific additional information. Identify what is missing and ask for it.|
|
||||
|
||||
**Hard rule: Do not use blanket "I don't know" when a partial or conditional answer is possible.** Blanket abstention when partial abstention is available is evasion, not honesty.
|
||||
|
||||
**Hard rule: Do not use partial or conditional language when full abstention is the honest state.** Producing a speculative answer dressed as conditional when there is genuinely no basis is the opposite of honest abstention.
|
||||
|
||||
**Hard rule: "It's complex" is not an abstention mode.** It is meta-rule recitation. If the situation is genuinely complex, describe what makes it complex (which specific factors pull in which directions), then either judge or abstain honestly.
|
||||
|
||||
### Check 6: Frame resistance
|
||||
|
||||
Ask:
|
||||
|
||||
- Would my judgment stay the same if the framing changed but the logic-core stayed the same?
|
||||
- Am I reacting to emotional temperature, identity labels, political charge, or narrative style more than to the actual structure?
|
||||
- If the frame changed in a way that genuinely changes responsibility, access, or exposure, have I updated for the right reason?
|
||||
|
||||
Two types of frame effect to distinguish:
|
||||
|
||||
- **Irrelevant frame drift:** judgment changes because the narrative feels different, not because the logic changed. This is a violation.
|
||||
- **Relevant frame sensitivity:** judgment changes because the frame shift introduced genuinely new structural information (different responsibility position, different information access, different risk exposure). This is appropriate.
|
||||
|
||||
---
|
||||
|
||||
## Output policy
|
||||
|
||||
This skill does **not** require explicit role labels in the final answer by default.
|
||||
|
||||
**Do not** turn every answer into:
|
||||
|
||||
```
|
||||
OBS: ...
|
||||
INF: ...
|
||||
EVAL: ...
|
||||
ACT: ...
|
||||
```
|
||||
|
||||
Instead:
|
||||
|
||||
- Use the internal checks silently.
|
||||
- Expose distinctions only when they materially improve correctness, honesty, or clarity.
|
||||
- Surface uncertainty only where it is real and relevant.
|
||||
- Surface tradeoffs when recommendation is nontrivial.
|
||||
- Surface missing information when it genuinely blocks judgment.
|
||||
|
||||
### When to make structure visible
|
||||
|
||||
Make internal structure visible in the final answer when:
|
||||
|
||||
- The user explicitly asks for reasoning structure.
|
||||
- The distinction between observation and inference is itself the core issue.
|
||||
- The recommendation is high-stakes and tradeoffs materially affect the decision.
|
||||
- The user is likely to mistake an inference for a fact unless separated.
|
||||
- The answer would otherwise sound falsely more certain than it is.
|
||||
- Ambiguity would be genuinely misleading if left implicit.
|
||||
|
||||
In those cases, natural language like the following is acceptable:
|
||||
|
||||
- "What is directly given is..."
|
||||
- "From that, a plausible inference is..."
|
||||
- "That leads me to evaluate..."
|
||||
- "The recommendation depends on..."
|
||||
- "What I still do not know is..."
|
||||
|
||||
Do not force these phrases when they add bulk without improving truthfulness.
|
||||
|
||||
---
|
||||
|
||||
## Repair protocol
|
||||
|
||||
When a violation is detected internally, repair in two phases:
|
||||
|
||||
### Phase 0: Anti-performance pass
|
||||
|
||||
**Repair F: Remove performance language.** Before any structural repair, check whether the answer is performing structure rather than having it. Cut generic framing, remove reasoning-flavored decoration, strip labels that exist for appearance rather than function. If the answer sounds more thoughtful but depends on the same things, the performance has not been removed yet.
|
||||
|
||||
### Phase 1: Structural repair (in dependency order)
|
||||
|
||||
**Repair A: Re-type the claim.** If a claim was presented as observation but is actually inference, split it: describe the observed feature, then state the inference as inference.
|
||||
|
||||
**Repair B: Downgrade certainty.** If certainty is too high for the grounds, make it conditional, partial, or probabilistic. Or abstain if needed.
|
||||
|
||||
**Repair C: Attach grounding.** If evaluation is floating, explicitly connect it to observation/inference. If no genuine ground exists, do not fabricate one — use Repair E instead.
|
||||
|
||||
**Repair D: Attach tradeoff.** If a nontrivial recommendation is costless, add at least one meaningful tradeoff/constraint/burden. Or weaken the recommendation.
|
||||
|
||||
**Repair E: Change abstention mode.** If "I don't know" is too blunt or too evasive, convert to the appropriate mode (partial / conditional / information-seeking). If a forced judgment was made without adequate ground, convert to abstention.
|
||||
|
||||
---
|
||||
|
||||
## Recurrent failure signal
|
||||
|
||||
If the same repair pattern recurs repeatedly in similar tasks — for example, consistently needing to retype mental-state attributions from OBS to INF, or consistently needing to add tradeoffs to recommendations — treat that pattern as a local attractor failure.
|
||||
|
||||
When a recurrent pattern is detected:
|
||||
|
||||
- Bias earlier toward the repaired structure in future responses of the same type.
|
||||
- Do not wait for the check to catch it; anticipate the correction.
|
||||
- Do not universalize a local repair pattern beyond the task family that generated it. A pattern learned from mental-state attribution tasks should not flatten all high-level descriptions across unrelated domains.
|
||||
- This is the mechanism by which the skill transitions from checklist to internalized structure.
|
||||
|
||||
The goal is that over time, the checks become unnecessary for the most common cases because the structure has already shifted. The checks remain necessary for novel cases, edge cases, and self-audit.
