Competition rules and FAQ

How an entry is evaluated

Commitment and queue

• Commitment. You commit on chain with a hotkey registered on subnet 92: reax-spec/1:<track>:<repo>@<40-hex revision> for model tracks, reax-eng/1:s1-fast:<image>@sha256:<digest> for the engine. One competition commitment per hotkey; a new commitment from the same hotkey replaces the old one. Earlier commitments are handled first, and a commitment for a track that is not active is ignored.
• Queue. Our competition worker runs an evaluation cycle about once an hour and evaluates up to 8 models per hour across the tracks, on one fresh draw per track. New entries wait in the order of their on-chain commitment; first evaluations of entries whose hotkey is already past its immunity period go first. Engine entries are measured in batches, with a new cycle every 6 hours. Statuses are explained on the live page.

Draws, first stage and confirmation

• Draws. Each draw has 300 fresh sealed items. Score = 0.75 × competence + 0.25 × calibration. An item is used at most 24 times and then retired; only retired items are published later. A confirmation draw is disjoint from your first draw. Items, seeds and the scores on private items are never published.
• First stage. A challenger has to beat the champion on the same draw: z ≥ 2 and a gain of at least 1.5 points pass. A close call (1 ≤ z < 2) gets one more first-stage draw, and z < 1 is not better. z is the gain divided by its standard error; evidence is pooled only over draws against the same champion.
• Confirmation. It is a second, disjoint fresh draw at least 12 hours after the first draw (counted from the first draw, also after a close call). You are crowned at a pooled z ≥ 2.75 over all your draws against that champion, with a pooled mean gain of at least 1.5 points.
• Crowning and ranks. The champion gets 75% of a track’s share, the former champion crowned most recently before it 17%, the one before that 8%. Ranks follow the order of crowning, not the scores. An entry at or below the entry floor earns nothing from the track.
• First champion of a track. It is decided only after all entries of the opening cohort have been decided; the cohort then plays off on a shared fresh draw. A near tie (within 1 standard error) gets one more shared draw, and a remaining tie goes to the earlier commitment.

S_eff and the gap penalty

• 20% of each 300-item draw are cross-author real-data items that are kept private and count in the score. The gap is your score on our generator-family items minus your score on the cross-author items.
• S_eff = S − min(10, 0.5 × max(0, gap − G_med − A)) with A = max(8, 2 × SE). The penalty is capped at 10 points in the first week. S_eff is the score every decision uses. For the generalist track the frozen median gap G_med is 14.84.

Keep-alive

• An entry that is pending confirmation (up to 24 hours after its first pass) or a new champion waiting for approval (up to 24 hours after crowning) gets the weight of a typical inference miner, if it scored above the entry floor and the non-procedural floor (generalist: 37.4). At most 8 waiting entries per track are paid this way, together at most a quarter of the track share (the 2-per-coldkey limit applies first).
• An entry still queued for its first evaluation gets a provisional keep-alive of half the median, from about 2 hours before its hotkey’s immunity ends, for at most 12 hours per hotkey and one per coldkey per track, together at most 5% of the track share.
• Generalist entries above 10B parameters need a score at least 1.5 points above their own unmodified base to qualify.
• For 7 days from the generalist opening, the champion and approved former champions of demand forecasting and risk assessment get a bounded legacy keep-alive, as pure entrants only.
• A keep-alive is not an award. It does not reserve a UID or guarantee that a hotkey is never deregistered.

