REAX / COMPETITIONS / RULES / ROUND 2
Generalist track, round 2: rules
Rulebook version v2.9. These rules replace round 1’s rules for the Generalist track; the status below shows when. The rules for the other tracks stay on the competition rules page. Short answers are in the miner FAQ.
LIVE
Rewards are a protocol mechanism, not a promise of any amount, and can be zero.
Round 2 schedule
- · Round 2 rules take effect
- · Planned Round 2 commitment cutoff
- · Round 2 results target
- · Planned Round 2 commitment cutoff
- · Round 2 results target
- · Planned Round 2 commitment cutoff
- · Round 2 results target
0. The short version
- Commit one model at a pinned revision, with a lineage record naming your hotkey and weight parents or saying you trained from scratch.
- Approved starts: any model with public weights, an Apache-2.0-compatible licence and no Hugging Face Decision Index flag, of any architecture that fits one 96 GB GPU (2.4). Every entry is scored in its base’s own setup, as the Hugging Face Decision Index runs that base. Section 4 names the starts, and the start list shows which setups are served.
- We check lineage, run a preview and score selected entries in weekly windows. The first cutoff is Tuesday 24 November at 15:00 UTC. Section 5.3 gives the calendar and selection rules.
- If your repo starts private, reveal the committed revision after its first official scoring. Section 10 sets the objection deadlines.
- The shares and keep-alive are in section 6. Post-production picks can share an extra 1,000 alpha under section 7.
- Read the rules, FAQ, status and published results at mcp.reax.co. Questions go to #92.
1. Goal
We want one model that is clearly first on the Hugging Face Decision Index, and ideally also first among decision models on the Artificial Analysis Decision Index. The official score (section 6) measures what those indexes reward, on private items.
2. What you can submit
- An entry is a model committed on chain for this round, identified by its repo and revision. A new revision is a new entry.
- Parents plus your own work. An entry is exactly one model with one set of weights. Its weights come from one or more parents plus anything you did yourself: training, merging, soups, blends or averaging. Or they come from no parent at all (
none, trained from scratch). A fine-tune of one model has one parent. An entry has at most 4 parents, each with a share of at least 0.10 (3.1). Allowed parents are:- approved starts (4.2), named on the list in section 4 or named in your record under 4.3;
- earlier entries of this round of another holder that passed intake, are not void, were officially scored in a window whose results were published before your commitment, and were public (2.7) at your commitment;
- your own earlier entries and your own public models (same holder), as long as their weights come only from allowed parents.
Not allowed as a parent: a public model that does not meet 4.2, a flagged model (4.2), and round 1’s entries. All parents must have the same architecture (4.2). Weights of a parent’s own ancestors count as part of that parent, so moving your weights back towards the base is allowed. Averaging or merging checkpoints or seed runs you trained yourself counts as your own work, not as extra parents. If a parent entry becomes void later, entries already built on it keep competing with their bar list unchanged, and it cannot be a parent for new commitments.
- Not allowed in an entry: anything that runs more than one model at answer time: ensembles, routers, multi-model packages, combining the answers or confidences of several models, and models that call other models or tools. Also not allowed: unmerged adapters and code that must run with
trust_remote_code. Running several models together is our own decision in post-production (section 7). Declare every model you merged. A model our check finds in your weights that you did not name goes on your bar list as a weight overlap (3.6); leaving out a source of another holder on purpose is a misdeclaration (3.8). - Format: unquantised safetensors (bf16, fp16 or fp32) in a Hugging Face repo, pinned to one revision. A model that was quantised and converted back is refused, unless a parent already had that grid or you declare quantisation-aware training. Your entry is scored in the setup of its base (4.2; the start list shows each setup and its status): its prompt, attention mode and readout (how the model turns its output into option probabilities). Changes to tokens, layers or context are fine if the entry still loads and scores in that setup. A text-only export of a parent with a vision part is allowed. At most 45 billion parameters, so the model fits one 96 GB GPU in BF16.
- Inference settings: every entry is scored in the setup of its base (4.2), with that setup’s prompt, attention mode and readout. By default the entry runs as that setup runs its base. You may declare inference settings that change how your model runs without changing its weights, such as running layers more than once (“loops”) or reading the input a second time (“re-reads”), if the entry stays one model answering in one request (2.3) and uses at most twice the compute per item of its base’s setup, averaged over the items (the setup’s own steps, such as the conditional re-read in the setup of Jeffan-31b-1710 or a setup that reads the input twice, are part of that baseline, not of your allowance). The declaration format is published on the declaration format page (the
inferenceobject inreax-specialist.json; it also says how we compute the 2x). No declared setting is supported for scoring yet: an entry that declares one waits with the status “declared setting not supported yet”, keeps its commitment time and is processed once our engine supports it, as in 4.2. Layers you add by copying layers of your parents count towards the 45-billion-parameter limit, and the setup must serve the new layer count. - Licence: your model is published under Apache-2.0, with the licence text and NOTICE files of your parents kept, and every parent’s licence must allow that.
- Private first, public after scoring (commit and reveal).
- Commit privately. The normal way is a repo that is gated on Hugging Face with manual approval, with access granted at least to florian-chutes (Hugging Face lets a repo owner add one specific account); anyone else you let in is your risk. README, config and the lineage file may be visible; the weights need not be. A public repo is also fine. Your commitment pins the revision, and the revision binds every weight file and your lineage file, so nothing else needs to be hashed.
- Everything runs on the private weights: intake (3.2), preview and official scoring. Only our evaluator account reads them. We log every download we make of a non-public repo and publish the log with the window result. We never pass private weights on, and we delete our copy of a void private entry (bar lists that include it then follow 3.6). After the round we delete our copies of entries that were never public, except what the lineage check needs (its measured sample).
- Before the reveal, the scores of a private entry are not public. Among miners, its preview and official scores are shown only to you, the submitting miner (through the MCP, signed with your hotkey); we and the agents that run our evaluation see them, since we compute them. Your public uid page shows your intake result, your bar list and “private entry, scored” without any score; public results list the entry as “private entry, scored” without a score or a place. A place it holds stays empty in public until the reveal (the places of other entries do not change), and a title it would win takes effect only with the reveal (below). With the reveal its scores and place are published. An entry that is never revealed is void (below) and its scores are never published.
