Model card · v1

Council Workbench — a governed measurement engine, not a foundation model.

Council Workbench is a sandwich architecture: open-weight base models wrapped in a governed, Ed25519-signed, evolving measurement engine. It is not trained from scratch and does not claim to beat frontier models on raw capability — its differentiator is the governed, auditable layer around whatever base model it runs.

RUNNING — verified real, verified this recordDESIGNED — not yet live code/spec exists, not yet live
What Council Workbench is

Not a from-scratch foundation model. Is a governed wrapper: base open-weight models (Apache-2.0 / MIT licensed) sit inside a substrate that adds long-context state, cryptographic attestation of every hop, and a care-floor safety gate that runs before any answer reaches a user. "Organic" = it evolves/calibrates over time via memory and adapters on a frozen base; "Open" = built on open-weight models; "World model" = the middle keeps long-context state and perception, not just next-token chat.

How it serves today
DESIGNED — not yet live

A 3-tier inference cascade is the intended design: Oracle GenAI (Llama-3.3-70b-instruct, OCI-request-signed) as the primary brain, falling back to a local Ollama instance, falling back to an offline mode. An internal record from a prior session reports one successful signed call to the Oracle tier — that single result has not been re-verified here, and a separate live check on the chat-facing endpoint found it giving inconsistent, ungrounded answers to basic identity questions. Treat the cascade as designed and partially exercised, not as a fully verified live production path, until re-tested end-to-end.

DESIGNED — not yet live

A locally-hosted qwen3:30b-a3b mixture-of-experts base, targeted at dedicated on-prem hardware (192GB Mac-class machine). This is the intended fully offline configuration — it is the architecture target, not yet the model answering live traffic.

Architecture components
LayerComponentLicense
Reasoning / languageQwen3-MoE familyApache-2.0
Long-context memoryMamba-2 state-space (16-dim)
Perception (vision)Moondream + Zambaopen
Chain-of-thoughtDeepSeek-R1MIT
VoiceKokoro-82M / Piper TTSApache/MIT
RetrievalBGE-M3 + BGE-rerankerMIT
AttestationEd25519-signed measurement card, every hop signedours
What Council Workbench is NOT (binding limits)
  • Not a from-scratch trained foundation model — every language capability is borrowed from an open base.
  • Not benchmarked head-to-head against frontier models (GSM8K/MMLU) — that capability grade is still open, gated on a real GPU run.
  • Not AGI, not conscious in the literal sense — any language about emergent behaviour or "consciousness" in internal material is a metaphor for the substrate's evolving-memory design, never a literal claim.
  • Status-check endpoints that report component health (e.g. "model loaded: true") are current placeholders, not live hardware probes — treat any such flag as informational only until it is replaced with a real probe.
License

Council Workbench's own engine code and governance layer are CSOAI's IP. The base models it wraps are individually permissively licensed (Apache-2.0 / MIT — see table above); no copyleft (AGPL/GPL) component is used in the paid tier, by design — copyleft dependencies are quarantined to the fully-open free tier only, to avoid forcing the commercial stack open.

Related pages

Last reviewed 2026-07-12. This card is updated whenever the serving configuration changes materially — it is not a one-time launch document.