Architecture summary

                    +-----------------------+
                    |  Customer (any tier)  |
                    +----------+------------+
                               |
                               v
                    +-----------------------+        +----------------------+
                    |  Gateway / Router     |<------>|  Billing bridge      |
                    |  (LiteLLM front)      |        |  (Stripe / x402)     |
                    +----------+------------+        +----------------------+
                               |
            +------------------+------------------+
            |                                      |
            v                                      v
   +-----------------+                +-----------------+
   |  Operator(s)    |                |  Validator(s)   |
   |  vLLM/SGLang    |                |  replay sampler |
   |  TEE attested   |                |  Yuma scoring   |
   +--------+--------+                +--------+--------+
            |                                  |
            v                                  v
                +---------------------------------+
                |  Chain (Substrate)              |
                |  pallet-{model-registry,        |
                |   operator-stake, job-market,   |
                |   yuma-consensus, bme,          |
                |   slashing, pouw-mint,          |
                |   attestation-registry,         |
                |   oracle-twap, nonce-vault,     |
                |   treasury-ext}                 |
                +---------------------------------+
                                |
                                v
                +---------------------------------+
                |  Explorer · Subsidy dashboard   |
                |  Indexer (Subsquid)             |
                +---------------------------------+

Verification stack (8 layers)

1. GPU device certificate + stake binding — hardware-rooted identity. Same device cert across coldkeys → both slashed 100%.
2. TEE attestation — Intel TDX + AMD SEV-SNP + NVIDIA H100/H200/B200 CC. Multi-vendor mandatory at launch.
3. Deterministic kernels — SGLang det-mode + batch-invariant kernels. Enables cheap replay.
4. Stake-weighted validator sampling — Yuma. ≥10% sample rate floor (red-team rule 2). Per-detection slashing.
5. opML challenge window — high-value jobs only (Phase 4).
6. zkML — small heads (moderation, routing, classification).
7. cuPOW kernel — minting only (5% supply, optional, deferred to Q4 2028).
8. Watermarks — output provenance.

Off-chain services

• Worker daemon (infer-worker-vllm) — primary; vLLM V1 plugin shim.
• Gateway / Router — OpenAI-compatible; routes by capability + tier + latency + price.
• Validator network — replay sampling, Yuma scoring, slashing evidence submission.
• Attestation service — multi-vendor PKI aggregator; produces combined signed reports.
• Weight CDN — content-addressed; triple-source (S3 + R2 + IPFS).

Tokenomics summary

• OROG — native token. Required to burn for any paid inference.
• CUC — non-transferable, account-soulbound compute credit. Minted on burn.
• BME — burn-mint equilibrium. Year 1 cap 8% supply, Year 2 4%, Year 3+ demand-elastic (180-day rolling cap).
• Split — 75% operator / 15% verification / 5% treasury / 5% governance.
• Subsidy target — Y1 <2×, Y2 <1.2×, Y3 <1×.