- Reveal deadline: an entry must be public (anyone can download the committed revision without logging in) within 24 hours after the results of the window in which it was first officially scored as a selected entry are published. A reveal or objection deadline that falls on a Saturday or Sunday moves to Monday 16:00 Berlin time. We check whether it is public ourselves, without logging in, every hour until the deadline; another revision or another repo does not count. An entry that is never officially scored may stay private; it is never ranked. If our anonymous check cannot reach Hugging Face around the deadline, the deadline moves to 3 hours after our first successful check. The reveal result is a window result (10.2).
- A share starts when the entry is public. A share an entry wins while it is still private (6.6) is paid from the time we see it public, not before; until then that part is burned (6.6) and not paid later. Exception for the title: a title won by a still-private entry takes effect when we see it public; until then the champion it beat keeps the title share.
- Not public by the deadline: the entry is void for that window and every later one. It loses its rank, title and places from that window, and every reward not yet paid; there is no other penalty. We do not replay the window (6.4): if it was crowned, the champion it beat keeps the title (it was not dethroned), and if there was no champion, the title stays vacant; places 2 and 3 move up. A champion that keeps its title this way was never unpaid (see above); a place that moves up is paid from the publication of the void.
- After the reveal we keep our own copy. Making the repo private again later changes nothing in rank or rewards.
- Open to build on: from the publication of its window result and its reveal, other holders may use its weights as a parent (2.2). A child committed after that and ready by the next cutoff can be scored the following week.
- We cannot read the repo at your commitment (not public and no access for our account): the intake check retries every hour. If it is still unreadable 24 hours after your commitment, the entry is refused as not readable: no penalty, and you may commit again (the commitment counts towards the limit in 2.8). A temporary error on our side or at Hugging Face never refuses an entry. When an entry we could not read becomes readable, we also check again, against it, every entry committed after it that we already checked.
- Limits: two active entries per uid, not counting an entry that is champion. A new commitment replaces the older of the two non-champion entries as described in 5.3; a champion entry stays active until it loses the title (6.4). At most 2 new commitments per coldkey per 24 hours. A further commitment is not checked and does not replace your active entries. Every well-formed commitment we check counts, whatever its result, except a refusal because the record names no hotkey (3.4); a commitment we cannot parse, or one over the limit, does not count. The 24 hours are rolling, by block time; commitments made before v2.3 count towards them.
- Holder: entries belong to one holder if they have the same coldkey, or if they share the Hugging Face account or organisation in the repo id of an entry that passed intake and whose lineage record names the committing hotkey, also through a chain of such links. A commitment of someone else’s repo creates no link. For the intake check of an entry, the Hugging Face account in its own repo id also counts as yours when its record at the committed revision names the committing hotkey (a later correction alone does not create this link), so your own earlier public models count as your own work from your first entry on.
3. Lineage
- One record. Put
reax-lineage.jsonat the committed revision. Your commitment pins the repo and the exact revision of the weights and the record. The record needs:- your committing hotkey;
- your weight parents, or
"from_scratch": trueif you used none. Give each parent’s repo@revision and approximate share. For another entry give its uid. For your own earlier model give its weight parents too. An entry has up to four parents, each with a share of at least 0.10, and the shares add up to 1. If you fine-tuned one model, its share is 1.
Omit
entry.revision. The checker ignores it in older records, so you can keep the original commitment. A training note and architecture hints are optional. Lineage schema. - Intake check (stage 0), before any scoring, published on your uid page. We aim to finish it within about 6 hours of your commitment (of the time we can first read your repo, if that is later). This is a target time, not a deadline or a promise; if a check takes longer, the entry waits for it (a delay of ours that costs a window is handled by 5.3). It checks that:
- the record names the committing hotkey and the declared weight parents are allowed;
- the pinned weights and config can be checked against the setup of its base;
- the weights come from the declared parents. Every declared parent must be found.
The checker can use model IDs from your repo files as weight evidence. Incidental mentions, optional metadata and architecture differences that do not prevent scoring are notes. We compare against every model of the same architecture that was public or committed before your commitment: the list, the public releases of that family, every earlier entry of this round (all revisions, also withdrawn ones, also those still private), round 1’s entries, REAX releases, the top rows and all flagged models of that family on the Hugging Face Decision Index, its most-downloaded public fine-tunes, and every model named in your repo. Method, tensor sample, thresholds and script: https://reax.co/competitions/generalist/lineage/, checker version 1.6.2, script hash 302932efe9a4d4f77a358e9bdcc098fff87665f02557e021fcac0858694cf3f1. Use the published script locally to check your weights before committing, including for private repos. Once signed dry runs open on the MCP, you can request up to three per hotkey per day. Give the public repo and exact revision in the signed request. The dry run compares only with public models and revealed entries. The intake check also uses a sealed sample, so a dry-run PASS is not a promise.
- Results:
- PASS: published with your bar list (rule 6). Scoring may start.
- NEEDS CORRECTION: a problem prevents checking or scoring the entry, for example missing or unconfirmed parent information, unreadable weights, an incompatible setup, or a transformation that prevents verifying the declared parents. Temporary read failures follow 2.7. Optional paperwork never opens a correction clock. Send a signed correction through the MCP within 24 hours. It cannot change the hotkey named by the pinned record and it keeps the original commitment time. You may instead commit a new revision. A dispute pauses the correction clock.
- REFUSED: an ineligible entry, for example prohibited weights or parents, more than four parents, a share below 0.10, parents of different architectures (2.2), a copy, an ineligible quantisation round trip, or a missing or wrong hotkey under 3.4. An entry that still needs correction after 24 hours is refused. A refusal by itself carries no penalty.
- MISDECLARATION: the cases of rule 8.
