Tranche 7 — Model Plane (tracking)

[mdheller]

Tracking epic for the Model Plane (Tiered On-Device Intelligence with Governed Distillation), spec v0.1 §XIV.

sourceos-spec

• T7-1 schemas/InferenceReceipt.json — done in feat(schemas): Model Plane inference-provenance contracts (T7-1/T7-3/T7-4) #240
• T7-2 schemas/ModelManifest.json + AdapterManifest.json — DECISION NEEDED (overlaps SourceOSModelCarryRef; name collides with AdapterDescriptor) → separate issue
• T7-3 schemas/EscalationDecision.json — done in feat(schemas): Model Plane inference-provenance contracts (T7-1/T7-3/T7-4) #240
• T7-4 schemas/AdapterPromotionDecision.json — done in feat(schemas): Model Plane inference-provenance contracts (T7-1/T7-3/T7-4) #240
• T7-5 add inference_provider to AgentPassport.agent_class — DECISION NEEDED (estate already models providers as separate InferenceProvider object) → separate issue
• T7-6 docs/adr/NNNN-model-plane.md — done as ADR-0015 in feat(schemas): Model Plane inference-provenance contracts (T7-1/T7-3/T7-4) #240

ontogenesis

• T7-7 Domains/model-plane/inference-provider.ttl
• T7-8 shapes/model-plane/residency-constraints.shacl — cross-document invariants schema if/then can't express (served==receipt reconciliation, ledger hash-chain continuity, base-digest match, biometric-never-crosses-socket)

source-os

• T7-9 docs/model-plane/architecture.md
• T7-10 profiles/model-plane/{constrained,standard,workstation,cluster-node}.nix
• T7-11 modules/modelplaned/ (catalog + residency)
• T7-12 modules/inferenced/ (serving)
• T7-13 modules/embeddingd/
• T7-14 modules/visiond/ — no network namespace
• T7-15 modules/distilld/ (governed distillation)
• T7-20 docs/seam-registry.md — add SEAM-014..017

agent-machine

• T7-16 Model plane capability grants + socket auth

mcp-a2a-zero-trust

• T7-17 inference_provider trust tier + residency forbidden-capability rules

synapseiq

• T7-18 Model plane client for embedding + classification enrichment

workstation-contracts

• T7-19 Conformance: receipt emission, unsigned-model refusal, residency enforcement

---

Reconciliation note: the Agent Machine / Model Carry family (InferenceProvider, ModelResidency, SourceOSModelCarryRef, ExternalModelProviderProfile, AgentMachineReceipt) already landed and must be reused across all downstream tranche items, not re-created.

Open design questions (spec §XIII): confidence-estimation method (the load-bearing escalation gate, self-reported); OQ3 base-model license under MIT/Apache-only — Llama/Gemma/Qwen-large fail it.

[mdheller]

✅ T7-1, T7-3, T7-4, T7-6 merged to main via #240 (squash 24f0044).

Copilot review addressed before merge:

• SEAM-015 teeth tightened on InferenceReceipt — off-device receipts now require a non-empty escalationChain and a non-null escalatedFrom (previously minItems never fired when the property was omitted).
• Unified EscalationDecision.grantRef → capabilityLeaseRef to match InferenceReceipt (one name for the urn:srcos:lease: reference).

Remaining in this repo: T7-2 (#242) and T7-5 (#243) — both decision-gated. Downstream tranche items (ontogenesis, source-os, agent-machine, mcp-a2a-zero-trust, synapseiq, workstation-contracts) unchanged.

[mdheller]

✅ T7-2 (#242) and T7-5 (#243) resolved and merged to main.

• feat(schemas): Model Plane store manifests (T7-2, resolves #242) #245 (squash) — ModelManifest + ModelAdapterManifest (SEAM-014/017 + SPDX-license teeth; named to avoid AdapterDescriptor collision; complements SourceOSModelCarryRef). ADR-0016.
• feat(schemas): Model Plane provider classification + review-hardening (resolves #243) #246 (squash) — T7-5 DECISION: add inference_provider to AgentPassport.agent_class? #243 resolved without a 6th agent_class: model daemon = system_core passport + InferenceProvider record, linked via new InferenceProvider.passportRef. Plus review-hardening teeth: ledger hash-chain (ledgerSeq now required on all three provenance types; non-genesis needs ledgerPrevHash) and a biometric hard boundary on EscalationDecision. ADR-0017.

Non-schema review findings (cross-doc SHACL, served==receipt conformance, spec-text contradictions) tracked in #247.

sourceos-spec Tranche 7 is now complete (T7-1..T7-6). Remaining T7 work is downstream: ontogenesis (T7-7/T7-8), source-os (T7-9..T7-15, T7-20), agent-machine (T7-16), mcp-a2a-zero-trust (T7-17), synapseiq (T7-18), workstation-contracts (T7-19).

