Chat with the accepted shard graph

The provider research profile connects the existing lightweight client to the complete answering graph. It requires that profile's freshly pinned native genesis, available provider offers and running full-replay auditors. It is not enabled by the public 0.4.0 genesis. The accepted-LLM trial passed its frozen two-customer latency and owner-recovery gates. Those experimental endpoints are retired. The subsequent operated alpha is available during its published funded window, with complete deployment evidence.

These commands require the matching research source release. The existing neuroshard-ai 0.4.0 package on PyPI does not include --hosted-config. From the reviewed checkout, install the lightweight client in a separate environment:

python3 -m venv .venv-client
.venv-client/bin/python -m pip install .
source .venv-client/bin/activate

Create a local wallet with neuroshard wallet create --home ./customer. Run a synchronized local full node for the reviewed hosting genesis. The client uses that node as its authority; these ABCI queries do not have light-client proofs. Put the node's loopback RPC, chain ID and manifest digest in hosting.json:

{
  "node_rpc": "http://127.0.0.1:26657",
  "chain_id": "CHAIN_ID_FROM_THE_REVIEWED_GENESIS",
  "manifest_root": "SHA256_OF_THE_CANONICAL_GENESIS_MANIFEST"
}

Read the complete price before submitting:

neuroshard chat "What can this network help me with?" \
  --home ./customer --hosted-config ./hosting.json --max-tokens 64 --quote-only

The quote includes bounded neural execution, all selected providers' fixed fees, full verification funding and a transaction-fee allowance. It expires after 64 blocks. Discovery does not reserve capacity. No model or GPU dependencies are loaded by the client. The local full node is a separate service.

The operated alpha profile additionally quotes the maximum capacity occupancy charge. Its admit_work transaction reserves the request, providers and a complete native audit together. Unused occupancy is refunded; occupied blocks and transaction fees remain spent. This profile requires its own genesis; it does not change an existing hosting chain.

The signed request asks native settlement to select available offers atomically inside that price ceiling. Two customers may read the same indicative quote; their committed reservations acquire different available capacity. A quote is not a guarantee of a particular provider key.

On the earlier hosting profile, fund_hosted_audit binds the verification reservation to this customer, graph and complete conversation hash. Another customer cannot front-run that budget, and its selected coordinator cannot attach it to unrelated neural work. The generic sponsor-funded fund_audit transaction is not accepted for hosted chat. The operated alpha disables this separate funding step: it binds the same customer, graph and request inside atomic admission against standing audit services.

After funding the wallet, replace YOUR_LIMIT with the maximum total NEURO you authorize. The client refuses a quote exceeding that limit before payment:

neuroshard chat "What can this network help me with?" \
  --home ./customer --hosted-config ./hosting.json --max-tokens 64 \
  --max-price YOUR_LIMIT --session ./conversation.json --wait-seconds 600

The terminal displays provisional text during generation. --json emits signed provider update contents and native result records as JSON lines. Each update replaces the previous draft; incremental token decoding is not always stable at the character boundary. Planning and intermediate worked text are excluded from the chat display. For a deterministic combined reply, each completed answer part becomes visible while the next part runs. Neural composition and structured rendering still require their complete inputs before visible output. Empty intermediate answers do not produce question-only previews. A malformed later answer retracts the provisional reply.

Drafts are unverified. Agreement among assigned provider signatures alone cannot authorize payment. The final reply and refunds come from native settlement after complete paid replay. A lost stream or customer disconnect does not cancel the native job, establish a failed audit or trigger another payment. Full replay and native audit windows also mean settlement can arrive much later than visible text. No public latency guarantee has yet been established.

Every signed payment is journaled before broadcast. If the wait ends or the client is interrupted, use its printed request ID to recover the same signed operations:

neuroshard chat --resume REQUEST_ID --home ./customer \
  --hosted-config ./hosting.json --wait-seconds 600

Do not replace an uncertain request with a newly signed payment. On the earlier profile, an audit offer that has not acquired a job is cancelled when its quote expires. Unused job, provider and audit budgets return through native expiry; transaction fees remain spent. A rejected audited claim does not authorize automatic new verification funding. Provider replacement can restart the pinned request within its native attempt, price and expiry bounds. The client follows a committed replacement and clears the obsolete draft. In the operated alpha, a recovery controller can request native replacement after the assignment deadline; the customer's client does not need to select or administer the replacement machines.

Only settled, nonempty replies extend the conversation file. Reusing --session adds the next user turn. The client pins the graph and tokenizer; a changed version requires a new conversation file. It never silently truncates context or treats an unverified draft as assistant history. Recovering an old request cannot overwrite later recorded turns. Run one process and outbox per wallet; different customer wallets may reserve different available provider capacity.

Visibility and storage: prompts, prior conversation, neural-call token IDs and final replies are public ledger data. This includes intermediate model calls even though the chat display hides them. Assigned providers, auxiliary model owners and auditors receive the conversation needed for execution. TLS protects transport, not those recipients. Do not submit private information. The ledger does not offer erasure. Local request journals, settled conversation files and provider transcripts persist until their owners remove them. Draft delivery keeps only the latest text per assignment, in bounded provider memory; native blocks and final execution records have separate retention obligations.

Checks cover token-by-token observation against full-model generation, delivery failure without numerical changes, wrong-customer and obsolete-epoch rejection, complete spending caps, quote cancellation, and lost-acknowledgement recovery after native expiry with all unused budgets refunded. The separate accepted-model trial settled ten requests after thirty complete numerical replays and met its frozen latency/load gates. This small sample does not establish a public latency guarantee.