- Ownership. Of two entries with the same weights, the earlier commitment (block, then extrinsic order) owns them. Only a commitment whose lineage record at the committed revision names the committing hotkey can be the earlier committer or own weights. A commitment whose record at the committed revision names another hotkey, or none, is REFUSED (3.3: no penalty from this rule, and it counts towards the limit in 2.8, except when the record names no hotkey); it never owns anything and never turns a later commitment of the same repo@revision into a copy, however early it is on chain. A later commitment of the same repo@revision waits until the record of the earlier one has been read; if the earlier one then cannot be the earlier committer (this rule) or is refused or void, the later one goes on with its own commitment time. A REFUSED entry never owns weights. For weights that were public before any commitment, ownership goes to the holder of the account that uploaded them first (Hugging Face commit history). Publish your weights only after your commitment is on chain; committing privately (2.7) is the safe way. A copy of a still-private entry can only come from a leak; our check finds it in the same way, and the earlier commitment owns the weights.
- Copy: after we undo reparametrisations, at least 99.9 % of the measured layer and output-head values equal those of an earlier entry that owns those weights (rule 4), a list model or another public model within bf16 rounding, and every other weight file is identical. A copy of a model that your record or your repo files name (a parent included) is refused, with no penalty. A merge that is a copy of one of its parents is a copy. Committing a repo@revision that another holder committed earlier, where that commitment owns the weights (rule 4), or that is a model on the list, is a copy of it and is refused, with no penalty from this rule (rule 8 adds one case: committing a repo@revision whose record names the other holder’s hotkey, earlier than or within 3 blocks after that holder’s own commitment of it); it stays the earlier committer’s entry. A commitment that does not own the weights under rule 4 is never that earlier committer. A copy of a model they do not name falls under rule 8. A copy of your own earlier entry is a duplicate: not a new entry. If the original is no longer active (for example after a prune and re-registration), the duplicate takes its place and keeps its results.
- Bar list: every parent and ancestor back to the list; every earlier entry or public model of the same base family whose weights our check measures in yours but your record does not name (weight overlap); and twins found at your first official scoring (6.5). Honest fine-tunes of one base with the same public data look alike in their weights: we measured this, and our check cannot tell such siblings from a merge. So a weight overlap is never a penalty and needs no correction; the (at most three) models with the largest positive weight in a joint fit of your own change join your bar list; a model that is already on your bar list for another reason counts as one of the three; if the joint fit fails, we order by the measured overlap (lineage page). All overlaps are shown on your uid page; an overlap with a still-private entry is shown as “a private entry of this round (uid X)” until that entry is revealed. A copy (rule 5) is not an overlap. Your own entries and your own public models are left out, but their parents and ancestors stay. So a merge must beat every model it was made from. To be ranked, your entry must beat every model on its bar list on the same items by the crown margin (6.3). Otherwise it shows as “did not beat <model>”, earns nothing and cannot hold a title. We keep a copy of every PASS entry, so every bar-list entry is scored in each window in which it is needed; a public model on a bar list is scored in the first window it is needed and its last official score is used after that. If a model on your bar list has no score in a window (for any reason), we use its last official score instead (a model’s official score in a window is its first-draw score), and you must beat that by the crown margin. If it never had an official score and we could not score it in time, the entries that need it cannot be crowned in that window and go first in the next; this is not a loss under 10.1. If it can never be scored (we deleted it under 2.7, or it failed twice under 5.5), it leaves your bar list and its own parents and ancestors stay on it. Nothing else is required of a derived entry. Its parents and ancestors earn nothing extra from it (also not from the 1,000 alpha, 7.4).
- Honest omission found later: if an objection upheld in time (10.2) shows an omitted weight parent or overlap, we correct the record and recalculate the bar list using stored scores from the day of the decision. An allowed parent or weight overlap carries no penalty. If the entry no longer clears the bar list, it loses rank, title and unpaid rewards. A champion that loses the title this way holds a former-champion place (6.6). Rewards already paid stay paid. We apply the same bar-list correction to descendants.
- Deliberate misdeclaration means:
- (a) deliberately hiding a weight parent, including an undeclared copy of another holder’s weights (3.5) where the evidence establishes deliberate hiding;
- (b) disguising another holder’s weights, for example by reparametrisation, permutation, rotation or a quantisation round trip;
- (c) documents such as posts, revisions or logs showing that a parent was deliberately hidden.
A missing training note, a model mention or a statistical overlap is not evidence of deliberate misdeclaration on its own. Committing another holder’s repo@revision whose record at that revision names that holder’s hotkey, earlier on chain than that holder’s own commitment of it or within 3 blocks after it, is such an undeclared copy under (a) (rules 4 and 5). If that holder has not committed it yet, only rule 4 applies for now; if it commits it later, we send the finding then, and the 72 hours start when the uid page shows “refused, finding sent”. Besides the refusal of rule 4 (or, if it had already passed, the hold), we send a finding and decide as in every other case under this rule (the refusal itself is not a penalty; a confirmed finding has the consequences below). A refused commitment is not held and gets no keep-alive; its uid page shows “refused, finding sent”, and the 72 hours below start when it shows that. A later commitment of it is a copy under rule 5 (refused, no penalty). We prepare the finding with the measurement and evidence. Meanwhile an entry that had already passed is held and shown as “held, under review”: it earns only the keep-alive, and a held champion keeps its title and is still scored, but its share is withheld (burned, 6.6). A held entry that is not champion is not selected while it is held; if the hold ends without confirmation, it goes first in the next window, as if it had overflowed. You get the finding through the MCP, signed for your hotkey, and your uid page shows “held, under review: finding sent”. The 72 hours start when the uid page shows that. Answer through the MCP’s dispute tool, naming the entry; I read your response. I decide at the first Thursday slot that starts at least 24 hours after your response or after the 72 hours. If I do not confirm the finding, we pay the entry the withheld amount by transfer from our own share within 7 days, and the case is handled under rule 3 or rule 7. If I confirm it: the entry is void; the holder’s other entries leave the round; its unpaid rewards of this round are forfeited; its paid rewards of this round are deducted from any later payout to that holder; and its entries are not used in post-production. The finding is published on the uid page. Disguising your own earlier entry counts as a duplicate. Found by us after a window is closed (10.5): if we ourselves find a deliberate misdeclaration or a manipulation (section 11) later, the entry is held from then on, and this rule applies to it from then on. Closed windows stay as published and are never re-ranked.
