po4yka/llm-wiki-skills/llm-wiki-model-policy

LLM-Wiki Model Policy

Goal

Create a practical model/data policy for ingest, query, lint, capture and publication workflows.

When to use

• The user asks which models or providers are allowed to process a given wiki source or folder.
• The user wants to decide what must stay local-only versus what can go to a cloud model.
• The user is setting up or revising the cheap/local vs. heavy/cloud model split across ingest, triage, synthesis, query, lint and embedding tasks.
• The user asks how to record model provenance (which model produced a page) in wiki frontmatter.
• The user needs escalation rules for low-confidence outputs or sensitive raw sources before publication.

Inputs

• Data sensitivity and domain.
• Current model providers and local models.
• Tasks: triage, ingest, query, synthesis, lint, embedding, reranking.
• Privacy, legal, cost and latency constraints.

Procedure

1. Classify data

Use:

public | internal | sensitive | regulated | unknown

Map folders and capture channels to these classes.

2. Classify tasks

Separate:

• capture cleanup;
• triage;
• source extraction;
• synthesis;
• query answering;
• linting;
• embeddings;
• reranking;
• publishing.

3. Assign model tiers

Create a matrix:

Use local-only defaults for sensitive or unknown material unless the user explicitly approves another policy.

4. Record provenance

Recommend frontmatter fields:

ai_model: ""
agent_version: ""
ai_confidence: 0.0
processed_at: YYYY-MM-DD
model_policy: local-only|cloud-allowed|redacted-cloud|unknown

5. Define escalation rules

Examples:

• cheap/local model for triage;
• stronger model for synthesis;
• human review for low confidence;
• no cloud for sensitive raw sources;
• redact before cloud when allowed.

6. Re-verify current provider claims

Browse official provider docs for current retention, privacy, pricing, model availability and API behavior when those facts matter.

Output

## Model policy summary
 
## Data classes
 
## Task/model matrix
 
## Local-only boundaries
 
## Frontmatter requirements
 
## Escalation rules
 
## Claims to re-verify

Safety gates

• Do not assume provider privacy or retention rules from memory.
• Do not route sensitive data to cloud models without explicit approval.
• Do not hide model provenance for generated pages.
• Do not optimize cost at the expense of trust boundaries.