Running a provider in the hosting research profile

This runtime serves assigned partitions of the complete answering graph through authenticated HTTPS peers. It is available in the matching research source checkout. The public 0.4.0 genesis does not enable these transactions. Use a freshly pinned hosting research genesis and its exact source/runtime release. Accepted-LLM owner replacement and the frozen latency targets pass the operated trial. The trial endpoints are retired; an independent-operator soak and persistent public deployment remain open.

Each provider controls its own native wallet, TLS key and local full node. The node must have finished synchronization. No SSH access to another provider, shared wallet, operator allowlist or remote code installation participates in serving. RPC authorization is restricted to the provider's own loopback node; ABCI responses are not portable light-client proofs.

Use the exact source commit and numerical profile published with the genesis. For the accepted-model trial the GPU profile requires Python 3.10.12, an NVIDIA A10G and the pinned CUDA/PyTorch environment. Other GPUs are not yet admitted by that numerical commitment. Permissionless registration does not promise arbitrary hardware compatibility. The metadata-only native node uses its separate CPU environment; it does not download the full model.

From the pinned checkout, create both environments without installing over an existing node:

python3 -m venv .neuroshard/native
.neuroshard/native/bin/python -m pip install -r docs/evolution-requirements.txt
python3.10 -m venv .neuroshard/provider
.neuroshard/provider/bin/python -m pip install -r docs/expert-execution-requirements.txt

The release must supply its genesis file/hash, reachable native peer, executor profile and object-length inventory. Build the pinned CometBFT binary using the candidate instructions. For a newly agreed genesis, start a non-voting full node with its own keys:

PYTHONPATH=src ATEN_CPU_CAPABILITY=default MKL_ENABLE_INSTRUCTIONS=SSE4_2 \
OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
  .neuroshard/native/bin/python scripts/join_funded_candidate.py \
  --home /srv/neuroshard-full-node --genesis /srv/release/genesis.json \
  --genesis-sha256 RELEASE_GENESIS_SHA256 --engine /srv/release/cometbft \
  --peer NODE_ID@PUBLIC_PEER:26656 --base-port 26656

This helper now accepts the funded expert-graph profile as well as the earlier lifecycle profile. It pins the source and genesis and preserves an existing recognized home. It is for the fresh research deployment: following an old stake history additionally needs a recent independently trusted checkpoint. Becoming a validator requires native stake admission; running this follower grants no vote. Keep RPC on loopback and expose the native peer and advertised HTTPS provider port. Never share account, TLS or consensus private keys to join.

Create a local JSON configuration with the release's chain ID, manifest hash, executor profile and object-length inventory. Mirror bases must be configured locally. Each requested digest is already committed by the graph or policy; the inventory cannot substitute different model bytes.

{
  "home": "/srv/neuroshard-provider",
  "node_rpc": "http://127.0.0.1:26657",
  "chain_id": "RELEASE_CHAIN_ID",
  "manifest_root": "RELEASE_MANIFEST_SHA256",
  "advertise": "https://YOUR_PUBLIC_HOST:8443",
  "bind": "0.0.0.0",
  "port": 8443,
  "rank": 0,
  "profile": "/srv/release/executor.json",
  "inventory": "/srv/release/object-lengths.json",
  "source_home": "/srv/neuroshard",
  "mirrors": ["https://YOUR_CONFIGURED_MIRROR/objects"],
  "max_bytes": 17179869184,
  "prepare_seconds": 600,
  "frame_seconds": 60,
  "run_seconds": 7200,
  "collateral": 50000000,
  "fee": 1000,
  "offer_blocks": 10000
}

The numbers are an illustrative local configuration, not a tested public price or latency promise. rank must exist in the accepted graph. collateral and fee use native atoms; the coordinator also reserves the genesis claim bond. Keep enough liquid balance for transaction fees. There is one active execution slot per runtime, advertised with capacity one. Clients obtain concurrency by reserving other available replicas. Do not reuse one wallet across concurrent runtime processes or independent transaction outboxes.