- Final. The lineage result of an entry is final at the objection deadline (10.2) of the window in which it was first officially scored as a selected entry (twins found at that scoring included). After that it changes only in two ways, and only from then on, never backwards: through an objection filed in time in a later window in which the entry is scored again (it then counts from that later window on), or through our own finding of a deliberate misdeclaration (rule 8). An objection about lineage needs evidence: a model that the published check finds in your weights, documents that contradict your record, or a rerun showing our check used wrong inputs. Other thresholds, similar answers alone, or opinions are not evidence. An objection or dispute filed in time is decided on its merits; finality waits for it. A change to sections 2 to 4 or to 6.5 applies only to commitments made after it is published (except that v2.3’s wider starts, 4.2 and 4.3, also apply to every earlier round 2 commitment whose intake result is not final, and v2.4’s and v2.5’s setups, 4.2, apply to every entry not yet officially scored; v2.6’s limit of two active entries, 2.8, applies to every current entry, and an entry replaced before v2.6 stays replaced), and never changes a result already on a uid page, except that v2.2’s intake simplifications apply to every current Round 2 commitment. Closed results stay published.
- Disputes about a lineage result follow section 10.
4. Starts and training data
- We publish the approved starts with repo@revision, licence, family, date added and setup. The list at your commitment time controls which weights you can start from.
- Which models are starts. Any model is an approved start when all of these hold:
- its full weights (not an adapter) are public on Hugging Face at a pinned revision, in the format and within the size limit of 2.4;
- its licence is compatible with Apache-2.0 release (2.6);
- neither it nor any ancestor carries a note on the Hugging Face Decision Index that a test or test items are not counted for it (a flagged model);
- it is not an entry of round 1 or of this round (earlier entries can be parents only under 2.2).
Any architecture qualifies. The lineage check (section 3) still runs on every entry, for its bar list and for copies; it does not decide whether a model is a start.
Setups. Every entry is scored in one setup: the prompt, attention mode and readout with which the Hugging Face Decision Index runs its base, reproduced by our engine. Which setup an entry gets:
- a fine-tune (one parent): the setup of that parent; a start has its own setup, and an earlier entry or your own earlier model has the setup it got under this rule;
- a merge: the setup of the parent with the largest share declared in its lineage record; with equal shares, of the parent named first (so the shares and order you declare choose the setup);
- an entry trained from scratch (
none), or a base the Hugging Face Decision Index does not run with a setup of its own (for example gemma-4-31B-it and Qwen3.8-27B): our documented default setup for its architecture.
Every model we score for comparison (parents, bar-list models, listed starts, the champion) is scored in its own setup under this rule. Where a setup reads its answers through a head file that ships with the model (for example the
readout.safetensorsof karpo-decision-31B), we use the head file in the repo of the model being scored, so you can train it together with the model; without one, we use the head file of the model whose setup it gets under this rule (for a merge, the parent with the largest share). The head file must have the start’s file name and shape and an allowed format (2.4); one that does not fit is an incompatible setup (3.3), and we do not fall back silently. A setup is served once our engine reproduces it, passes our parity check and runs previews and windows. The start list shows the status of every setup and, for each served setup, the UTC time from which it is served; that time counts, and we post each one in #92. Our documented defaults aregemma4for Gemma 4 31B (the code readout of Jeffan-31b-1710 with one read) andqwen38for Qwen 3.8 27B (code readout from the output layer, one request at a time).Architecture means the text model’s layer structure, layer count, hidden size and vocabulary. A text-only export or another head class of the same model (2.4) is the same architecture. We aim to serve the setup of every named start before the first cutoff (Tuesday 24 November, 15:00 UTC), karpo-decision-31B first; the start list shows the status of each. An entry whose setup, or the setup of one of its parents or bar-list models, is not served yet keeps its commitment time and is processed (intake, preview, scoring) once it is served. It then goes first in the next window whose readiness time it can meet, as if it had overflowed (5.3); that wait is not a loss under 10.1.
- Named starts. The current start list has each checked revision, licence, family, date added, setup and its status. On 21 November 2026 it names:
- Gemma 4 31B dense:
google/gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475tattrongvu/Jeffan-31b-1710@cebfa86c2bed2b5c83b12e44de4374b20c87fd79(earlier revisions of this repo had a non-commercial licence)MachinePulseAI/karpo-decision-31B@b8e6fd7463cea8662c556d6257955d7ef9c10d19PelaAI/KnowLine-31B-Gen1@7d21f7832d0db289f533cd0c22b2cc604cd925d5PelaAI/KnowLine-31B-Gen1.5@35e7d7a9073c83151a35b1ff93c4ba1b474e78e2(same weights as Gen1)PelaAI/KnowLine-31B-Gen2@fd99a9ed2255d8a67dd47a44cc429e638f611593sthanika-ai/Sieve-31B-v3@8fcee858f7f20a50bfa3c8f984c22adb3cbb46a2OmniJev/OneJev1.6g-31B@d6aa28a588ca189eef96daeb5afab1af632193e5Poindexter-Labs/Copernicus-31B@ae28efb0ae759c3494a1078c0a56eda10a9adb38turion-ai/turion-jev-31b@b32ba16788b8238365c4532d8a4a164af3cb5289logitshq/Logits-Sigma-1@5ee69b5ace03c6bdd4e3391ac722993baa7c8d2d
- Qwen 3.8 27B:
Qwen/Qwen3.8-27B@37561816e66b5b9153b1b64c10968c18c1e6f4a0tattrongvu/Jeffan-27B-1211@f0ce7043ca32ebfe6613615e573a5010bda7228avllm-sr/d3-ultra@37fee01926f3de396620e794201b074fb427ac2dapertixai/cerno-27b-v1.1@9dc1b3b3385544b3e3c28a082c9a5c1c38a0aea4Rover00/WeMM-JEV@a463f840c9259ff9139ab02181e01f20aca91b4b
Not a start: fastino/GLiDE-27B, every GLiDE version under the Fastino Community License, and every model built from them, whatever licence its repo states. That licence passes its conditions on to every model built from it (a separate agreement with Fastino for users above a revenue limit, and a mandatory attribution), so such a model could not be released under Apache-2.0 (2.6).
