Native training-to-inference result
Checklist item 3 passed on 2026-09-16. The operated candidate completed actual expert training, funded full execution audits, exactly-once issuance, a separate quality promotion, and inference paid from earned NEURO. The public evidence release contains the signed ledger, numerical reports, available-tensor catalog and replay instructions. A machine-readable result binds the source, genesis, graph and archive checksum.
| Check | Measured result |
|---|---|
| Accepted training | 140 four-update windows; 560 distinct updates |
| Native issuance | Exactly 560 NEURO |
| Single-answer quality | 7/64 before, 57/64 after |
| Composed-answer quality | 0/32 before, 23/32 after |
| Retained inputs | 2,048 unchanged committed computations |
| Earned-token inference | Two neural calls; 1,313 atoms paid, 11,615 refunded |
| Application replay | 1,728 signed transactions; 16,359 headers; four validator states |
| Available retained tensors | 2,856 objects, 228.26 GB; anonymous access verified |
The quality result reproduces the previously published expert and its original acceptance rules. It is not a new learning cohort. Retention preserves the original computations, including their errors. The paid response was neuroshard-ai; Ed25519; every actual call and its payment is in the ledger.
The forged measurement was rejected by actual execution audits. Withholding checkpoint bytes blocked payment until restoration. Repeated work did not mint again. A native restart and recovery of settled update 472 onto replacement hosts preserved the same genesis, checkpoint and paid-work history. The resumed job completed only the remaining 88 updates. Application replay matched the exported header commitments and saved states; it does not authenticate an independent remote consensus quorum or repeat GPU execution.
All validators and workers belonged to one administrator. The initial serving check used five owners across three AWS availability zones; recovery used one zone. The finite experiment's GPUs, volumes and temporary security group were retired after preservation. The three integration attempts cost a conservative $31.96 in compute, including failures and recovery. Storage, earlier learning and the separate backend preflight are outside that estimate.
The six-item checklist is now 1/6 complete. Automatic useful composition, further admitted learning cohorts, independent reliable providers, complete operating economics and usable public chat remain open. The candidate result does not change the deployed 0.4.0 network or establish a live global assistant.