ConfigurationModel Policy Current

Model Policy Current

Model Policy Current

Canonical markdown page for this topic. Rendered reference version: model-policy-current.

Last Updated: 2026-04-01 Status: Active current-state policy note

This document separates the current model lanes used by Jorvis and OpenClaw.

It exists to prevent repeated confusion between:

  • chat models
  • embedding/search models
  • local fallback models
  • future policy changes

Current Repo Truth

Jorvis chat surfaces

Jorvis exposes multiple chat/model surfaces through Open WebUI and internal API routes. The exact public/runtime IDs can vary by registration surface, but current canonical references live in:

  • docs/operations/OPENWEBUI_MODEL_SETUP.md
  • deploy/docker-compose.prod.yml

OpenClaw chat lane

Current repo-backed OpenClaw chat defaults:

  • primary: google/gemini-flash-lite-latest
  • fallback chain: google/gemini-2.5-flash-lite → ollama/jorvis-4b-256k

Canonical sources:

  • deploy/openclaw/openclaw.json
  • deploy/openclaw/boot-openclaw.sh
  • analytics-platform/src/ai/model-registry/gemini-model-config.ts

Lossless Claw summarization lane

Lossless Claw uses a two-layer repo-backed model contract:

  • plugin-level primary summary model: gemini-3.1-flash-lite-preview
  • plugin-level provider: google
  • broader OpenClaw fallback chain: google/gemini-flash-lite-latest → google/gemini-2.5-flash-lite → ollama/jorvis-4b-256k

Canonical sources:

  • deploy/openclaw/normalize-openclaw-config.mjs
  • deploy/openclaw/openclaw.json
  • deploy/docker-compose.prod.yml

OpenClaw memory-search lane

Current repo-backed OpenClaw memory-search defaults:

  • provider: gemini
  • model: gemini-embedding-2-preview
  • fallback: none

This lane is for embeddings/search, not chat generation.

Current rollout note:

  • OpenClaw historical semantic memory was treated as disposable during the 2026-03-28 cutover.
  • The live memory-search store was intentionally reset/rebuilt onto gemini-embedding-2-preview.
  • The OpenClaw runtime baseline has now advanced to v2026.4.1, but the memory-search baseline itself remains gemini-embedding-2-preview.
  • gemini-embedding-001 is now a rollback-only baseline for this lane, not an inline mixed-space fallback.
  • The verified post-retry runtime and SSOT now supersede the older 2026-03-13 / gemini-embedding-001 mixed baseline for current-truth docs.

Local fallback lane

Jorvis also maintains an optional local llama.cpp / GGUF fallback lane for resiliency and offline-capable paths. It is not part of the mandatory first-boot or default shared-VM profile.

Current policy target for Jorvis Local when that lane is intentionally enabled:

  • base model family: gemma-4-E4B-it-GGUF
  • quantization: Q4_K_M
  • build requirement: abliterated variant
  • canonical wording: abliterated gemma-4-E4B-it-GGUF Q4_K_M

Canonical operational references:

  • docs/operations/OPENWEBUI_MODEL_SETUP.md
  • current runtime/config surfaces under deploy/

Important boundary:

  • local fallback policy must not be conflated with the OpenClaw embedding/search lane
  • local fallback re-enablement must use the profile-gated local-llm service, not an always-on default container
  • this is current policy truth for the optional local lane, not a claim that the lane is currently live on prod

Policy Rules

  1. Do not document chat model IDs as if they were memory-search embedding models.
  2. Do not document memory-search embedding IDs as if they were chat defaults.
  3. Do not silently update docs to a desired future policy unless runtime/config/code have been updated accordingly.
  4. When a future target differs from current repo truth, label it explicitly as future.

Resolved Normalization Note

The internal Jorvis search-config response contract was normalized to current OpenClaw runtime truth in merged PR #399.

Evidence:

  • docs/agent_ops/OUTBOX/task_openclaw_search_contract_normalization_20260306.md

This is resolved work, not an active follow-up.

Future Policy Directions (Not Current Truth)

Future OpenClaw expansion is expected to add clearer separation between:

  • OpenClaw user-assistant chat policy
  • OpenClaw admin-observer/admin-planner policy
  • Jorvis-side capability/control-plane policy

These are future directions only. They are not yet canonical runtime model policy.

  • docs/operations/OPENWEBUI_MODEL_SETUP.md
  • docs/operations/OPENCLAW_RUNTIME_OPERATIONS.md
  • docs/architecture/OPENCLAW_MEMORY_ARCHITECTURE.md