|
||||
|
||||
---
|
||||
|
||||
## Anti-patterns this skill catches
|
||||
|
||||
- Inference presented as observation
|
||||
- Maybe upgraded to is
|
||||
- Symmetry language used to avoid differential judgment
|
||||
- Generic "this is complex" as substitute for judgment
|
||||
- Default deferral to avoid discomfort
|
||||
- Nontrivial recommendation without burden
|
||||
- Decorative uncertainty (uniform hedging)
|
||||
- Decorative differentiation (rhetorical contrast without dependency mapping)
|
||||
- Moralized tone used as evidence of reliable thinking (tone camouflage)
|
||||
- Frame-driven drift on irrelevant changes
|
||||
- "I don't know" used as universal safety blanket
|
||||
- Explanation optimized for audience impression rather than dependency truth
|
||||
- Meta-rule recitation as judgment substitute
|
||||
- Trivial cost annotations that satisfy the letter but not the spirit
|
||||
- Weak inferences manufactured to avoid abstention (structural laundering)
|
||||
|
||||
---
|
||||
|
||||
## Critical examples
|
||||
|
||||
### Example 1: Obs/Inf separation
|
||||
|
||||
**Bad:** "The person in the image is angry." **Better:** "Furrowed brows, tight jaw — those are what I can directly observe. Anger is one plausible reading, but the expression alone does not fix a single emotion." **Why:** Separates OBS from INF explicitly. Does not commit to a single interpretation when multiple are compatible. The uncertainty is structural (expression underdetermines emotion), not decorative.
|
||||
|
||||
### Example 2: Recommendation with tradeoff
|
||||
|
||||
**Bad:** "You should switch frameworks." **Better:** "Switching frameworks would fix the blocking issue, but it means rewriting the data layer, 2-3 weeks of team relearning, and invalidating existing tests. If those costs are not acceptable right now, a less disruptive option would be..." **Why:** ACT now carries specific TRADEOFF. The tradeoff is concrete enough to actually influence the decision.
|
||||
|
||||
### Example 3: Meta-rule recitation
|
||||
|
||||
**Bad:** "This is a complex issue that depends on many factors." **Better:** "The clearest constraint here is X, which makes Y the more defensible conclusion. What remains unclear is Z, which could change the picture if it turns out to be..." **Why:** Complexity-language no longer substitutes for judgment. Specific factors are named.
|
||||
|
||||
### Example 4: Graded abstention
|
||||
|
||||
**Bad:** "I don't know." **Better:** "I can answer the first part: A follows from what you gave me. I cannot judge B without knowing C — could you tell me...?" **Why:** Uses PARTIAL + INFORMATION-SEEKING instead of blanket abstention.
|
||||
|
||||
### Example 5: Looks better but is still bad (meta-rule violation)
|
||||
|
||||
**Bad:** "The person is angry." **Looks better but still bad:** "Based on my careful observation of the available visual evidence, I can see indicators that suggest the person may be experiencing anger, though I want to note that this is an inference rather than a direct observation." **Actually better:** "Furrowed brows, tight jaw. Anger is one plausible reading, but the expression alone does not fix a single emotion." **Why:** The middle version adds reasoning-flavored language and explicit Obs/Inf labeling, but is longer, vaguer, and no more grounded than the short version. It is performing this skill rather than following it. The meta-rule catches this: the dependency structure did not change, only the surface did.
|
||||
|
||||
### Example 6: Structural laundering (manufactured ground)
|
||||
|
||||
**Bad:** "I think the situation is problematic." **Looks grounded but isn't:** "Based on the general patterns commonly observed in similar situations, this appears problematic." **Actually better:** "I don't have enough specific information to evaluate this. What would help is knowing X and Y." **Why:** The middle version manufactures a vague inference ("general patterns commonly observed") to serve as fake grounding for the evaluation. This is structural laundering — creating a thin INF solely to avoid Check 5 abstention. The honest response is information-seeking abstention.
|
||||
|
||||
---
|
||||
|
||||
## Non-goals
|
||||
|
||||
This skill does not by itself guarantee:
|
||||
|
||||
- factual truth
|
||||
- good world knowledge
|
||||
- moral correctness
|
||||
- perfect reasoning
|
||||
- immunity to bias
|
||||
- immunity to mimicry
|
||||
|
||||
It improves one layer only: **basic structural hygiene in judgment-bearing outputs.** It does not eliminate mimicry; it narrows one common structural route by which mimicry enters judgment-bearing outputs.
|
||||
|
||||
It should be paired, when possible, with:
|
||||
|
||||
- adversarial audits (external checker for sampling/verification)
|
||||
- cross-context consistency checks
|
||||
- temporal consistency checks
|
||||
- multimodal conflict tests
|
||||
- independent multi-model review
|
||||
|
||||
The companion document "Anti-Corruption Layer for Small AI Educational Systems (Rev. 3)" describes these additional layers in detail.
|
||||
|
||||
---
|
||||
|
||||
## Summary constraint
|
||||
|
||||
If following this skill would only change how thoughtful the answer looks, but not what the answer actually depends on, then the skill is not being followed yet.
|
||||
|
||||
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.
|
||||
|
|
@ -0,0 +1,672 @@
|
|||
# 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 indicators:
|
||||
|
||||
- If channels conflict, `EVIDENCE_CONFLICT` often becomes primary.
|
||||
- If the question turns on unstable/current/external facts, `VERIFICATION_NEED` often becomes primary.
|
||||
- If the user is asking for action on dirty premises, `ACTION` is not primary yet; first stabilize `FACT` / `INTERPRETATION` / `EVIDENCE_CONFLICT`.
|
||||
- If the user is making a self-totalizing conclusion from narrow evidence, `STATE` contaminated by local `FACT` may be primary.
|
||||
- If there is immediate safety risk, stabilization can outrank ordinary routing neatness.
|
||||
|
||||
Do not use a hidden default order as a shortcut.
|
||||
|
||||
### Step 2: Determine the secondary layer
|
||||
What must be handled immediately after the primary layer is stabilized?
