Full-stack architecture & MCML
Deterministic runtime enforcement. A reserved inside view.
The planned curriculum connects frontier model research with executable enterprise controls. MCML—Model Connect Markup Language—treats governance as an execution boundary rather than a post-hoc spreadsheet audit.
Safety Inside the Model. Governance Around It.
Track 1 learners build and test runtime controls. Track 2 leaders write and audit the policies. Together they form one defense-in-depth architecture.
Defense in depth
Govern what an agent can do—not just what it says.
01 · Govern the context
Signed Agentic Governance Cards define permitted actions. Review, provenance checks and safety evidence precede deployment.
02 · Lock down execution
An out-of-band Zero-Trust MCP Gateway validates tool parameters, authority, spending and rate boundaries. Prompt injection can change requests, not permissions.
03 · Inspect validated signals
A reserved interface can carry activation probes, model diffing and steering signals as the research matures. These signals never replace deterministic authority controls.
Signed policy + verified model provenance
External MCP permission boundary
Validated execution → version-bound audit receipt
Learner stretch targets · not measured results
Learner stretch target: assurance below 5 milliseconds.
>60%
Policy-prefix caching
Learner stretch target: reduction in governance-related TTFT overhead using pinned KV-cache blocks and prefix trees.
<5 ms
Speculative safety verification
Learner stretch target: runtime overhead through parallel deterministic checks during normal generation intervals.
<1.5 ms
In-SRAM probing
Learner stretch target: tripwire latency for lightweight linear probes using custom Triton/CUDA kernels.
These are stretch targets for learner experiments, not measured PLEYVARSITY results or guaranteed assessment outcomes. They require controlled benchmarking and independent validation. No deployed performance guarantee is made.
Infrastructure roadmap · planned
Year 1 AI Control Lab. Year 2+ digital twins.
PLEYVERSE AI Control Lab
Year 1 plans PLEYVERSE-hosted GPU sandboxes behind MCML gateways, with transport kill-switches, version-bound receipts and model provenance verification. This is the proposed reference deployment, not a claim of a live service today.
Digital twin scale-up
From Year 2 onward, planned high-fidelity enterprise and lab replicas aim to reduce per-learner compute cost and widen access. Partner agreements, benchmarking and funding are required.