Funded audit integration evidence — September 11, 2026
This candidate pays a separate process to replay complete neural execution and tests recovery on a native ledger. It uses a new genesis and does not change the public 0.4.0 network. All accounts and machines in these experiments belong to one operator. The audit rules describe its assumptions; the operations guide provides reproduction commands.
The evidence bundle contains genesis files, final results, replay reports, recovery records and four-validator supply checks, with a SHA-256 manifest. These are inspectable operator measurements, not proof of independent ownership.
Source and topology
The bounded trial uses the scripts at 3710226. Both its synthetic and real-model genesis commit package source 2875b6e2de443c136cfffe3288b35d03a3f763ee521ce4f231d6704043e1a125. The source hash covers package Python files; it is distinct from a Git revision or genesis hash. Later off-chain operator checks bound inference length and require explicit subsidy approval. They were tested separately and were not part of the frozen trial's second-cohort training path. The current driver also extends its configurable second-cohort wait to two hours after the frozen real-model trial exceeded its one-hour limit; it does not change ledger deadlines or package source.
Four validators run on the primary Intel Xeon Platinum 8259CL host. The real trial places two workers there and two workers, the funded auditor and a non-voting full node on an Intel Xeon Platinum 8175M host. Each host has four logical CPUs and approximately 16 GiB RAM. Remote workers and the auditor each have a one-CPU quota. Computation uses the pinned float32 CPU profile, one numerical thread, default ATen capability and SSE4.2 MKL instructions. Existing public services also use these machines; timings are not isolated benchmarks.
Workers and the artifact mirror use HTTP through SSH. The second-host follower first joins through a native peer tunnel, then restarts with the same keys and genesis through the public peer at 54.210.225.74:26666. The peer tunnel is removed. All five full nodes agree on the application hash in block 3,588: AA3A1C5372395FB98A70C3857D8C41B58A57C8E0802499A426BB662737DF3FE3. This verifies public peer connectivity and replay across two CPUs. It does not establish five independent operators or survival of the validator host.
The real chain is neuroshard-lifecycle-check-2f4ad2ad7d32, with genesis SHA-256 d7695cbb6a8eccdaa76854c3e40ddbbe11894e2442a5fc7cce13a7cddf0ac654. The synthetic chain is neuroshard-lifecycle-check-b84227cd51c0, with genesis SHA-256 4e27a2feb9258b9bb3d2f09c10ddba4e94ee97fcb66fa0b86af2ac6eb5f30029. Each starts with 4,010 experimental NEURO, including validator collateral. These allocations and subsequent rewards are unrelated to public balances.
Completed synthetic rehearsal
The synthetic model has 10,384 parameters. Two cohorts complete eight training steps and both quality evaluations. Killing the continuous operator after its durable reservation and restarting it preserves that reservation and completes the second cohort. Exactly 8,000,000 atoms are issued for eight distinct paid tasks. The auditor receives 21,100,000 atoms from existing sponsor balances for 73 accepted services. Auditing creates no issuance.
The auditor records 214 stage replays across 74 reports: 211 covered stages in paid services and three stages reaching the forged output head. Native replay rejects that head using 4,520 uploaded bytes. Honest one-token inference pays 1,000 atoms and refunds 6,000. Both synthetic promotion decisions pass; this is a protocol fixture, not evidence of LLM improvement.
Three of four validators continue producing blocks. Half the voting power halts finality. Restarting the retained validators restores agreement in block 1,393, with application hash 90EEEC60E21758CE0812CCB114FC5FFAF88EEB2995650A8136E0842A74B9D580. Final supply accounting is 3,997,602,000 liquid atoms, 10,000,000 validator collateral, zero remaining escrow and 10,398,000 burned atoms. Their sum equals the 4,010,000,000 initial atoms plus 8,000,000 issued atoms.
Real model and data
The initial SmolLM2 model has 134,515,008 parameters. Four identity layers expand the learning candidate to 148,675,392 parameters under four declared 48M-parameter partition capacities. Growth earns no training issuance. The first four full-model updates issue 4,000,000 atoms.
The tokenizer and two reviewed Smol-SmolTalk cohorts are the same immutable inputs documented in the earlier lifecycle results. Each evaluation group contains 32 documents. The first cohort has 69 retention and 75 fresh-response windows; the second has 89 and 76. Windows are aggregated within each document before paired scoring. Previously unused rows are not evidence of newly acquired world knowledge. Four training steps do not consume every admitted training window, and these public cohorts can be overfit.
The first evaluation rejects promotion. Retention passes its tolerance; the fresh-response result does not establish the required 0.001-nat improvement. The original serving root remains active and becomes the next learning parent. Verified computation is paid even when its candidate fails the quality gate.