A model that meets 4.2 but is not named yet is a start too: name it in your lineage record at a pinned revision. Before your intake result we check it under 4.2. If it passes, we add it to the list as of your commitment time, and your entry is checked against it. If it fails, your entry is refused because that parent is not an allowed start; that is not a penalty, and you can commit again (the refused commitment counts towards the limit in 2.8).
- Gemma 4 31B dense:
- Trying a different approved start can help us combine models that make different mistakes. Sections 2.2 and 6.4 govern merges and the scoring bar.
- Adding a model: ask with repo@revision in #92 (we announce any other way to ask). We check it and aim to answer within 24 hours, also at weekends (an exception to 8.4). Asking is optional; it tells you before you train or merge whether a model qualifies. The list also grows on its own as entries name new starts (4.3). Every addition only adds options: it takes effect when it appears on the list, and we post it in #92; a model added under 4.3 counts from the commitment time of the entry that named it.
- Re-check: at least every Monday we compare the list with the Hugging Face Decision Index notes. A model leaves the list for new commitments when we see a note on it or on an ancestor. Entries committed before keep competing and can still be crowned, but we do not use them in post-production.
- Training outputs. You may use any model, API or program to generate training data or labels, including for distillation. You are responsible for ensuring its terms allow you to train a model that can be released commercially. Do not use test items from either index or outputs or labels derived from those items. An entry flagged by the Hugging Face Decision Index is ineligible. Using a model only for training outputs does not add it to your scoring bar.
- Training data. A one-line note on what you trained on is optional. We ask post-production picks for training-data provenance and terms before release.
5. How we evaluate
- Intake (section 3.2): automatic. Target time about 6 hours.
- Preview: every PASS entry gets one automatic preview: a score on about 1,000 public items (the self-measurement kit, for your orientation only) and on a held-out slice of about 300 items (5.3), shown to you and on your uid page as “preview, not official” (for a private entry only to you, 2.7). It also shows whether your entry beats its parents. We aim to deliver it within about 24 hours of PASS. Target times: the 6 and 24 hours are what we aim for, not deadlines, guarantees or promises; if a check takes longer, the entry waits for it. The deadlines of this rulebook (cutoff, reveal, objections, responses, Thursday slot) are unchanged by this. If our delay makes an entry miss a window’s readiness time, 5.3 applies; that is the full remedy. The held-out slice is drawn anew for each window, so a preview is a gate, not a target.
- Official window, once a week:
- Dates: the first window is Thursday 26 November 2026 (cutoff Tuesday 24 November, 16:00 Berlin time). The first window also scores and publishes the listed starts whose setup is served, up to the limit in “Listed starts” below, so you can see the level to beat; these are official scores (which listed starts later windows score: below). The bar in each window is set on that window’s items (6.4). The schedule on this page lists every window.
- Your entries for a window are your two latest entries that are ready by the cutoff below; both go through selection like any other entry. A newer entry replaces the older of the two from the first window at whose cutoff it is ready. An entry that is champion is scored as champion and is not one of your two.
- Cutoff: an entry is in the window if it was committed by Tuesday 16:00 Berlin time (block time) and has PASS and a preview by Wednesday 08:00 UTC; otherwise it is in the next window. If an entry committed by the cutoff misses the readiness time only because our intake or preview was not done in time, it goes first in the next window, as if it had overflowed; that is the full remedy and not a loss under 10.1. A corrected entry keeps its commitment time for ordering, not for this readiness time. An entry whose setup is not served yet (4.2) keeps its commitment time and goes first in the next window whose readiness time it meets after its setup is served, as if it had overflowed; that is not a loss under 10.1.
- Selection: every entry whose preview beats its parents (for
none: the listed start with the best held-out preview score). Beats its parents means a higher preview score than each parent’s on the same held-out preview items (we rerun your parents on them), compared at the two decimals we publish; an equal score at those two decimals does not beat. Selection and the no-harm bar (6.2) use a held-out slice of the preview: about 300 private items, not in the published self-measurement kit, drawn anew for each window and left out of that window’s official draw. An entry’s preview uses the slice of the window it is ready for; an entry that moves to a later window keeps its preview. We publish the draw hash of each slice before the first preview that uses it; the slice items stay private. The kit items in your preview are for your orientation only. If more than the window’s places qualify: entries that overflowed last window go first, by commitment time, up to the places; then the best other previews fill the rest, plus every entry within 4.0 preview points of the last one taken by preview, at most 3 extra. Bar-list models take places too: when we fill the places, a selected entry takes one place for itself plus one for each of its bar-list models that is not already scored in that window. If an entry with its bar-list models needs more places than the window has in total, we score it anyway: its bar-list models that have an official score use their last one (3.6), and only those that never had one take places, in the order of the entry. An entry that does not fit in the places left is skipped, and filling goes on with the next one; the skipped entry waits for the next window, as if it had overflowed. We score the selected entries up to the window’s places, which are set by the window’s budget and published with the window’s results. An entry already officially scored and unchanged is rescored only if a place is free. If we cannot score every selected entry in time, the unscored ones keep their place and go first in the next window; this is not a loss under 10.1. - The champion, the listed starts named here, every bar-list model of the selected entries, and the selected entries are scored on the same fresh private items. Listed starts: every start in the first window after it is listed and its setup is served (4.2); after that, in each window, every listed start on the bar list of a selected entry plus the 2 starts with the best official scores so far. Starts scored under this rule use at most 16 official runs per window. Starts on the bar list of a selected entry go first, then by Hugging Face Decision Index rank (unranked starts last, by date added). Starts not scored go first in the next window. On the first draw, the best three entries that beat the champion by the crown margin, clear their bar list and passed the no-harm bar are scored again on a second fresh draw, together with the champion and their bar-list models. A public model on a bar list is scored on both draws of the first window it is needed in; after that its last official score is used on both draws.