Derived entries, exact copies and the engine track

• Exact copies (same tensor hash) are refused; the earliest commitment owns a hash. Re-sharding or renaming files does not hide a copy. For generalist entries the check also uses a hash of the backbone that leaves out the readout head.
• Derived entries. If you declare another entry as your base, or our checks find that your weights or answers derive from an earlier entry (the champion included), your entry is derived. It must beat its parent by the full decision margin and, on competence alone, by at least 1.0 point and 2 standard errors, so a copy that only recalibrates the same answers cannot be crowned. Independent fine-tunes of the official base are not affected.
• Near-champion rule. A challenger very close to the champion (cosine similarity above 0.995, relative distance below 0.15, or the champion’s own weight change plus added noise) must also show a net correctness gain and clear a margin floor of max(computed margin, 3.0 points) on every deciding draw.
• Engine track. Every submission is measured against the reference engine on the same GPU in the same session. It must pass the correctness gate and be at least 3% faster than the reference; a challenger takes the title only with at least 3% more than the champion, confirmed in a second fresh session, and ties go to the earlier on-chain commitment. The image label ai.reax.engine.hotkey binds a digest to the hotkey that committed it. When crowned, the champion gets the whole engine share (20% of miner emission), which then halves linearly over 14 days. The previous champion gets 17% and the one before 8%, each for up to 7 days after losing the title, after its own award review. The opening round ended with two co-champions, UID 4 and UID 198, who split the champion share 50/50 as a one-time exception.

Coldkey limits, dual-role and multi-model hotkeys

• Coldkey limits: at most 2 open submissions per track and 4 new submissions per 7 days for each coldkey. Generalist entries are counted separately from demand forecasting and risk assessment.
• Dual-role and multi-model hotkeys (from block 9331987, 21 October 2026, about 13:00 UTC). From that block one hotkey can serve inference and hold one competition entry at the same time, and one hotkey can carry several inference models (2 to 4, commitment reax-s1m/1), each with the same per-model cap as a separate hotkey. A hotkey publishes one axon and still holds only one competition commitment. The keep-alive of a dual-role entry is a top-up over its serving weight. Before that block a competition hotkey does not serve traffic. See DESIGN-MULTIROLE.md.

Awards

• What an award approval is. A winning entry is paid only after it has been checked and approved: the engine champion, and every paid rank of a model track (ranks 1, 2 and 3). The approval covers exactly one entry, the track, the hotkey and the submission (repo@revision, or the image digest for the engine). A new revision or digest needs a new approval. Until it exists, the share goes to inference, nothing is burned, and the next rank does not move up.
• Who decides. The subnet owner. Approvals are signed with the owner’s key and published in the public repo (deploy/approvals/approvals.json); validators accept only a valid signature for exactly that entry.
• The review. The automatic copy and near-duplicate checks must pass first; an entry whose base cannot be verified is pending verification and is neither evaluated nor paid. Then we look at provenance (where the weights or the image come from), originality, and credit for merged or reused work. On the generalist track, per-area scores are information only; a model far below the reference in one area gets a closer look.
• While a champion waits. On a model track the former champion keeps the champion share for up to 48 hours while a new champion awaits approval. On the engine track the most recently crowned approved former champion is paid the champion’s part meanwhile.
• Timing. Award decisions are made between 10:00 and 22:00 Berlin time. The expected decision time for each champion is shown on the live page; it is an estimate, not a promise.
• Where to see status. Queue position, status, pending confirmation times and recent draws per entry: reax.co/competitions/live.

Validators

• Current release: v1.2.0-rc10.post1, a validator-only follow-up to v1.2.0-rc10 (same weights version). It carries the generalist track (opening 21 October 2026, 13:00 UTC), multi-model and dual-role hotkeys from block 9331987, the 7-day legacy keep-alive and the engine former-champion rules. Please run it before 13:00 UTC on 21 October 2026. Validators that are not upgraded cannot read multi-field commitments and will not score those hotkeys.
• How to upgrade. In your reax-subnet checkout, fetch and check out the release tag, rebuild the images and restart the stack from deploy/validator-1.2 (VALIDATOR.md). Competitions run in remote mode: a competition worker holds the sealed pools and the docker socket, and the validator only reads its state.
• Owner approvals. Keep deploy/approvals/approvals.json current (git pull of a checkout, or copy the file into your approvals directory) and check it with reax-award-approve verify. The file is re-read when it changes; no restart is needed.
• Releases and release notes: github.com/reaxlabs/reax-subnet/releases.

How to talk to us

Rule and status questions: ask in #92 on the Bittensor Discord, so everyone sees the answer. Direct messages only for private matters: keys, ownership proofs, code releases. We answer once an hour between 06:30 and 21:30 UTC.