Native provider join and recovery preflight

The frozen CPU trial passed all three cases: initial service, loss of the coordinator, and loss of a required backbone owner. Fresh provider keys joined through native registration, restored their own committed partitions over HTTP, and executed through authenticated HTTPS peers. Recovery preserved the original request and graph, changed the assignment epoch, and paid each request once.

Case	Assignment epoch at settlement	Complete numerical replays	Case duration
Ordinary request	0	3	104.81 s
Coordinator killed and replaced	1	3	247.55 s
Backbone owner killed and replaced	1	3	245.58 s

The failed owners were killed after native acceptance and creation of their execution-start marker. Each replacement used a fresh key and endpoint with an empty model cache. The protocol restarted the pinned request; it did not recover an in-memory activation stream. Case durations include fresh process startup, funding, three full numerical replays and native settlement windows. They are not measurements of LLM generation latency.

Every case paid 7 execution atoms and refunded 987 unused execution atoms. Its separate provider reservation paid 510 atoms and refunded 490. The three requests issued zero tokens. These fixture prices are not production tariffs. All four validators agreed on the completed result. Fresh application replay reproduced 660 headers, 84 accepted transactions and all four stored validator states, including collateral, audit payments and refunds.

This is a 33,760-parameter synthetic graph on one physical host under one administrator. Preparation uses the existing deterministic fixture's two tail revisions. Its ordinary request emitted a terminal EOS token; the experiment checks execution and settlement, not useful answering or accepted-LLM quality. There were no new instances or GPUs. The 628.52-second trial used a 7 GiB memory limit, 1.5-CPU quota and 20-minute cutoff. Every trial process stopped; the host's two existing public services remained running.

The method was committed before execution as 71a6d7fde8fdc5c67a43424cdd7fed251e81e87e. The source commit that introduced the provider implementation, 7dc8dd6, passed all five CI checks, including 989 tests on each of Python 3.10 and 3.12. Later resident-model reuse and customer-quote changes have their own focused checks and were not part of this frozen native run.

Public artifacts, each verified by complete public readback:

• Evidence, model fixture, native blocks, four saved states and replay driver, SHA-256 8c30280e59ecabbf53bde07d8bd932a2bbabd8a95c0c2b5956783c48d3faf7a5.
• Matching source archive, SHA-256 abf642c08f95b2d70cb150132f31e866467c9ffe1718102c9db6ae7ea3c9ba62.

Verify both archive hashes and the evidence file inventory before use. With the declared research CPU dependencies installed, application replay requires no signing key, GPU or running native node:

PYTHONPATH=/absolute/path/to/matching-source/src \
  python /absolute/path/to/evidence/replay_provider_preflight.py \
  --home /absolute/path/to/evidence

The replay checks application transitions and saved states; it does not independently authenticate CometBFT commit signatures or redo neural execution. The separate complete numerical reports cover the three inference claims.

TODO 4–6 remain open. The accepted LLM still needs the frozen operated recovery/load trial and complete cost accounting; chat delivery needs its public integration. Independent administration and voting-power distribution still require operators outside this administrator's infrastructure.