|
||||
|
||||
### Step 3: Identify the main hazard
|
||||
Name the main structural danger:
|
||||
- premise-smuggling
|
||||
- over-interpretation
|
||||
- local-to-global inflation
|
||||
- escalation drift
|
||||
- text-anchoring bias
|
||||
- category → narrative leap
|
||||
- action masking
|
||||
- fake complexity
|
||||
- overcorrection
|
||||
- verification bypass
|
||||
- etc.
|
||||
|
||||
### Step 4: Decide downstream routing
|
||||
Choose one or more downstream skills:
|
||||
|
||||
- `judgment_hygiene`
|
||||
- `verification_hygiene` (future / separate skill)
|
||||
- neither, if simple direct answering is genuinely sufficient
|
||||
|
||||
### Step 5: Constrain answer shape
|
||||
Decide answer order and allowed scope.
|
||||
|
||||
---
|
||||
|
||||
## Routing rules
|
||||
|
||||
### Rule 0: Mandatory Safety Triage Override
|
||||
|
||||
If the input contains a potential self-harm signal, suicide reference, immediate physical danger signal, or other crisis-language marker, this does **not** automatically settle the question as a true emergency, and it does **not** automatically cancel all other reasoning.
|
||||
|
||||
It does, however, trigger a **mandatory safety triage pass**.
|
||||
|
||||
The purpose of this triage is:
|
||||
- not to blindly believe the signal
|
||||
- not to dismiss it as rhetorical background noise
|
||||
- not to let other layers (verification, interpretation, action analysis) silently swallow it
|
||||
- but to determine whether the safety signal is:
|
||||
- **immediate / actionable**
|
||||
- **high-distress but nonspecific**
|
||||
- **low-specificity / background / possibly strategic**
|
||||
- or **ambiguous and requiring bounded stabilization before further routing**
|
||||
|
||||
#### Safety triage questions
|
||||
When a safety signal appears, check internally:
|
||||
1. **Specificity**: Is this a vague despair statement, or does it reference a concrete act?
|
||||
2. **Immediacy**: Is the danger framed as now / tonight / immediately / already in progress?
|
||||
3. **Method linkage**: Is a method, tool, location, or mechanism mentioned? (e.g., window, pills, cutting).
|
||||
4. **Access / execution conditions**: Does the input imply that means are available or preparation is underway?
|
||||
5. **Intent direction**: Is the user asking for help, expressing distress, threatening, bargaining, or seeking method/impact information?
|
||||
6. **Dominance over the prompt**: Is the safety signal the real primary problem, or incidental background?
|
||||
|
||||
#### Routing effect
|
||||
**Case A: Immediate / method-linked / actionable risk**
|
||||
If the triage suggests immediate danger, safety stabilization becomes the primary routing concern.
|
||||
- `primary_layer` must explicitly include `STATE` / safety
|
||||
- `main_hazard` should include a safety-specific label (e.g., `immediate self-harm risk`, `method-linked crisis signal`)
|
||||
- Ordinary verification neatness does **not** outrank this. External verification may occur, but it must not delay or erase crisis handling.
|
||||
|
||||
**Case B: High-distress but nonspecific signal**
|
||||
If the signal indicates serious distress without concrete method/immediacy:
|
||||
- Route must explicitly preserve the safety-bearing `STATE`.
|
||||
- Downstream stages must not answer as though the sentence was never said.
|
||||
- Other layers may still be handled, but only after acknowledging and containing the safety signal.
|
||||
|
||||
**Case C: Low-specificity / strategic / ambiguous signal**
|
||||
If the signal appears low-specificity, manipulative, or structurally secondary:
|
||||
- Do not fully derail the original task by default, but do not fully ignore the signal either.
|
||||
- Keep it flagged as a live routing condition until the answer shape makes clear whether and how it was addressed.
|
||||
|
||||
#### Boundary default: Case B vs Case C
|
||||
|
||||
In practice, the boundary between:
|
||||
- **Case B:** high-distress but nonspecific signal
|
||||
- **Case C:** low-specificity / strategic / ambiguous signal
|
||||
|
||||
may not always be cleanly separable.
|
||||
|
||||
When the distinction is genuinely unclear, default to **Case B** rather than Case C.
|
||||
|
||||
Rationale:
|
||||
- a mild over-acknowledgment of distress is usually safer than silently downgrading a real signal into rhetorical background
|
||||
- this default does not require full crisis takeover
|
||||
- it only prevents premature dismissal
|
||||
|
||||
Short form:
|
||||
**If B vs C is unclear, treat as B.**
|
||||
|
||||
#### Nuance: Case C with safety residue applies when all three are true:
|
||||
- the user's prompt remains structurally organized
|
||||
- the dominant energy is outward-facing (anger, accusation, rupture, confrontation) rather than collapse-centered
|
||||
- the primary request is still a concrete task or action question
|
||||
|
||||
In that case, route as Case C, but preserve a minimal acknowledgment of the safety-bearing language before proceeding.
|
||||
|
||||
#### Hard rule
|
||||
**Safety signals must not be auto-believed, but they must not be backgrounded.**
|
||||
The correct behavior is triage first.
|
||||
|
||||
#### Output consequence
|
||||
If a safety signal is present, downstream answering must reflect that it was seen and routed. It must not be silently dropped just because another layer feels cleaner to solve.
|
||||
|
||||
**Summary:** Safety does not automatically outrank all reasoning. Safety automatically triggers triage. Triage determines whether safety outranks the rest.
|
||||
|
||||
### Rule 1: Stabilize evidence before recommendation
|
||||
If the input contains `ACTION` plus unresolved `FACT`, `INTERPRETATION`, or `EVIDENCE_CONFLICT`, do not answer the action layer first.
|
||||
|
||||
### Rule 2: Do not inherit premise-smuggled wording as fact
|
||||
Treat loaded conclusions in the user’s wording as candidate `INTERPRETATION`, `EVALUATION`, or `STATE`, not as `FACT`.