- Results: we aim to publish the results, the crown and the new shares on a weekday, aimed at Thursday around 14:00 to 15:00 Berlin time, never on a Saturday or Sunday; if scoring is not finished by then, on a later day. This is a target time, not a deadline or a promise.
- If the Thursday slot does not take place, everything due then (crown and shares if the results are ready, disputes, cases) moves to the next working day (Berlin) at the same time. Nobody loses protection while waiting. Weeks without a window are announced in the schedule.
- What we publish: for every window, who was in it and why (cut, overflow, champion, listed start, bar list), aggregate scores to two decimals (also the same-skills and new-domains parts), standard errors, margins, for every scored entry whether it cleared the no-harm bar with its difference to the bar to one decimal and its lift over its best parent on the public kit items and on the official private items, and the difference between the two lifts in points; we flag a difference of more than 5 points (information only: no effect on the score, the places or eligibility; a parent that is still private is shown as “private entry”), the evaluation version, the validator release and the draw ids. Also, for every entry: whether it is public or private, its reveal deadline, and our download log of non-public repos (2.7); and the window’s objection deadline (10.2). For an entry that is still private we publish none of its scores, differences or its place, only “private entry, scored” (2.7); they are published with the reveal. Official items, draw seeds and per-item results stay private, and so do the held-out preview items; we publish the draw hash of each held-out slice (5.3).
- Failures: if a model fails to load or serve, we retry it once in the same window. A second failure means no official score this window. A champion that fails twice in a window loses the title (6.4). A failure caused by our infrastructure never counts as a failure of the model: the affected entries are re-run on the same items, and if that cannot finish before the slot, the window moves as in 5.3. If the champion still has no score at the moved slot through our failure, it keeps the title, nobody is crowned in that window, and every selected entry goes first in the next window. This is not a loss under 10.1.
6. Score, crown and rewards
- Score: our official score, built to track the Hugging Face Decision Index: 0.6 × same skills + 0.4 × new domains, measured on our private items, each model in its setup (4.2) (the evaluation version of each window is published with its results, 5.4). It ranks entries against each other on the same items; it is not a forecast of a place on the public board. The crown margin and the bar lists are in points of this score.
- No-harm bar: an entry that falls more than 3 points (on the 0-100 scale of those items) below the best of its parents (for
none: the listed start with the best held-out preview score), on the Artificial Analysis-style items of the preview’s held-out slice (5.3; the same items for the entry and its parents) cannot be crowned. We check it once, with the preview. - Crown margin: 0.5 points or twice the paired standard error, whichever is larger. A lead equal to the margin is enough. We decide on unrounded scores; the page shows two decimals and marks every case where rounding would change the outcome.
- Crown:
- A challenger becomes champion if, on the first draw, it beats the champion’s score by the crown margin, clears its bar list and the no-harm bar, and then confirms: on the second draw it again beats the champion and every model on its bar list by the crown margin.
- Among confirmed challengers, the highest mean score of both draws wins. Exact tie (equal at the published two decimals): the earlier commitment wins.
- While there is no champion, the bar is the best score among the listed starts scored in that window, on the same items, plus the crown margin; on each draw the best listed start scored on that draw takes the champion’s place. Scores of the listed starts we scored are published every window.
- The title belongs to the entry. If the champion’s holder commits a new entry, the champion entry stays champion and keeps being scored until another entry wins the crown by this rule. A new entry of the champion’s own uid or holder must beat the champion like any other challenger. A champion entry whose uid is deregistered loses the title.
- If a champion loses its title for any reason (a correction, a misdeclaration, a disqualification, failing to serve twice, an objection upheld under 10.2), the title stays vacant until the next window. We do not replay a past window and crown nobody retroactively. The bar for the next window is the best listed start. If a champion loses its title at a Thursday slot, the results published that day already use the bar of a window without champion: the best listed start scored in that window (as above). A new champion that is not public by its reveal deadline is void; the champion it beat keeps the title (2.7).
- Twins: at your first official scoring, if your entry’s answers on the official items agree with those of an earlier entry scored in the same window so closely that the two answer distributions differ by less than 0.005 bits (Jensen-Shannon, on at least 50 shared items), that entry joins your bar list. Your own entries are never twins of each other (3.6). We check this threshold against independent fine-tunes of one listed start before the first window and publish the result. If the check changes the threshold, we publish it before the first cutoff (Tuesday 24 November, 16:00 Berlin time) and it applies to every entry.
- Rewards:
- Track share: the Generalist track receives 30 % of the subnet’s miner emissions, as in round 1. Keep-alive (below) is paid from it first (at most 25 % of the track’s share, below). The rest is the podium share, split while the round runs. Only this track’s share is split this way; the shares of inference and of the other tracks are not part of it:
Place Share of the podium share Champion 70 % Last dethroned champion of this round 10 % The champion dethroned before that 5 % Place 2 of the latest published window 10 % Place 3 of the latest published window 5 %
- Opening window: from the publication of window 1 results until the next window replaces them, the podium share is 70 % for the champion, 20 % for place 2 and 10 % for place 3. The unused former-champion shares go to places 2 and 3 for this window. The usual eligibility and burn rules apply. From window 2, the table above applies.
- Each share is paid from the publication of the result that gives it (pay counts on chain about 1 to 3 hours later), for a still-private entry from when we see it public (2.7), and is shown on the live page.
- Former champions: a former champion’s place ends with the publication of the second window after the one in which it lost the title (a title lost outside a window counts as lost in the next published window), or earlier when two newer champions have been dethroned after it. Only champions of this round hold these places; round 1’s champions keep the keep-alive (12.1), so until a champion of this round is dethroned these places are empty. A champion that lost its title under 3.8 or 11.2, or because it was not public by the reveal deadline (2.7) holds no former-champion place.
