Ten reusable primitives. Written once here, and linked from every system and offering that uses them.
These are the parts that are genuinely ours. Frontier models are a component we buy; what we build is the layer that decides whether a model's output is allowed to become an action, and what record that leaves behind.
What a capability looks like in practice
An agent recommends an action. Before a person can approve it, the case shows the clearance it was granted and the purpose that clearance was scoped to, the lineage from raw ingestion through transformation, the confidence attached to the recommendation, and the alternatives that were considered and dropped. Governed Autonomy, Evidence Ledger, Signal Firewall, and Multi-Agent Orchestration are not four features here — they are one behaviour.
§ 01
State & evidence
01
State reconciliation and drift detection between what was predicted, what the sensors observe, what the simulation produced, and what an operator confirmed. When those four disagree, the system says so rather than picking one silently.
Manifold · PRODIGY · SerpenOps
02
A hard separation between verified information, lower-confidence signal, and anything a model generated. Risk and confidence are scored on separate axes, so a severe but unconfirmed signal reads as investigate, never as act.
Cortex · Global Health IQ · Intelligence Hub
03
End-to-end provenance for information, transformations, decisions, and the actions that followed. Six months later, someone will ask why a call was made. The ledger is the answer to that question, written at the time.
All four platforms
§ 02
Reasoning & action
04
Oversight, scoped permission, escalation paths, and recorded reasoning around anything a system does on its own. Access is granted for a stated purpose against a named data boundary — not to a role in the abstract.
Aya · Work OS · PRODIGY
05
Routing and coordination across specialised reasoning and execution agents — planner, executor, memory retriever, verifier, router, self-critic — with the discarded paths kept alongside the chosen one.
Aya · OpenAya · Work OS
06
Turns an ambiguous institutional problem into structured objectives, constraints, evidence requirements, and work that can actually be executed. Most failed deployments failed here, before anyone wrote code.
Aya · Cortex · Mission Sprints
§ 03
Physical & data substrate
07
Assets, relationships, movement, and geography understood together — where a thing is, what it touches, and what changes when it moves.
Manifold
08
Bring-your-own-storage with controlled persistence, so applications reach institutional data without ever holding the provider's tokens.
Vault
09
Operation through low-connectivity, degraded, and intermittent conditions — which is when institutions need the system most.
Manifold · SerpenOps
10
Sensors, feeds, APIs, telemetry, databases, documents, and the institutional systems already in place — combined without replacing any of them.
All four platforms