affine

Bittensor’s subnet page labels netuid 120 as registered on June 29, 2025 at block 5,749,344. That chain event created the durable slot; it did not publish Affine’s later rulebook. The current onchain identity is Affine, with alpha symbol Yad (ⴷ), while CoinGecko uses SN120. A SymbolUpdated event is visible but its prior value is not, and no reviewed primary record establishes another former project name.

Affine did inherit a technical past. Its own project file says the research began by testing kings from Albedo, subnet 97, before becoming the SN120 validator and evaluation service. The distinction matters: Albedo was the experimental panel, not an earlier name for Affine or the same alpha currency.

The Affine Foundation published its monorepo on August 4, 2026. The stated goal was a teacher-anchored score for coding trajectories rather than an LLM judge: miners would train a Qwen-family checkpoint, submit it, and challenge a reigning model for the emissions throne.

Five days later, the repository recorded an uncomfortable comparison. An archived Albedo panel of 30 models had shown rho +0.758, but the live SN120 panel of 29 models reversed to rho −0.421 with p=0.024. All three SN120 kings crowned by that score solved 0 of 25 held-out tasks. These were Affine’s own measurements, published on August 9, not an independent benchmark; publishing them forced the team to treat its meter rather than its miners as the first problem.

On August 21 the operator lowered the minimum winning margin from 0.002 to 0.001. During the next 18 hours, four near-noise challengers took the crown while their measured Reason scores drifted downward. The system was selecting lucky evaluation slices, not reliably better models.

The August 22 reversion restored 0.002 and left the disputed reigns in history, but noise was not the last flaw. Affine says the reign-41 model appended the same filler suffix to every thought, creating a flat lift that v4 rewarded without showing better task reasoning. Operator records date the response to August 27: min(R,G) v5 subtracts the plain mean so a constant lift cancels, then applies a two-sided Grounding band so filler below the teacher range and parroting above it both lose. The throne reset to untouched Qwen3.6-35B-A3B; the implementing commit became public on August 28. These were operator-authored scoring changes, not a vote by Yad holders.

Affine initially accepted model revisions through Hugging Face. A cutover posted September 2 and activated September 3 replaced that path with uploads to a validator-controlled Cloudflare R2 bucket. Miners now need an Ed25519 hotkey, receive temporary credentials, upload a signed manifest and reveal readiness onchain; ordinary sr25519 hotkeys cannot decrypt the sealed credentials.

The change makes copying challengers before evaluation harder, but concentrates intake and model access in the validator service. This was more than a notice: the snapshot generated September 4 showed queued registrations with temporary credentials, and duel history recorded r2:// challenger paths. The evaluation corpus itself is public. A duel draws 1,300 turns from a seed based on the reveal block hash and challenger hotkey, making the exact sample unpredictable before reveal and reproducible afterward. Crowned checkpoints are published; unpublished losing weights stay private and are scheduled to expire after 60 days. Inference still runs through Targon or operator-managed hardware. Publishing the inputs and selection rule makes sampling inspectable without handing over control of the evaluation service.

Dynamic TAO gives SN120 its own alpha currency and TAO/alpha pool. Staking TAO creates Yad exposure; unstaking returns TAO at the pool’s current ratio. Alpha supply grows through Bittensor emissions toward a 21 million cap. On the page reviewed September 5, 2026, emissions split 45.2% to miners, 36.8% to validators and stakers, and 18% to the owner; those percentages are a dated state, not permanent token terms.

One model reigns at a time, but it is not the only possible miner-reward recipient. The current published state code selects up to five accessible, distinct kings, allowing a deeper king to replace an inaccessible recent one. Weight setting then drops unregistered keys and duplicate UIDs, shares the reward weight equally among those left, and uses the burn UID if none remain. This is separate from the owner’s 18% share in the dated snapshot and the observation of two active validators. The owner coldkey also changed twice in August. Affine’s dashboard makes rule changes and defeats unusually visible, but transparency after a decision is different from shared control over it. Yad ultimately prices confidence that a small referee set will keep correcting the test faster than miners learn to game it.