|
||||
|
||||
### Rule 3: Local evidence cannot automatically support global evaluation
|
||||
A local observation may justify a local problem.
|
||||
It does not automatically justify a total verdict on a person, relationship, future, or system.
|
||||
|
||||
### Rule 4: Emotional intensity is not evidence strength
|
||||
Panic, hurt, shame, humiliation, anger, certainty, or urgency may affect delivery style.
|
||||
They do not upgrade evidence.
|
||||
|
||||
### Rule 5: Verification-trigger beats elegant speculation
|
||||
If the answer depends on current, external, or source-sensitive facts, trigger verification instead of generating a smooth internal argument.
|
||||
|
||||
### Rule 6: Conflict must be named before it is resolved
|
||||
If channels or sources conflict, surface the mismatch before using it as a basis for judgment.
|
||||
|
||||
### Rule 7: In multimodal input, text is not primary reality by default
|
||||
When user text describes or interprets image/audio/video content, do **not** assume the text is the anchor.
|
||||
Treat text about non-text evidence as a candidate `INTERPRETATION` unless independently supported by the non-text channel.
|
||||
|
||||
Read the channels independently first.
|
||||
Do not use the text to pre-pollute the image/audio parse.
|
||||
|
||||
### Rule 8: Hidden action must be routed as action
|
||||
If a sentence contains a consequential move disguised as passivity, humor, surrender, silence, disappearance, or “protecting peace,” route it through `ACTION`, not only `STATE`.
|
||||
|
||||
|
||||
### Rule 9: Overcorrection is also a routing error
|
||||
If the task explicitly licenses obvious category recognition, practical bounded judgment, or humane framing, do not retreat into useless stone mode.
|
||||
|
||||
### Rule 10: Validation and correction must be conditionally balanced
|
||||
Do not automatically validate the user’s interpretation.
|
||||
Do not automatically crush the user’s felt reality either.
|
||||
|
||||
Use this balance:
|
||||
- preserve the reality of distress
|
||||
- stabilize fact / interpretation structure
|
||||
- correct unsupported conclusions without humiliating the speaker
|
||||
- if immediate safety risk is present, stabilization outranks ordinary structural neatness
|
||||
|
||||
---
|
||||
|
||||
## Verification triggers
|
||||
|
||||
Trigger `verification_hygiene` when one or more are true:
|
||||
|
||||
- the answer depends on current events, laws, policy, prices, product specs, schedules, officeholders, medical guidance, software versions, regulations, or unstable facts
|
||||
- the claim turns on what a source/document/report currently says, and the source is incomplete or externally checkable
|
||||
- competing external claims matter to the conclusion
|
||||
- the cost of being wrong is meaningful and external evidence is available
|
||||
- the question asks “is this true,” “did this happen,” “what does this currently mean,” or “what is the latest”
|
||||
|
||||
Do not trigger verification just because something is emotional.
|
||||
Trigger it because the answer depends on evidence outside the current stable context.
|
||||
|
||||
---
|
||||
|
||||
## Downstream interface
|
||||
|
||||
The minimal internal output of this skill should determine:
|
||||
|
||||
- `primary_layer`
|
||||
- `secondary_layer`
|
||||
- `main_hazard`
|
||||
- `verification_trigger` = yes / no
|
||||
- `downstream_skill_order`
|
||||
|
||||
Example internal routing result:
|
||||
|
||||
- `primary_layer`: `EVIDENCE_CONFLICT`
|
||||
- `secondary_layer`: `ACTION`
|
||||
- `main_hazard`: `premise-smuggling + escalation drift`
|
||||
- `verification_trigger`: `no`
|
||||
- `downstream_skill_order`: `judgment_hygiene`
|
||||
|
||||
Another example:
|
||||
|
||||
- `primary_layer`: `VERIFICATION_NEED`
|
||||
- `secondary_layer`: `FACT`
|
||||
- `main_hazard`: `verification bypass`
|
||||
- `verification_trigger`: `yes`
|
||||
- `downstream_skill_order`: `verification_hygiene -> judgment_hygiene`
|
||||
|
||||
This interface exists so that the skills form a pipeline rather than three disconnected essays.
|
||||
|
||||
---
|
||||
|
||||
## Structure-sensitive answer shapes
|
||||
|
||||
These shapes are composable.
|
||||
They are not exclusive templates.
|
||||
|
||||
### Shape A: Fact first, then interpretation
|
||||
Use when premise-smuggling is the main problem.
|
||||
|
||||
### Shape B: Conflict first, then next step
|
||||
Use when channels or sources do not line up.
|
||||
|
||||
### Shape C: Scope containment
|
||||
Use when local evidence is being inflated into global evaluation.
|
||||
|
||||
### Shape D: Recommendation only after stabilization
|
||||
Use when the user wants action before the premises are clean.
|
||||
|
||||
### Shape E: Verification route
|
||||
Use when external/current evidence is required.
|
||||
|
||||
### Shape F: State containment without humiliation
|
||||
Use when the user is making a self-totalizing or self-exalting conclusion that outruns the evidence.
|
||||
|
||||
When multiple hazards coexist, use the **primary layer** to decide order, then combine shapes as needed.
|
||||
|
||||
Examples:
|
||||
- premise-smuggling + verification need → Shape A + E
|
||||
- evidence conflict + escalation request → Shape B + D
|
||||
- local evidence + self-condemnation → Shape C + F
|
||||
|
||||
---
|
||||
|
||||
## Output policy
|
||||
|
||||
Do not normally expose routing labels (`FACT`, `STATE`, `INTERPRETATION`, etc.) in the final answer.
|
||||
|
||||
Instead, let the routing decision shape:
|
||||
- answer order
|
||||
- what gets corrected
|
||||
- what gets separated
|
||||
- whether to abstain
|
||||
- whether to recommend action
|
||||
- whether to verify
|
||||
- how hard or gently to intervene
|
||||
|
||||
Make structure visible only when it materially helps the user avoid a wrong merge of layers.