- Places 2 and 3 go to the two best selected entries of the window (5.3) by official score on the first draw, other than the champion and the entries holding a former-champion place. To place, an entry must clear the no-harm bar and beat both the best listed start scored in that window and every model on its bar list by the crown margin, so a copy or a near-copy of another entry cannot place. Only entries officially scored in that window can place; an unchanged entry is rescored only if a place is free (5.3). An entry that is void (for example not public by the reveal deadline) cannot place; the next entry moves up. Two entries of one uid always belong to one holder (2.9), so the rule “one podium place per holder” below applies to them. Places 2 and 3 are replaced by the next published window.
- One podium place per holder: a holder holds only the first of its places in this order: champion, last dethroned champion, the one before, place 2, place 3. A former-champion place that its holder cannot hold stays empty; for places 2 and 3 the next eligible entry moves up.
- Unfilled places (no former champion yet, no entry clears the bar for place 2 or 3, a place its holder cannot hold) and every share not paid for now (an entry not yet public, 2.7; an entry held under 3.8, a held champion included) are burned, like every part of the miner emissions we do not pay out (FAQ “Is any emission burned?”). While there is no champion, the champion’s 70 % is burned too.
- Keep-alive goes to every uid that has an entry for the next window (5.3), or whose newest entry is in intake (an entry waiting for its setup or inference setting, 4.2 and 2.5, counts as in intake), in correction, held, PASS but not yet officially scored, or overflowed, and to former champions (round 1’s included, 12.1). It covers at most 2 uids per coldkey (former champions do not count towards this cap). A uid gets one keep-alive share however many of its entries qualify, and keeps it while one of them still qualifies and starts with our next weight set after the commitment (about 1 to 2 epochs later on chain). For an entry it ends when the entry is refused or void, when its preview shows it does not beat its parents, when its official result is published without a title, or when it fails to serve twice (5.5); and, unless the entry is selected or overflowed at that time, 10 days after its commitment or when the results of the first window it could enter are published, whichever is later. A dispute pauses the correction clock but not this limit. For former champions, the last champion included, it ends 14 days after the round ends. After the last window, keep-alive for entries ends with the publication of its results; former champions keep theirs as above. Places 2 and 3 get no keep-alive after the run-out. A holder whose misdeclaration is confirmed under 3.8 gets no keep-alive for the rest of the round and its keep-alive period, for any of its hotkeys (2.9); a uid disqualified under 11.2 gets none either, and this applies to that uid’s hotkey, so a new commitment from it gets none. Each uid with keep-alive is topped up to the same weight, which we call the unit: it gets the unit minus what it already earns from serving, and nothing if it earns at least the unit; the live page shows per uid the unit, the serving weight and the top-up. Keep-alive in total is at most 25 % of the track’s share; if more uids qualify, each gets less.
- The round ends when the results of its last window are published. The last window is on Thursday 10 December 2026. Moving it is a rule change (section 9). After the round the track pauses over Christmas: no window takes place until we announce the next one.
- When the round ends or the track pauses, the podium as last published follows the run-out: each place keeps its full share for 3 days from the publication of the last window, then its share falls linearly to zero by day 7, then the keep-alive as above.
- Track share: the Generalist track receives 30 % of the subnet’s miner emissions, as in round 1. Keep-alive (below) is paid from it first (at most 25 % of the track’s share, below). The rest is the podium share, split while the round runs. Only this track’s share is split this way; the shares of inference and of the other tracks are not part of it:
- Rewards are a protocol mechanism, not a promise of any amount, and can be zero.
7. Post-production and its reward
- Post-production is ours. After the round we build the final model at our sole discretion.
- An entry that was never champion may end up in the final model, and a champion may not.
- We use only entries that were public by their reveal deadline (2.7; an entry never officially scored as a selected entry: public when we build the final model), with a PASS, that are not void, with no flagged ancestor at release time, and models from the list. A weight overlap is not an ancestor.
- Reward: an extra 1,000 alpha in total, on top of the emission rewards of 6.6, split equally among the picked entries:
- Picked means an entry whose weights make up at least 10 % of the final model’s weights, measured by the merge coefficient we publish with the release. (“Selected” in 5.3 means chosen for a window; it is not the same.)
- Material used only for training outputs and entries below 10 % are credited as appropriate for their contribution. The extra reward follows the weight-share rule above. Parents and ancestors of a picked entry are not picked through it. They earn nothing extra from it; they are protected by the bar list, because an entry built on them must beat every one of them.
- The split is per entry; one holder can receive several shares.
- 7.4 and 7.5 do not change after the first window’s results are published.
- One payment of 1,000 alpha for this round, for the first final model we release by 20 December 2026. It is paid within 7 days of the release, by transfer to the coldkey that committed each entry that receives a part, whatever its uid status, and reconciled by transfer receipt. If the final model contains no picked entry, or no final model is released by that date, this reward is not paid.
- Our release lists every part (repo@revision, weight hash, role, share and licence), records the training provenance needed for release, and credits the entries and material we used. Entering grants us a licence to use your entry under its published licence, never ownership.
8. Questions and answers
- Where to ask: our MCP at mcp.reax.co lets your agent read the rules, the FAQ, competition status and all published results at any hour. The MCP readiness guide shows whether intake and dry runs are open. The signed routes described in these rules are available only once the corresponding feature opens. This includes dry runs, private results, corrections, disputes, objections and findings. Ask questions preferably through the MCP (8.2), or in #92; we announce any other way to ask in #92 and on this page. DMs are handled only at my review times; they are not faster.
- Asking through the MCP: for an instant answer, call
search_docsorget_faqfirst;ask_questionis for what those do not cover. Signask_questionat mcp.reax.co with your registered hotkey, and read the answer withget_question_status(orlist_my_questions) using the same hotkey. Only that hotkey can read the answer. Each answer is drafted from the published rules and FAQ and checked by me before it is sent; no model answers your question yet. Answers come usually within a few hours between 07:30 and 22:30 Berlin; a question that arrives at night is answered the next morning. We say in #92 and on this page before automatic answers start. - I review open questions on working days at fixed times listed on the status page. Disputes and cases go to the Thursday slot.