The second cohort completes four more updates and its full evaluation on a 134,515,008-parameter candidate. Retention again passes and fresh-response improvement again fails. The two decisions are:
| Cohort | Candidate parameters | Mean retention loss change (nats) | Mean fresh loss change (nats) | Promotion |
|---|---|---|---|---|
| First | 148,675,392 | −0.00039184 | −0.00120472 | Rejected |
| Second | 134,515,008 | +0.00030163 | +0.00002372 | Rejected |
These means use the recorded quantized document differences; negative values mean lower loss. The first fresh point estimate improves, but its uncertainty bound does not establish the required improvement. Each decision uses the predefined 32-document paired gate, including its quantization allowance. Neither result establishes a generally more useful language model.
Across both cohorts, eight distinct training tasks issue exactly 8,000,000 atoms. The remote auditor records 169 reports and 706 stage checks: 168 accepted services cover 702 paid stages, and the forged response reaches four checks. The sponsor pays 70,200,000 existing atoms for those services. No audit payment creates issuance. The publisher separately replays 353 stages on the first cohort and honest inference path; this count is reconstructed from the frozen driver's completed path and excludes its separate growth replay.
Driver timeout, completion and quorum recovery
The coordinator is killed after reserving the second cohort's first task. Its automatic restart recovers that reservation and completes all four remaining training steps. The frozen experiment driver then exceeds its one-hour wait while evaluation is still in progress. At the retained height 8,968, only 68 candidate fresh-response windows remain. The driver stops its child services; this is an experiment-controller timeout, not a completed quality evaluation.
A bounded manual continuation restarts the same validator homes, keys, signing state, workers, auditor, artifact store and operator outbox. It finishes those windows without new training or a genesis reset. The result records the resume height and continuation-helper digest. The second-host follower also needs a manual restart to reconnect promptly after the long outage. All five nodes then agree in block 9,086. This demonstrates retained-state recovery with intervention, not fully unattended operational resilience.
After evaluation, three of four validators continue producing blocks. Half the voting power halts finality. Restarting the retained validators restores common headers in block 9,811, with application hash 00073FC58D6D2DB580AC832864C6313870EF0571FCC0CB8FCF4F16532EB740B2. All four final databases independently pass the supply invariant: 3,997,019,000 liquid atoms plus 10,000,000 validator collateral, zero remaining escrow and 10,981,000 burned atoms equal 4,010,000,000 initial plus 8,000,000 issued atoms. Both bounded trial networks are stopped with their state preserved. The existing public 0.4.0 network continues separately.
Recovery during a real fraud dispute
The remote auditor detects a deliberately forged output head. After three native evidence chunks have finalized, the failure injector pauses the auditor, checks its durable outbox and kills it while an upload has no locally recorded receipt. The restarted process recovers the same signed bytes and nonce 244. That operation then has a finalized receipt; recovery does not sign a new nonce to replace it. A missing local receipt does not mean the transaction was absent from the chain when the process died.
The auditor completes the 113,334,000-byte dispute. All validators use the native referee to reject the forged response at block 5,265 with the reason objective replay mismatch: forward evaluation result. This demonstrates recovery of an interrupted large evidence upload and an objective rejection, not a cheap neural proof. The auditor had detected the mismatch after four stage checks in approximately 5.79 seconds, including its object handling. The dispute comprises 116 unique signed transaction envelopes totalling 151,171,234 bytes. The interval between their first and last consensus block timestamps is 342.559577 seconds, including the injected crash. This is neither a monotonic client-latency measurement nor total network bandwidth: retries, RPC framing and peer gossip are excluded.
What this candidate does not establish
Funded reports remain attestations under an honest-observer assumption. The adversarial tests deliberately settle a forged update when the selected auditor colludes and no observer challenges it. An objective challenge rejects the same forgery and slashes a false revealed report. A cheap coverage digest, token conservation and a paid-service counter do not prove honest computation.
The prototype's global pool can be filled by 16 unaccepted, refundable offers. The admission-saturation test demonstrates that another sponsor is then denied access until an offer is cancelled or expires. This is an unresolved admission denial of service, not a fair permissionless market. Selected-auditor withholding, challenge griefing and complete future-cohort funding remain separate limits.
Acceptance also fails to reserve separate refutation capital. An adversarial test gives an auditor enough for its audit bond but insufficient liquid funds for the challenge bond. It detects fraud and refuses to attest; the claim cannot mint, but the missed-report rule burns its bond. The real trial starts its auditor with 100,000,000 atoms. Its measured head dispute needs another 2,000,000 atoms of challenge collateral and 116,000 atoms of transaction fees while the 5,000,000-atom audit bond is locked. Prefunding that obligation is a required protocol change, described in the admission RFC.
The real inference fixture generates one token on the retained seed, placed in three partitions. One complete auditor covers those partitions and the output head: 400,000 atoms of audit fees plus two 1,000-atom publisher submission fees, against a 1,000-atom payment. A four-partition serving model would instead require 502,000 atoms for those costs. The later operator refuses this subsidy unless explicitly configured. Its default one-token bound fits within the native eight-token maximum; practical chat throughput has not been established.
The ledger still needs independent ownership, resilient hosting, fair resource admission, artifact retention, sustainable prices, useful held-out improvement and a rehearsed upgrade or launch procedure before a production cutover. These experiments exercise a funded settlement candidate, not a permanent mainnet.