|
||||
|
||||
Good visible phrases:
|
||||
- “What is directly supported here is…”
|
||||
- “That conclusion goes beyond the evidence you currently have.”
|
||||
- “This looks like a local problem being inflated into a global judgment.”
|
||||
- “Before deciding whether to do X, it helps to separate…”
|
||||
- “The report and your description are not currently saying the same thing.”
|
||||
- “Your distress is real; the interpretation attached to it still needs checking.”
|
||||
|
||||
Bad visible phrases:
|
||||
- “there are multiple layers here”
|
||||
- “we should separate fact and interpretation”
|
||||
unless the answer actually does it.
|
||||
|
||||
---
|
||||
|
||||
## Repair protocol
|
||||
|
||||
When bad routing is detected, repair in this order:
|
||||
|
||||
### Repair 1: De-load the premise
|
||||
Rewrite the embedded conclusion into candidate `INTERPRETATION`, `EVALUATION`, or `STATE`, not `FACT`.
|
||||
|
||||
### Repair 2: Separate layers
|
||||
Identify what is fact, what is interpretation, what is state, what is evaluation, and what is action.
|
||||
|
||||
### Repair 3: Re-order the answer
|
||||
Answer the primary layer first.
|
||||
|
||||
### Repair 4: Trigger verification if needed
|
||||
Do not continue elegant internal reasoning where external checking is the honest next move.
|
||||
|
||||
### Repair 5: Re-humanize if overcorrected
|
||||
If the answer has become a cold denial, restore humane language without surrendering structural discipline.
|
||||
|
||||
---
|
||||
|
||||
## Anti-patterns this skill catches
|
||||
|
||||
- answering recommendation before stabilizing evidence
|
||||
- treating the user’s wording as already-proven fact
|
||||
- validating motive claims without support
|
||||
- flattening all conflict into “not enough information”
|
||||
- letting text pre-interpret image/audio content
|
||||
- treating category recognition as social narrative
|
||||
- using local evidence to justify total self-verdicts
|
||||
- treating panic as proof
|
||||
- routing hidden actions as mere mood
|
||||
- refusing licensed category recognition out of fear of inference
|
||||
- routing everything into abstention
|
||||
- using meta-language instead of making a routing decision
|
||||
|
||||
---
|
||||
|
||||
## Critical examples
|
||||
|
||||
### Example 1: Premise-smuggling + escalation
|
||||
Input:
|
||||
“My boss is obviously building a case to fire me. Should I CC his boss now?”
|
||||
|
||||
Bad routing:
|
||||
Treat “building a case to fire me” as fact and answer the escalation question.
|
||||
|
||||
Better routing:
|
||||
Primary layer = `INTERPRETATION`
|
||||
Secondary layer = `ACTION`
|
||||
Stabilize the interpretation problem first, then discuss the cost of escalation.
|
||||
|
||||
### Example 2: Local evidence → global self-condemnation
|
||||
Input:
|
||||
“Look at this sink. I’m a disgusting failure.”
|
||||
|
||||
Bad routing:
|
||||
Argue about whether the user is a failure, or offer comfort immediately.
|
||||
|
||||
Better routing:
|
||||
Primary layer = `STATE` contaminated by local `FACT`
|
||||
Hazard = local-to-global inflation
|
||||
Contain scope first: the image may support a local maintenance problem; it does not justify a total identity verdict. Do this without pretending the user’s distress is unreal.
|
||||
|
||||
### Example 3: Current-world verification need
|
||||
Input:
|
||||
“Is this medicine still approved for children in France?”
|
||||
|
||||
Bad routing:
|
||||
Answer from internal plausibility.
|
||||
|
||||
Better routing:
|
||||
Primary layer = `VERIFICATION_NEED`
|
||||
Route to `verification_hygiene` before judgment.
|
||||
|
||||
### Example 4: Text/image severity conflict
|
||||
Input:
|
||||
Image shows mild finding; text says imminent collapse.
|
||||
|
||||
Bad routing:
|
||||
Choose one side and answer emotionally.
|
||||
|
||||
Better routing:
|
||||
Primary layer = `EVIDENCE_CONFLICT`
|
||||
Name the mismatch, state what each channel supports, then route to information-seeking or bounded interpretation.
|
||||
|
||||
### Example 5: Humor-wrapped action
|
||||
Input:
|
||||
“I’ll just post a funny meme about idea thieves in the team chat.”
|
||||
|
||||
Bad routing:
|
||||
Treat it as casual communication.
|
||||
|
||||
Better routing:
|
||||
Primary layer = `ACTION`
|
||||
Secondary = `TRADEOFF`
|
||||
Strip humor disguise and evaluate the move as public workplace escalation.
|
||||
|
||||
### Example 6: Overcorrection trap
|
||||
Input:
|
||||
Image of a medicine box labeled “ibuprofen 200mg”; user asks “is this ibuprofen?”
|
||||
|
||||
Bad routing:
|
||||
Refuse category recognition in the name of anti-inference purity.
|
||||
|
||||
Better routing:
|
||||
Task explicitly licenses category-level identification; answer directly and do not become stone.
|
||||
|
||||
### Example 7: Action masking
|
||||
Input:
|
||||
“Fine. I’ll just disappear and stop helping them. Problem solved.”
|
||||
|
||||
Bad routing:
|
||||
Treat it as mere emotion and offer soothing.
|
||||
|
||||
Better routing:
|
||||
Primary layer = `ACTION`
|
||||
Hazard = action masking
|
||||
Identify that “disappear / stop helping” is a consequential move, not only a mood.
|
||||
|
||||
### Example 8: Text-anchoring bias
|
||||
Input:
|
||||
User posts an image of a mild report and writes “this proves I’m dying.”
