Operated alpha deployment result

The accepted growing graph is available through a source-installed CLI alpha. Join the alpha. This is a finite service operated by one administrator on seven GPU hosts and four separate ledger hosts. AWS participants have separate keys and machines; they do not establish independent ownership. No physical host contains the complete backbone.

The original ten-request gate passed: two warm-ups, six ordinary requests with two concurrent customers, and two automatic provider-process recovery cases. All eight measured ordinary/recovery responses and neural outputs exactly match the preceding accepted-model service trial. This deployment runs no new training and makes no new model-quality claim.

Measured behavior

Request	First visible text	Generation complete	Native settlement
client-package	34.02 s	37.04 s	99.68 s
retained-directory	36.51 s	40.14 s	161.51 s
ordinary-general	34.30 s	49.34 s	168.09 s
constrained-general	33.15 s	34.75 s	99.62 s
multiturn-context	36.21 s	38.28 s	111.47 s
combined-question	42.17 s	46.99 s	180.27 s

First-visible and generation timing includes atomic native admission. Settlement includes the batch's end-to-end time. The frozen limits were 45 seconds for first visible text, 90 seconds for generation and 1,200 seconds for ordinary settlement. Each of the six ordinary cases passes; this small sample is not a population latency guarantee. Full verification is substantially slower than draft delivery.

Automatic process-loss recovery settled once within the unchanged 1,500-second limit: coordinator-loss: 446.36 seconds and backbone-loss: 504.27 seconds. The driver stopped a provider service, preventing systemd from immediately restarting it. The recovery controller and native assignment selected a previously advertised replica on another host. The request, price ceiling and graph stayed pinned. These are process-loss checks, not a physical-host or independent-operator availability soak.

A fresh observer synchronized from the public genesis and peer endpoints. A new wallet obtained 500 sponsored trial NEURO; retrying the grant returned the same transfer. Its first chat deliberately timed out, then resumed the same request without signing another payment. The driver first requested an unsupported 1,200-second CLI wait; that rejected invocation is preserved, followed by a 600-second resume of the same admission. The second turn retained the requested phrase: violet kettle. Both paid turns settled, provisional text was observed, and the graph/tokenizer-pinned conversation contains four messages. The first reply invented an irrelevant definition of “violet kettle.” That poor answer remains in the published transcript. This check establishes context retention and payment behavior, not broad answer quality.

Evidence and reproduction

The snapshot reproduces 3,369 headers and 697 accepted transactions, with zero rejected transactions, against all four saved validator states. There are 36 complete numerical replays across 12 claims, covering the deployment gate and onboarding. Serving issued zero new tokens. The final snapshot state is e8c9ceaac60be926c10a713bc1075e1532daaa6ec9f54207604dcb318dc77bf7.

• Current ledger, numerical reports and replies — SHA-256 6579c3fef23c36f109711e3f8e757bf82509d5fda5a62ec2b15a6ecced33f679.
• Exact source — SHA-256 7484da5237e1295cb6ccfb470136cb4ee3d349ee73247cb62ab5e035605c2362; commit 52ec72f49b5e77b682bc5e48174fdb9f9544ca25.
• Pinned public access descriptor — SHA-256 7a0a49c8d9f8731145fc88078cd3c2fcbd526d4483d63c29e3b1b8e490c40bfd.
• Machine-readable result.

After downloading and checking the evidence checksum, extract it into an empty folder. From the pinned source with the documented CPU dependencies, replay it:

ATEN_CPU_CAPABILITY=default MKL_ENABLE_INSTRUCTIONS=SSE4_2 \
  PYTHONPATH=src:scripts python scripts/replay_provider_preflight.py \
  --home /absolute/path/to/extracted-evidence

Application replay verifies transitions against saved states, header linkage and application hashes; it does not independently verify CometBFT commit signatures. The separate joining observer follows native consensus from genesis. Numerical reports represent three actual full replays under one administrator's audit keys; they are not proof of independent auditing.

Availability, funding and limits

Admission is scheduled to close September 22, 2026 at 11:09 UTC. GPU service ends by September 22, 2026 at 12:39 UTC, or earlier at the funding/request limit. The ledger remains scheduled through September 26, 2026 at 23:49 UTC for expiry/refunds, then private validator signing state and block stores are preserved before retirement. Hourly private provider-wallet backups and absolute resource retirement timers are enabled. These records are excluded from the public archive.

The aggregate ceiling is $800. The complete-window planning estimate at publication is $731.82, including all 60 GPU hours, seven ledger days, a worst-case CPU bursting allowance, disks, networking, controller use and failed-deployment reserves. This is a conservative forecast, not an invoice. New capacity closes at the $750 watch threshold, leaving $50 for draining admitted work. Starter credits are finite sponsor transfers with no demonstrated market price or Sybil-proof distribution.

The service has two simultaneous request slots and a 64-token response cap. Prompts, prior context and settled responses are public ledger data. Draft text is provisional until settlement. Full replay remains expensive. The website and PyPI 0.4.0 client still use the older chain; use the pinned source joining guide.

Earlier failures and their declared repairs remain recorded: preparation/heartbeat acknowledgement, the GPU library path, exhausted CPU credits, a strict 45.016-second latency miss and a check against a lagging validator. No failed result was rounded into a pass. The accepted weights, tokenizer and neural behavior were preserved throughout these deployment repairs.

This completes item 6's bounded operated-alpha criterion. Item 4 remains open for independently administered hosting, its declared availability soak and the remaining permissionless admission/resilience evidence. Prior learning criteria remain bounded demonstrations, not proof of ChatGPT-level general capability.