Honest status ledger
The honest view of what is actually live, what is built-and-tested, what is design-only, and what is blocked on an open dependency. Nothing here is inflated — a capability with no shipped evidence is never listed as done.
Read this first. Cognitum MaaS is a governance / cost / safety / multi-tenancy platform, not a capability-frontier model. It rides commodity models at a measured Pareto cost-corner. The load-bearing claim is the cheap-vs-frontier finding — measured, not asserted: on everyday agentic work, cheap models are frontier-class at a fraction of the cost, with a real structural ceiling on the hard tail. The platform's value is orchestration as a cost lever, not an accuracy lever. We do not claim beyond-SOTA capability, a trained coordinator, or any quality lift over the underlying models.
Legend: LIVE proven against the live service BUILT implemented + tested, not yet on live infra DESIGN specified, not implemented OPEN a real gap or external dependency
Completions API — the serving primitive
| Capability | Status |
|---|---|
| OpenAI chat completions (streaming + JSON) — primary serving path | LIVE |
| Anthropic Messages API (streaming + JSON), with real tool-use translation | LIVE |
| Legacy OpenAI completions shape | BUILT |
Model listing (four cognitum-* aliases, no raw vendor ids) | LIVE |
| Liveness endpoint | LIVE |
| Multi-tenant key auth (hashed lookup → scopes) | BUILT |
| Tier router — intrinsic difficulty (low / mid / high) | LIVE |
| Post-generation escalation (non-streaming) | BUILT |
| Family-correct metering + usage ledger | BUILT |
| Reserve-and-Commit budget (sharded, account cap, fail-closed) | BUILT |
| Scatter-gather rate limiter + tenant-scoped idempotency | BUILT |
Darwin-Loop pods
| Capability | Status |
|---|---|
Pod spawn + lifecycle (pod = {domain × host × tier}) | LIVE |
| Real pod executor + real artifact capture | LIVE |
| Runaway-cap pause (bounded spend on budget deny) | LIVE |
| Read-only pod poll (multi-tenant projection, no cross-tenant leak) | BUILT |
| Host axis — native in-process runtime | LIVE |
| Host axis — CLI hosts (detect-and-skip until containers provisioned) | OPEN |
| Behavioral bench over real artifacts | LIVE |
| Approval gate — park + reservation-held proposal + verdict | BUILT |
| Continuous scheduler-trigger cadence wiring | OPEN |
Flywheel, deploy & economics
| Capability | Status |
|---|---|
| Assumed → measured calibration from the usage ledger | BUILT |
| Hot-reloadable routing params + tier pools (no redeploy) | BUILT |
| Learned-router bootstrap (trained head replacing the heuristic) | OPEN |
| Managed-container service on Google Cloud (live) | LIVE |
| CI — emulator-first, zero-cost | LIVE |
| CD — keyless (workload-identity), deploy-from-HEAD, image-only | LIVE |
| Budget-defense model (Reserve-and-Commit + caps) | BUILT |
| Subscription / billing lifecycle & checkout | DESIGN |
| Human-facing surface adapters + business templates | DESIGN |
Known open dependencies
- Tenant ownership field is provisioned by an upstream gateway, not this service; until then isolation is exercised on seeded fixtures.
- CLI hosts are detect-and-skip until their runtime containers are provisioned.
- Continuous pod cadence — the tick job is built and tested; wiring a recurring trigger is a deployment decision.
- One optional inflight-escalation dependency is unpublished upstream, so that path fail-closes to the standard one-shot routing — the operative, safe state.
- Learned router needs a per-request quality label the ledger does not yet carry.
One-paragraph summary
Cognitum MaaS is a multi-tenant, tiered, metered, budget-governed, dual-protocol Completions API — deployed live on Google Cloud — plus a Darwin-Loop autonomous agent-pod system with a human approval gate, all riding the same authenticated seam. It is honestly a governance / cost / safety / multi-tenancy platform on commodity models at a measured Pareto cost-corner — not a capability-frontier model. The serving path, real pod loop, runaway cap, native host, behavioral bench, flywheel calibration, security hardening, and the approval-gate seam are live or built; the billing lifecycle, the surface adapters and business templates, and the CLI hosts remain design-only, gated on a small set of open dependencies. The API and its source code are private and auth-gated.