[mdheller]

✅ T7-7 + T7-8 merged in ontogenesis: SocioProphet/ontogenesis#129 (squash).

• T7-7 Domains/model-plane/inference-provider.ttl — mp:InferenceProvider OWL class (upper:System; not a 6th agent_class — composed via mp:passportRef per ADR-0017).
• T7-8 shapes/model-plane/residency-constraints.shacl.ttl — spec §VIII residency teeth as SHACL: dataResidencyClass enum, emitsInferenceReceipt must be true, and sh:sparql cross-field rules (on_device⇒no-network, escalation⇒off-device, biometric⇒on-device+no-escalation). networkCapability/escalationPermitted required (no absence-dodge); emitsInferenceReceipt maxCount 1 (no multi-value dodge). 6 negative fixtures + lane validator + make validate-model-plane.

Copilot reviewed and both findings were remediated before merge. Cross-document manifest-digest binding remains tracked in #247 (needs sourceos-spec manifests as RDF).

Remaining downstream: source-os (T7-9..15, T7-20), agent-machine (T7-16), mcp-a2a-zero-trust (T7-17), synapseiq (T7-18), workstation-contracts (T7-19).

[mdheller]

✅ T7-19 merged in workstation-contracts: SociOS-Linux/workstation-contracts#50.

Model Plane conformance lane: vendored the sourceos-spec schemas (InferenceReceipt/EscalationDecision/ModelManifest/ModelAdapterManifest) into schemas/model-plane/, tools/validate_model_plane.py (type-dispatched), 5 good + 4 bad fixtures under conformance/model-plane/, wired into make validate. Teeth (both ways): off-device receipt w/o lease, unsigned manifest, adapter w/o baseModelDigest, and permitted escalation w/o lease all rejected. Copilot review addressed (exit-2 usage vs exit-1 conformance so a broken fixture can't masquerade as a rejection).

Tranche 7 status: sourceos-spec T7-1..6 ✅ · ontogenesis T7-7/8 ✅ · workstation-contracts T7-19 ✅. Remaining downstream: source-os (T7-9..15 daemons+profiles, T7-20 seam-registry), agent-machine T7-16, mcp-a2a-zero-trust T7-17, synapseiq T7-18.

[mdheller]

Model Plane reality audit (L1/L2/L3) — is it live?

• L1 Contract-real ✅ — gates re-run on origin/main: ontogenesis lane rejects 6/6 bad fixtures; workstation-contracts lane 5 good pass / 4 bad reject; sourceos-spec if/then teeth proven. Wired into CI and firing.
• L2 Consumption-real ⚠️ nominal — sole substantive referencer is agent-machine (docs + schema + fixtures only). Its docs/inference-local-steering.md states production is blocked: no repository artifact registers a provider. No real emitter of InferenceReceipt anywhere. (agent-machine also carries its own contracts/inference-provider.schema.json — a divergence from sourceos-spec to reconcile.)
• L3 Runtime-real ❌ — no modelplaned/inferenced/visiond module in source-os; nothing deployed.

Verdict: real as contracts, not yet live as a system. Making it live = the T7-9..T7-18 tier (source-os daemons = emitter, agent-machine grants = gate, synapseiq = first consumer). Acceptance for 'live' = a real hash-chained receipt in $LEDGER_PATH validated by the T7-19 lane as real output, not a fixture.

[mdheller]

Gap (b) — InferenceReceipt emitter built (workstation-contracts #51, merged). Reference emitter + hash-chained append-only ledger: real receipt construction, sha256 I/O hashing, ledgerPrevHash chain, schema-conformance against the vendored InferenceReceipt schema, chain verification incl. tamper-evidence (self-test wired into make validate; single-writer fcntl lock). Honest L3 boundary UNCHANGED: no local model runs (no weights: ollama blobs empty, no gguf, ollama CLI absent), so the self-test uses labelled synthetic completions — the machinery is real and proven, a real LLM completion is not. To go live: provision a small MIT/Apache gguf (llama.cpp is installed) → emit_receipt() with real digest+completion → the existing T7-19 lane validates real output. That single step remains the L3 unblock.

[mdheller]

✅ L3 IS REAL — the estate has real inference receipts. workstation-contracts #52 merged. tools/inferenced_shim.py runs a real llama.cpp completion against a real Apache-2.0 model (Qwen2.5-0.5B) and emits a real InferenceReceipt — baseModelDigest = real sha256 of the weights that ran, input/output hashes of the real prompt/completion. 3 real receipts committed at evidence/model-plane/live-inference-ledger.jsonl, hash-chained, and make validate (CI) proves each schema-conforms to the vendored InferenceReceipt schema + unbroken chain. No synthetic data; model weights not committed.

The reality-audit ledger is now: L1 ✅ (contracts, gates fire), L2 ✅ (real emitter output validated by the T7-19 lane's schema), L3 ✅ (real receipt from a real completion). Remaining is productionization: wire every estate inference call (noetica/prophet-platform/agent-machine) to emit through this path, and a persistent (non-CI-snapshot) ledger service — the emitter + shim are the reference they call.

