Skip to content

Add LIP-0008: LLM tier discipline and dossier discipline#2

Open
danvoulez wants to merge 1 commit into
mainfrom
add-lip-0008-llm-tier-discipline
Open

Add LIP-0008: LLM tier discipline and dossier discipline#2
danvoulez wants to merge 1 commit into
mainfrom
add-lip-0008-llm-tier-discipline

Conversation

@danvoulez
Copy link
Copy Markdown
Contributor

Summary

Constitutional rule for how LLMs participate in the pipeline. Splits "LLM" into four distinct tiers, each with its own grammar and mandate, and defines Dossier as the only admissible shape crossing the Frontier boundary.

Sentence of stone

We do not ask Frontier to do Mini's job.
We do not ask Mini to do Operator's job.
We do not ask Operator to execute.
We do not ask Translator to authorize.
We do not ask any LLM to close evidence.

Each tier receives the smallest grammar it can honestly carry.
Frontier receives dossiers, not mess.

Four tiers

Tier Size Grammar Mandate
Mini 1.5–3.5B Operational classify, extract, propose tiny candidates
Operator 9–14B local Strong conduct session, route workorders
Translator 9–14B local + Frontier help Strong natural language → LogLine candidate
Frontier external Dossier-only yes/no on prepared case

What this PR does

  • Adds lips/LIP-0008-llm-tier-discipline-and-dossier-discipline.md
  • Updates README.md to list LIP-0008 as Proposed

What this PR does NOT do

  • No code patch (lands in PR 1 in constitutional-runtime)
  • No Dossier shape exists in code yet
  • No tier-aware admission in code yet

Connects to

  • LIP-0001 (canon hardening) — defines the doctrine this formalizes
  • LIP-0002 (if-doubt simulation) — translator can ghost when uncertain
  • LIP-0007 (three-layer hashing) — tuple/content/envelope = talent boundaries
  • constitutional-runtime/src/operational_grammar.rs — Mini tier ingress channel (already exists)
  • constitutional-runtime/src/strong_grammar.rs — Operator/Translator tier ingress channel (already exists)

🤖 Generated with Claude Code

Constitutional rule for how LLMs participate in the pipeline.

Four tiers, four mandates, three grammars:

  Mini       (1.5-3.5B)  Operational Grammar  classify, propose
  Operator   (9-14B)     Strong Grammar       conduct sessions
  Translator (9-14B)     Strong Grammar       natural -> LogLine candidate
  Frontier   (external)  Dossier Grammar      yes/no on prepared case

Core invariants:
- We do not ask Frontier to do Mini's job.
- We do not ask Mini to do Operator's job.
- We do not ask Operator to execute.
- We do not ask Translator to authorize.
- We do not ask any LLM to close evidence.
- Frontier receives dossiers, not mess.

Defines new types for code (PR 1 follow-up):
- LlmTier enum
- GrammarKind enum
- Dossier struct
- FrontierVerdict enum
- CapabilityManifest extension for tier/grammar admission

Connects to existing LogLine-Foundation work:
- Operational Grammar already exists in constitutional-runtime
- Strong Grammar already exists in constitutional-runtime
- Three-hash receipt (tuple/content/envelope) per LIP-0007 maps to
  three talent boundaries (human act / translator interpretation /
  transport)
- Entropy enters AdmissionContext as a measurable policy input

Minimal spike (GOLDEN-004) proves the 9-step receipt chain without
dispatch and without Frontier ever touching raw mess.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant