DCS·agiA research program with a built substrate.
We do not claim general intelligence. We claim a verifiable, sovereign substrate — designed for frontier capability and safety — that is built and verified today, running live-dark while it is hardened. Below is exactly what exists, and what is explicitly future and decision-gated.
The substrate
P0 → P6, built & verified — live-dark.
Each layer is implemented and tested. “Live-dark” means built and verified but intentionally not enabled in production until the gate is passed. P7–P8 are future work, gated on the compute raise and safety review.
Foundation spine
The base inference + training spine — the model and serving core the rest of the program builds on.
Vertical curation
Curated, attested data pipelines that specialise the spine into real domains rather than scraped generality.
Trust & cognition
The trust layer — provenance-aware reasoning and the cognition scaffolding that tracks why an output was produced.
Safety gates
Refusal and safety gating evaluated at 90/90; safety is a gate in the path, not a wrapper around it.
Regulated / hardened API
A hardened, regulation-aware API surface for exposing capability under policy.
Sovereign deployment
Run the full stack in your own environment, air-gapped if required.
Federation
Federate across sovereign nodes — capability without centralising data or control.
(future) Scaled training
Decision-gated. Larger training runs contingent on the compute raise and safety review.
(future) Autonomy research
Decision-gated. Bounded autonomy research behind explicit safety and governance gates.
Verticals
Nine verticals on one substrate.
The curation layer specialises the spine into nine real domains, several already shipping as products.
Safety & resilience
Gated, audited, simulated.
Refusal / safety 90/90
The safety gate is evaluated at 90/90 against a refusal/safety battery — measured, not asserted.
ISO-42001 pack
An AI-management-system pack mapped to ISO/IEC 42001, documenting governance for the program.
BFT / partition sims
Byzantine-fault-tolerance and network-partition simulations validate the federation layer under adversarial conditions.
Verification underpins all of it — see Verifiable AI for the cryptographic provenance layer.