[mdheller]

✅ Receipt-emitting gateway merged (workstation-contracts #53) — the primitive that makes the whole estate emit receipts. A drop-in OpenAI-compatible proxy: forwards /v1/chat/completions to the real provider and emits a real hash-chained InferenceReceipt (real digest, real I/O hashes, real usage token counts) per call. Proven live against llama-server (Qwen2.5-0.5B, Apache-2.0). Two reference emitters now exist — CLI shim (#52) + gateway (#53) — and make validate validates all evidence/model-plane/*-ledger.jsonl in CI (4 real receipts, schema-conformant, unbroken chains).

How the whole estate gets receipts: point each service's OPENAI_BASE_URL at the gateway (config, no code). That per-service rollout is the remaining step; the mechanism + proof are done.

[mdheller]

✅ Gateway now covers /v1/embeddings (workstation-contracts #54) — the estate's RAG/ingestion path emits receipts too. The gateway forwards both /v1/chat/completions and /v1/embeddings and emits real receipts (chat→task:chat.completion tier:T1; embedding→task:embedding tier:T0, outputTokenCount:null, output hash = the real vectors). Proven live: real llama-server nomic-embed-text-v1.5 (--embedding, Apache-2.0), real 768-dim /v1/embeddings → real embedding receipt. This is exactly what prophet-platform/apps/embeddings serves, so it's protocol-correct to route through the gateway.

Model Plane receipt coverage now spans the estate's two real inference shapes (chat + embeddings), both proven live + CI-validated (5 real receipts). Remaining = per-service deploy config: point each service's chat/embeddings endpoint at the gateway (noetica speaks Ollama-native, needs an Ollama-proto adapter or the sidecar; model-router is a spec, not a live proxy).

[mdheller]

✅ Gateway now speaks Ollama too (workstation-contracts #55) — unlocks the estate's Ollama-native consumers (noetica: OLLAMA_HOST). The gateway accepts /api/chat, /api/generate, /api/embeddings, /api/embed, translates to the OpenAI backend, emits the same real hash-chained receipt, and translates the response back to Ollama shape — so pointing a service's OLLAMA_HOST at the gateway gives it receipts, no code change. Proven live: /api/chat → Ollama {message,done,eval_count}; /api/embeddings → {embedding:[768]}, each emitting a receipt.

Gateway coverage complete for the estate's real protocols × shapes: OpenAI /v1 + Ollama /api, chat + embeddings. 7 real receipts across ledgers, CI-validated. Remaining = per-service deploy config (point OPENAI_BASE_URL / OLLAMA_HOST at the gateway) — a rollout/ops step, and a persistent ledger service vs the CI-snapshot evidence.

[mdheller]

✅ Deployable gateway + persistent ledger (workstation-contracts #56) — both remaining infra items done and proven. The gateway is now a container (Dockerfile + compose, runs non-root, /health) with a durable mounted ledger volume (SEAM-011 non-local ledger, vs the CI-snapshot evidence). Proven live in podman: built the image, ran it against a real llama-server, real /v1/chat/completions through the container → receipt written to the mounted volume.

Remaining = the per-service config rollout, with one safety prerequisite I found: repointing a service's OLLAMA_HOST at the gateway is only safe once the gateway passes through non-inference endpoints (noetica also calls /api/tags, /api/show, etc. — today those 404). So the honest next steps are: (1) add transparent passthrough for unhandled paths, then (2) the per-service config PRs (noetica OLLAMA_HOST, prophet-platform embeddings). I won't ship a redirect that breaks a service's model-listing.

[mdheller]

✅ Rollout enablers + first service wired. Gateway hardened into a true drop-in and the first service is wired:

• examples: add shared content/build/release family payloads #57 passthrough — non-inference paths (/api/tags, /api/show, /v1/models) forwarded transparently (upstream headers preserved), only inference paths emit receipts.
• docs(spec): place CompressionEvaluation in sourceos-spec contract lane #58 per-request model→digest — multi-model backends record the ACTUAL model per request (RECEIPT_GATEWAY_MODEL_DIGESTS map, authoritative; unknown model → forward, no fabricated receipt).
• Noetica overlay Noetica#601 — opt-in docker-compose.receipts.yml routes agent-machine's OLLAMA_HOST through the gateway (front of ollama) with real per-model digests + health-gated depends_on + pinned build. Ready + Copilot-addressed; blocked only by a pre-existing nextjs failure that also fails on main (unrelated agent-machine Ollama-fallback test — flagged separately, not force-merging over it).

Receipt mechanism is complete (OpenAI+Ollama, chat+embeddings, passthrough, multi-model, deployable+persistent-ledger). Remaining rollout: merge #601 once the pre-existing nextjs break is fixed; then prophet-platform embeddings (OpenAI /v1/embeddings — route-able the same way).