Operated alpha: retired GPU service

GPU serving and new admissions closed on September 20, 2026 at the owner's request.The deployment result preserves the complete service gate, failures, reproducible evidence and retirement record. The original September 22 GPU deadline was a maximum lifetime, not a promise of continued service. The ledger stays scheduled through September 26, 2026 at 23:49 UTC for settlement and refunds. The retired window served the already accepted growing graph and ran no new training jobs. Observer instructions remain useful while the ledger is online; the starter-credit and chat commands below are historical and do not provide a live GPU service. No replacement GPU deployment has been announced.

The trial supported two concurrent customers, with up to 64 output tokens. The six ordinary deployment cases returned first text in 33.15–42.17 seconds; full verification and payment settled later. All bootstrap hosts belong to one administrator. Trial credits have no demonstrated market value. The website and PyPI 0.4.0 package use a different chain; follow this source installation.

Run your own ledger observer

Use Linux x86_64 with Python 3.10–3.12, at least 4 GB RAM and several GB of free disk. This observer needs the pinned CPU dependencies, but no GPU or model weights. Review the descriptor and its exact source commit before installation. Download the reviewed source and public descriptor:

git clone https://github.com/neuroshard-ai/neuroshard.git
cd neuroshard
git checkout 52ec72f49b5e77b682bc5e48174fdb9f9544ca25
curl -fSL 'https://dwquwt9gkkeil.cloudfront.net/research/native-expert-live-20260916/objects/7a0a49c8d9f8731145fc88078cd3c2fcbd526d4483d63c29e3b1b8e490c40bfd' -o access.json
printf '%s\n' '7a0a49c8d9f8731145fc88078cd3c2fcbd526d4483d63c29e3b1b8e490c40bfd  access.json' | sha256sum --check
python3 -m venv .venv-alpha
.venv-alpha/bin/python -m pip install -r docs/evolution-requirements.txt
.venv-alpha/bin/python -m pip install .
source .venv-alpha/bin/activate
python scripts/join_operated_alpha.py --descriptor ./access.json --home ./alpha-node

The observer verifies native blocks from genesis. Keep it running and wait until it reports ready; its RPC listens only on your machine. The descriptor pins the genesis, manifest, executable checksum, complete answering graph and native peer endpoints. Merely starting an observer neither spends nor bonds tokens.

Historical starter-credit and chat commands

The starter-credit endpoint and GPU service are closed. These commands document the completed trial; retain an existing wallet when inspecting prior payments.

In another terminal using the same environment:

neuroshard wallet create --home ./customer
python scripts/operated_alpha_credits.py --descriptor ./access.json --wallet-home ./customer
neuroshard chat "What is the name of the NeuroShard client package?" \
  --home ./customer --hosted-config ./alpha-node/hosting.json \
  --max-tokens 64 --quote-only

The credit service receives only your public key. It sends at most 500 trial NEURO per key, with a global 10,000 NEURO sponsorship ceiling. This is not proof of unique people: multiple keys can exhaust the finite pool. A pending request must be retried with the same wallet; it never authorizes another grant.

Prompts, conversation context and settled replies are public ledger data. Read the complete quote, including verification and occupied capacity, before paying:

neuroshard chat "What is the name of the NeuroShard client package?" \
  --home ./customer --hosted-config ./alpha-node/hosting.json \
  --max-tokens 64 --max-price 500 --session ./conversation.json --wait-seconds 600

Provisional text appears before native settlement and remains unverified until then. If the command times out, use its printed request ID with --resume and the same wallet. Never sign another payment to replace an uncertain one. Reuse --session for another turn only after the preceding request settles. Capacity is limited to two simultaneous requests and the graph has a bounded context. If both slots are occupied, a quote can be refused; retry later. A quote refusal does not submit a payment. Full replay is expensive: the maximum verification budget is locked before execution, then unused funds return after settlement. A 500-credit grant can start a bounded request; how many later turns fit depends on actual charges and the next complete quote. It is not an unlimited chat subscription.

See chat and recovery behavior and the funding and trust assumptions. Running your own observer establishes your own ledger verification. Independent provider and validator operation also requires control of those keys, resources and administration; the fixed TODO does not count centrally controlled AWS hosts as outside operators.