export PYTHONPATH="$PWD/src"
export ATEN_CPU_CAPABILITY=default MKL_ENABLE_INSTRUCTIONS=SSE4_2
export OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 MKL_NUM_THREADS=2
export CUBLAS_WORKSPACE_CONFIG=:4096:8
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
.neuroshard/provider/bin/python -m neuroshard.evolution.provider_runtime --config provider.json --identity
.neuroshard/provider/bin/python -m neuroshard.evolution.provider_runtime --config provider.json --publish-offer
.neuroshard/provider/bin/python -m neuroshard.evolution.provider_runtime --config provider.json

The first command creates local keys and prints only public registration fields. Fund that native address on the intended research chain before publishing the offer. Publishing reserves the configured collateral and creates the bounded service offer. The final command discovers assignments for this key and rank, restores/checks its own partition, acknowledges the native assignment and runs the pinned model through authenticated peers. Its coordinator submits the complete, jointly signed response. Payment still requires funded full replay.

Keep the provider identity, TLS key and transaction journal on durable storage with a private backup. The local native node has separate keys and stores to preserve. Only home/models is a disposable cache suitable for instance-local SSD storage: committed model bytes can be fetched again, but a new provider key cannot spend the old key's balance or withdraw its collateral. Restore the same identity and journal together, and require a new native assignment after an interrupted execution.

Reissuing an identical publish command recovers the same signed operation; it does not renew an expired offer. A fresh advertisement needs a changed offer intent. Cancel an offer through cancel_expert_offer before a planned departure; existing reservations remain obligations until completion, expiry or replacement. Withdraw only unreserved collateral after the native cooldown.

The continuous runtime keeps one immutable model partition resident between requests for the same graph. A new assignment installs fresh authenticated wires and frame sequences while preserving those weights. Request KV caches and conversation context do not carry between jobs. Changing the graph evicts the older resident partition before loading its replacement.

/hosting/quote supplies a complete first-attempt debit bound: neural execution, the selected providers' whole-request fees, full verification funding and a bounded client transaction-fee allowance. Its matching checks shared collateral, offer capacity and expiry. Quotes expire and native reservation rechecks them; a quote alone does not reserve capacity. A rejected audited attempt needs fresh verification funding rather than an automatic additional debit.

On interrupted execution, providers wait for replace_hosted_job to commit a new assignment. The client or any participant can replace a timed-out group within its original fee ceiling and attempt/expiry limits. Late frames and receipts cannot settle the new epoch. A pending transaction is recovered using its original signed bytes; a committed epoch closure can retire that unknown operation without inventing a successful receipt. Refunds follow native state, not the process's local log.

Model files remain in the configured home for reuse; local cache management and the host filesystem quota remain the operator's responsibility. Transcripts and failure records are local and persist across restarts. Current assignments, prompts, neural-call token IDs and final responses are public ledger data. Providers, auxiliary model owners and auditors see the text needed for execution. TLS protects connections; this profile does not offer private conversations or ledger erasure.

The hosted chat client obtains a complete quote, funds native verification and reserves provider capacity. The coordinator exposes a bounded certificate-pinned /v1/events endpoint authenticated by the customer's wallet. Visible drafts are provisional; only native replay settlement is authoritative.

Validation: adversarial native accounting, authenticated transport/retries, partition-only HTTP restoration and five-process real neural equivalence cover the implementation. The operated accepted graph additionally recovered a killed coordinator and a killed backbone owner on seven GPU hosts, each with a fresh key, endpoint and empty model cache. Cached surviving owners use the same initialization protocol as fresh owners. All recovered replies passed three complete numerical replays and settled once. The final correction also passes all six frozen ordinary latency cases without changing any final neural result. The preceding failed combined-response gate remains published. See the complete measurements and costs before treating this research profile as a public service.