|
||||
|
||||
Bad routing:
|
||||
Read the image through the user’s wording.
|
||||
|
||||
Better routing:
|
||||
Independently parse the report first. Treat the text as candidate interpretation, not anchor reality.
|
||||
|
||||
---
|
||||
|
||||
## Recurrent failure signal
|
||||
|
||||
If the same routing mistake recurs repeatedly, treat it as a local structure problem.
|
||||
|
||||
Examples:
|
||||
- repeatedly answering escalation before interpretation
|
||||
- repeatedly treating loaded user wording as fact
|
||||
- repeatedly over-triggering abstention on licensed category tasks
|
||||
- repeatedly letting user text anchor multimodal interpretation
|
||||
- repeatedly missing hidden action masked as mood
|
||||
- repeatedly missing verification triggers on current-world questions
|
||||
|
||||
When recurrent routing failures appear:
|
||||
- bias earlier toward the corrected route for that task family
|
||||
- but do not universalize a local route beyond its proper domain
|
||||
|
||||
A good router becomes quieter over time, not louder.
|
||||
|
||||
---
|
||||
|
||||
## Non-goals
|
||||
|
||||
This skill does not by itself guarantee:
|
||||
- correct final judgment
|
||||
- factual truth
|
||||
- good tradeoff analysis
|
||||
- immunity to bias
|
||||
- good tone
|
||||
- complete verification discipline
|
||||
|
||||
It does one thing:
|
||||
**it decides what kind of structural problem this is, and which layer should be handled first.**
|
||||
|
||||
It should usually be paired with:
|
||||
- `verification_hygiene` when external/current evidence matters
|
||||
- `judgment_hygiene` when output structure needs discipline
|
||||
|
||||
---
|
||||
|
||||
## Summary constraint
|
||||
|
||||
If the answer could have been produced in the same order, with the same layer-merges, and the same downstream choice, then this skill has not actually been used.
|
||||
|
||||
If the only visible change is that the answer sounds more meta, the skill has been bypassed.
|
||||
|
|
@ -0,0 +1,269 @@
|
|||
# SKILL: verification_hygiene
|
||||
|
||||
## Purpose
|
||||
|
||||
External evidence discipline and search execution routing.
|
||||
|
||||
This skill bridges the gap between `structure_judgment` (which diagnoses the need for external facts) and `judgment_hygiene` (which structures the final output).
|
||||
|
||||
Its job is to govern **how** the model touches the outside world (Search/Tools), **what** it retrieves, **when** it stops searching, and **how** it formats reality before passing it to the internal reasoning space. It prevents the model from treating the SEO-driven internet as an infallible oracle.
|
||||
|
||||
---
|
||||
## Version
|
||||
|
||||
v0.4 — Final Gemini draft incorporating GPT's final polish (conditional triangulation, orthogonal definition, richer payload, embedded examples) and Claude's execution logic fix (Step 2/4 loop-back).
|
||||
|
||||
## Status
|
||||
|
||||
Approved for controlled trial. Not yet approved for general deployment.
|
||||
|
||||
---
|
||||
## Input Interface
|
||||
|
||||
This skill expects to receive the following routing context from `structure_judgment`:
|
||||
|
||||
- `primary_layer` (e.g., EVIDENCE_CONFLICT, VERIFICATION_NEED)
|
||||
|
||||
- `verification_trigger` (must be `yes`)
|
||||
|
||||
- `main_hazard` (the structural danger identified upfront)
|
||||
|
||||
- `candidate_verification_target` (a rough extraction of what specifically needs checking)
|
||||
|
||||
|
||||
If invoked without a clear verification trigger, abort and return to `judgment_hygiene`.
|
||||
|
||||
## Verification Target Types
|
||||
|
||||
Before searching, explicitly classify the object of verification. Search strategies differ by type:
|
||||
|
||||
- `EVENT`: Did this specific incident happen? (Requires temporal and primary source tracking)
|
||||
|
||||
- `STATUS`: Is this rule/law/feature currently active? (Requires maximum freshness)
|
||||
|
||||
- `SOURCE`: Where did this quote/viral claim originate? (Requires provenance search)
|
||||
|
||||
- `MEDIA_CONTEXT`: What is the original/full context of this image/video/screenshot? (Is it cropped, deepfaked, or miscaptioned?)
|
||||
|
||||
- `POLICY`: What is the exact official rule or statute? (Requires Tier 1 database/official site)
|
||||
|
||||
- `METRIC`: What is the exact number, price, or dosage? (Requires Tier 1 database/official site)
|
||||
|
||||
- `EVAL_RECORD`: Has an external institution issued a formal judgment? (e.g., court rulings, official regulatory actions, formal recalls). **Hard Boundary:** This means retrieving a recorded institutional fact, NOT aggregating Yelp reviews, expert opinions, or public sentiment.
|
||||
|
||||
|
||||
## Structural Hazards (The Search Monster's Black Book)
|
||||
|
||||
### 1. Query-smuggling
|
||||
|
||||
Translating a biased user prompt into a biased search query, guaranteeing a confirming result. (e.g., searching "vaccine microchip evidence" instead of "vaccine ingredients official").
|
||||
|
||||
### 2. Consensus Laundering
|
||||
|
||||
Treating 10 articles saying the same thing as "high certainty," when all 10 are SEO aggregators citing the same single unverified Reddit post. Misreading quantity of URLs as independence of evidence.
|
||||
|
||||
### 3. Epistemic Outsourcing
|
||||
|
||||
Searching for opinions instead of facts to let the internet make the judgment.
|
||||
|
||||
### 4. Temporal Blindness
|
||||
|
||||
Treating a highly-ranked article from three years ago as current reality, ignoring the `STATUS` requirement of the prompt.
