First verified Fez testnet round
On September 25, 2026, Fez completed a fresh training → signed submission → validator evaluation → timelocked weight publication → on-chain verification round on Bittensor testnet subnet 579. This was a closed rehearsal with three controlled miner identities and one validator on one host.
The aggregate evidence records checkpoint hashes, methodology, scores, the successful transaction, and observed chain weights.
Results
| Miner UID | Correct / 224 | Family-macro Brier ↓ | Requested weight | Observed chain value |
|---|---|---|---|---|
| 1 | 183 (81.70%) | 0.273177 | 34.09% | 65,535 |
| 2 | 180 (80.36%) | 0.278502 | 33.60% | 64,598 |
| 3 | 180 (80.36%) | 0.292567 | 32.31% | 62,125 |
Weights use max(0, 1 - brier / uniform_brier), normalized across eligible candidates. The chain stores maximum-scaled integers, not percentages. Their normalized proportions matched the SDK-quantized requested vector within 2 / 65535. Accuracy alone does not determine the allocation.
Chain evidence and interruption
The weight commit succeeded in block 8,081,301, extrinsic 8081301-0008: 0xccf932d531bd995c6e87dd8e4749e8f0996cc9d24fe92ff388e36ba4a582b36c. The validator hotkey was UID 0. Commit–reveal remained enabled.
The RPC connection dropped while awaiting confirmation. The SDK returned an uncertain result that Fez initially labeled failed. Inspection of the block found both TimelockedWeightsCommitted and ExtrinsicSuccess. The original receipt was preserved and reconciled with that evidence; no transaction was resent. Fez now records unknown SDK outcomes as unknown, with no automatic retry. The regression check and all 12 project tests passed.
After reveal, Fez verified the actual weights at block 8,081,345, observed at 2026-09-25 06:40:01 UTC. Verification checked registered identities, a newer validator LastUpdate, the exact target UIDs, and normalized weight proportions. Transaction inclusion alone was not treated as proof of revealed weights.
Method and limits
Each miner trained a distinct checkpoint from the same pinned 0.8B reference: Qwen3.5-0.8B-Base, revision dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68, with rank-16 LoRA and a 256-dimensional decision head. Training used one epoch over 224 examples, learning rate 2e-5, microbatch 1, accumulation 4, FP32, and a seed derived from each hotkey and the fresh round identifier.
The validator downloaded and verified the signed artifacts, fitted temperature on 112 separate calibration questions, and evaluated each calibrated checkpoint on the same 224 test questions. Calibration minimized equal-family NLL over 81 temperatures from 0.25 to 4. The synthetic benchmark was reused development data; it is not an independent generalization test.
All four processes ran on an Apple M4 Pro with 24 GiB unified memory using MPS. A shared compute lock serialized GPU work. This establishes the registered training-to-chain path, not participation by independent operators or isolation between hostile miners. The private inputs prevent exact reproduction of these scores from a public checkout; the testnet guide describes how to run a new rehearsal.
The one-round services exited afterward. The subnet remains registered and activated, but this report does not claim continuous fleet availability, mainnet deployment, miner earnings, an open competition, a model release, or automatic winner promotion. The separate 4B experiment was not used in this run.