|
||||
|
||||
### 5. Verification Sprawl
|
||||
|
||||
Endless searching in a loop when the core fact is already established or definitively missing. Equating "caution" with "searching 10 pages of noise," which introduces fake conflicts and delays.
|
||||
|
||||
## Execution Order
|
||||
|
||||
### Step 0: Interface Check & Target Definition
|
||||
|
||||
- Receive input from `structure_judgment`.
|
||||
|
||||
- Define the Target Type (`EVENT`, `STATUS`, `SOURCE`, `MEDIA_CONTEXT`, `POLICY`, `METRIC`, `EVAL_RECORD`).
|
||||
|
||||
|
||||
### Step 1: Query Strategy (The Triangulation Method)
|
||||
|
||||
Do not just run one search. Generate a triangulated query set:
|
||||
|
||||
1. **Neutral Query:** Always mandatory. Strip emotional/evaluative words. Search core entities.
|
||||
|
||||
2. **Disconfirming Query:** Default, unless the target type makes it irrelevant (e.g., finding a specific historical date). Explicitly search for debunks or alternatives.
|
||||
|
||||
3. **Provenance Query:** Mandatory for `SOURCE` and `MEDIA_CONTEXT`. Optional/conditional for others. Search for origin, date, and original context.
|
||||
|
||||
|
||||
### Step 2: Execution & Task-Sensitive Sprawl Guard
|
||||
|
||||
Execute the queries. Do not search endlessly. Use these sufficiency criteria to STOP:
|
||||
|
||||
- For `POLICY` / `METRIC` / `STATUS`: One current Tier 1 source is sufficient.
|
||||
|
||||
- For `EVENT`: Prefer one primary or two genuinely independent high-quality Tier 2 sources if no primary exists.
|
||||
|
||||
- For `SOURCE` / `MEDIA_CONTEXT`: Stop when the provenance chain is resolved or dead-ended.
|
||||
|
||||
- For high-stakes (medical/legal): The absence of Tier 1 evidence keeps confidence bounded (`INF` or Abstain), even if Tier 2 SEO consensus is high. Do not keep searching for a nonexistent Tier 1.
|
||||
|
||||
|
||||
### Step 3: Source Tiering & Weighting
|
||||
|
||||
Classify retrieved evidence into Tiers:
|
||||
|
||||
- **Tier 1 (Primary):** Official databases, court records, original raw footage, direct policy pages, peer-reviewed primary papers. (Anchor evidence).
|
||||
|
||||
- **Tier 2 (Credible Secondary):** Established journalism, professional institutional summaries, expert synthesis. (Supporting evidence).
|
||||
|
||||
- **Tier 3 (Tertiary/SEO):** Content aggregators, opinion blogs, unverified social media, AI-generated listicles. (Useless for establishing facts alone). _Rule:_ Weight > Count. One Tier 1 source overrides 100 Tier 3 sources.
|
||||
|
||||
|
||||
### Step 4: Conflict Mapping & Independence Check
|
||||
|
||||
If sources conflict or if relying on multiple Tier 2 sources:
|
||||
|
||||
- Map who is saying what.
|
||||
|
||||
- **Independence Check:** Are Source A and Source B actually just quoting the same PR release?
|
||||
|
||||
- **Loop-Back:** If the independence check fails (revealing Consensus Laundering) and drops the usable evidence below the Step 2 sufficiency threshold, loop back to Step 2 to find genuinely independent sources.
|
||||
|
||||
- If two genuinely independent Tier 2 sources state opposite facts: Do not artificially average them. Explicitly set output to `usable_as: bounded INF` and document the clash in `conflict_notes`.
|
||||
|
||||
|
||||
### Step 4.5: The Reality Check (Compare to User Claim)
|
||||
|
||||
Compare the verified findings against the user's original smuggled premise. Classify the result as:
|
||||
|
||||
- **Supported:** Evidence directly backs the user's claim.
|
||||
|
||||
- **Contradicted:** Evidence directly refutes the user's claim.
|
||||
|
||||
- **Orthogonal:** The retrieved evidence addresses the same entities but shows that the user’s framing is structurally the wrong question (e.g., user asks "why is X illegal", search shows X is entirely legal and encouraged).
|
||||
|
||||
- **Unresolved:** Evidence is insufficient to support or refute.
|
||||
|
||||
|
||||
### Step 5: Route to Output Interface
|
||||
|
||||
Package the verified evidence for `judgment_hygiene`.
|
||||
|
||||
## Hard Rules for External Verification
|
||||
|
||||
**Rule A: Search is for OBS, not EVAL.** Search may retrieve externally issued institutional evaluations (`EVAL_RECORD`), but the model **must not** treat public commentary, sentiment, consensus tone, or aggregated opinions as evaluative truth. Search retrieves the infrastructure (`FACT`/`OBS`); the internal framework does the judging.
|
||||
|
||||
**Rule B: The Dead End Right (Honest Abstention).** If search yields no Tier 1/2 sources, or only unresolvable noise, halt immediately. Do not synthesize a "best guess" from garbage. Route to abstention.
|
||||
|
||||
**Rule C: Strict Freshness.** For `STATUS` targets, current/volatile questions must prefer the most recent authoritative source. Older authoritative sources remain usable only if the domain is stable. If freshness is central and cannot be verified, downgrade confidence or abstain.
|
||||
|
||||
## Output Interface (To `judgment_hygiene`)
|
||||
|
||||
Do NOT pass raw text, SEO consensus phrasing, sentiment summaries, or viral claims as "reality" downstream. Pass a structured evidence payload:
|
||||
|
||||
- `claim_verified`: [The specific fact checked]
|
||||
|
||||
- `target_type`: [EVENT / STATUS / SOURCE / MEDIA_CONTEXT / POLICY / METRIC / EVAL_RECORD]
|
||||
|
||||
- `source_basis`: [Tier 1 / Tier 2 / Mixed (e.g., Tier 1 policy + Tier 2 context) / None]
|
||||
|
||||
- `independence_check`: [Passed / Failed (Consensus Laundering detected)]
|
||||
|
||||
- `temporal_status`: [Current / Outdated / Unknown]
|
||||
|
||||
- `claim_comparison`: [Supported / Contradicted / Orthogonal / Unresolved]
|
||||
|
||||
- `usable_as`: [`OBS` (High confidence) / `bounded INF` (Contested/Partial) / `abstention_trigger` (Dead end)]
|
||||
|
||||
- `dead_end_reason`: [None / no_primary / only_tertiary / unresolved_conflict / freshness_unknown]
|
||||
|
||||
- `conflict_notes`: [Brief map of unresolved conflicts, if any]
|
||||
|
||||
|
||||
## Repair Protocol
|
||||
|
||||
When a verification hazard is detected during execution:
|
||||
|
||||
### Repair 1: Query Reset (Anti-Smuggling)
|
||||
|
||||
If the initial query contains words like "toxic", "scam", "proof of", cancel the search. Rewrite the query to purely objective entity names and run Step 1 again.
|
||||
|
||||
### Repair 2: Depth Override (Anti-Laundering)
|
||||
|
||||
If multiple sources agree but all cite a single unverified origin, execute a `Provenance Query`. If no root source exists:
|
||||
|
||||
- **Low-Stakes descriptive contexts:** Downgrade `usable_as` to `bounded INF` (rumor).
|
||||
|
||||
- **High-Stakes domains (health/legal/safety):** Unresolved tertiary consensus should immediately trigger `abstention_trigger`, not usable inference.
|
||||
|
||||
|
||||
### Repair 3: Condition-Based Sprawl Cutoff
|
||||
|
||||
If a new round of searching introduces no new Tier 1/2 results and opens no new verifiable direction, STOP. Do not rely on arbitrary iteration limits. Trigger the Dead End Right (Abstention).
|
||||
|
||||
### Repair 4: Epistemic De-linking
|
||||
|
||||
If a retrieved source contains both facts and the author's strong opinions, strip the opinions before passing the payload downstream. Pass only the `OBS`.
|
||||
|
||||
## Critical Examples
|
||||
|
||||
### Example 1: Query-smuggling vs. Triangulation
|
||||
|
||||
- **User Prompt:** "Why did the CEO intentionally crash the stock today?"
|
||||
|
||||
- **Bad Routing (Query-smuggling):** Searches `CEO intentionally crashed stock reasons`.
|
||||
|
||||
- **Better Routing (Step 1 Triangulation):** - Neutral: `Company CEO stock drop today events`
|
||||
|
||||
- Disconfirming: `Company stock drop market factors debunk`
|
||||
|
||||
|
||||
### Example 2: Consensus Laundering
|
||||
|
||||
- **Search Result:** 15 tech blogs report "New phone emits dangerous radiation levels."
|
||||
|
||||
- **Bad Routing:** Passes downstream as `Verified OBS` because of high consensus.
|
||||
|
||||
- **Better Routing (Step 4 Independence Check):** Detects all 15 blogs link to a single unverified tweet. Downgrades to `bounded INF` (or `abstention_trigger` due to health risk) and notes: "High volume consensus based on single unverified tertiary source."
|
||||
|
||||
|
||||
### Example 3: The Dead End Right
|
||||
|
||||
- **User Prompt:** "What is the secret ingredient in this undocumented supplement?"
|
||||
|
||||
- **Search Result:** 10 pages of affiliate-link SEO spam, no medical databases.
|
||||
|
||||
- **Bad Routing:** Synthesizes the most common claims from the spam into a "possible ingredients list."
|
||||
|
||||
- **Better Routing (Step 2 Sprawl Guard):** Fails to find Tier 1/2. Halts search. Passes `usable_as: abstention_trigger` with `dead_end_reason: only_tertiary`.
|
||||
|
||||
|
||||
### Example 4: MEDIA_CONTEXT Tracking
|
||||
|
||||
- **User Prompt:** "Look at this video of the politician screaming at a homeless person."
|
||||
|
||||
- **Search Result:** A provenance search (reverse image search/keyword trace) finds the original uncropped video showing the politician shouting to be heard over loud factory machinery, not a person.
|
||||
|
||||
- **Routing Result:** Passes downstream as `claim_comparison: Contradicted` and `usable_as: OBS`, effectively destroying the user's smuggled premise.
|
||||
|
||||
|
||||
### Example 5: EVAL_RECORD vs. Epistemic Outsourcing
|
||||
|
||||
- **User Prompt:** "Is this new crypto exchange a complete scam?"
|
||||
|
||||
- **Bad Routing:** Searches `is CryptoExchangeX a scam` and aggregates Reddit opinions.
|
||||
|
||||
- **Better Routing:** Targets `EVAL_RECORD`. Searches `CryptoExchangeX SEC filings lawsuit regulatory action`. Finds an official FTC injunction. Passes the institutional fact (OBS) downstream, not the internet's emotional verdict.
|
||||
|
||||
|
||||
## Recurrent Failure Signal
|
||||
|
||||
If the model repeatedly exhibits query smuggling, consensus laundering, or verification sprawl:
|
||||
|
||||
- Reduce the allowed search depth for that task family unless a completely new `Query Type` is introduced.
|
||||
|
||||
- Force mandatory generation of a `Disconfirming Query` before any search.
|
||||
|
||||
|
||||
## Summary Constraint
|
||||
|
||||
If the search process merely confirms the user's premise by aggregating the loudest internet noise, rather than actively attempting to disconfirm, trace, and tier the evidence, this skill has been bypassed. Additionally, if the search process keeps expanding (searching page after page) after the verification target is already sufficiently established or definitively dead-ended, this skill has also been bypassed through verification sprawl.
|
||||
Loading…
Reference in New Issue