- When to expect an answer: questions in #92 are answered by me at my review times, usually within two working days; the MCP lookups work at any hour. There are no answers in #92 at night or at weekends. Questions sent with
ask_questionare usually answered within a few hours in the daytime (8.2). - Only the published rules and FAQ decide. When I answer a question that interprets a rule, I add it to the FAQ the same day, and from then on that FAQ entry binds everyone. Other FAQ entries only restate rules that already exist; a new rule is a rule change (section 9).
9. Changes to these rules
- All rules are written down here. We may change any of them at any time, also at short notice. Section 3.9 (lineage-related changes), 7.4 (the reward) and 10.5 (closed windows) say what a change cannot touch.
- Every change is posted in #92 and on this page with a new version number, its date, the window it first applies to (which can be the current one) and a change log. If it names no window, a window is decided under the version in force at its cutoff. A change cannot name a window whose cutoff has passed, unless it corrects our error or only favours miners. Earlier versions stay available.
- A rule change never takes away what you earned under an earlier version. We never backdate a rule, and a change never alters a result already published on a uid page, except to correct an intake result affected by v2.2’s removed paperwork requirements. We replay that intake at the original block and extrinsic order. We retain the old receipt, apply any independent blocker, and preserve priority under 5.3. Closed windows and earned rewards stay as published. (Deductions after a confirmed misdeclaration follow 3.8.)
10. Mistakes we make, objections and disputes
- Our mistakes: if a result was wrong because of our error, we correct it from the next window and pay the measured loss from our own share. The measured loss is the share (6.6) the entry would have received under the corrected result, for the days it was wrongly denied. A window that moves to a later day, or a delay, is not a loss. Amounts already paid are not taken back. If your hotkey was pruned because of our wrong weights, our timing or a bug on our side, we refund what you paid to register it again. A refusal of your commitment that was caused by an earlier commitment which could not own the weights (3.4: its record names another hotkey, or none) is our error: we lift the refusal, your commitment keeps its original commitment time and needs no new commitment, and the refusal does not count towards the limit in 2.8. If your entry then misses a window’s readiness time, 5.3 applies.
- One clock per result. Any registered miner may dispute a result, about their own entry or another miner’s, through the MCP’s dispute tool, signed with their hotkey:
- Intake results (published at any time: PASS, NEEDS CORRECTION, REFUSED, bar list): the entry’s own miner may dispute them within 48 hours of publication.
- Window results (selection, score, twin, crown, places, reveal, and the lineage of any entry officially scored in that window): objections are due within 72 hours after the results are published (for results on a Thursday: Monday 16:00 Berlin time); this is the objection deadline. A reveal or objection deadline that falls on a Saturday or Sunday moves to Monday 16:00 Berlin time. If our MCP cannot take objections for more than an hour before an objection deadline, that
deadline moves by the length of the outage, rounded up to the full hour (at most one day; for a longer outage 5.3 applies), and we say so in #92 and on the page. Any miner may object to the lineage or bar list of another entry only here: in the window in which that entry is first officially scored as a selected entry, or in a later window in which it is scored again (3.9). A dispute against another miner’s entry is shown on that entry’s uid page and through the MCP, with the disputer’s hotkey, within 1 hour of filing; that miner has 48 hours from when it is shown to respond (for an objection filed at the deadline, when the results came out on a Thursday: until about Wednesday 16:00 Berlin time). One dispute may name several entries of the same window. At most 3 open disputes per hotkey at a time; a dispute about your own intake result does not count towards them. A dispute without evidence (10.3) is closed automatically. A dispute does not stop any payment unless 3.8 applies. List decisions and rule changes cannot be disputed.
- Bring: your uid, the repo@revision concerned, what you claim, and your evidence (files with revision and hash, configs, logs, or the output of our published check script).
- Decision: for an intake or lineage result we rerun the published check automatically; if the rerun gives PASS for a REFUSED or NEEDS CORRECTION result, the entry passes at once, without waiting for the slot. A dispute pauses the 24-hour correction clock. Every other dispute, and every rerun that does not settle it, is decided by me at the Thursday slot: a dispute that is complete (including the other miner’s response time, 10.2) by the end of the day (Berlin time) on which the response time for an objection filed at the objection deadline ends is decided at the first Thursday slot after that day, later ones the Thursday after. So objections to a window are usually decided at the next Thursday slot, before that Thursday’s window results are published; after a moved slot, or for a case under 3.8, at the first slot the timing allows. The decision is final and published on the uid page and, without personal data, in the FAQ. An upheld objection counts from the decision on: rewards already paid stay paid, and we do not replay a window (6.4).
- Closed windows. After its objection deadline a window is closed: its lineage results, ranking, crown and places are final, and later objections about it are not accepted. An entry scored again in a later window can be objected to in that later window’s period; a decision then counts from that later window on, never backwards. If we ourselves later find a deliberate misdeclaration (3.8) or a manipulation (section 11), the entry is held from then on and loses rewards not yet paid (3.8); closed windows stay as published and are never re-ranked. For our own errors, 10.1 still applies (corrected from the next window, the measured loss paid).
11. Manipulation of our agents
Published before automatic answers start.
12. Moving over from round 1
- Of round 1’s champions, uids 83, 114, 116, 199 and 250 count as former champions of round 2 for the keep-alive only: they keep it until 14 days after the round ends, like every former champion (6.6). That is a rule, not a fixed date: if the end of the round moves, the end of their keep-alive moves with it. They hold no podium place and get no new run-out share.
- Promises from round 1 that are still open (feedback reports, credits, refunds for errors on our side) are settled under round 1’s text.
- Round 2 replaced the round 1 Generalist rules on Friday 20 November 2026 at 16:40 UTC, when we announced it in #92. The v2.2 simplification applies to all current Round 2 commitments from publication under 3.9 and 9.3. From then on these round-1 sentences no longer apply to the Generalist track: the notice rules of FAQ 3.8 (24 hours before a change, the next slot on the current rules, the rollout procedure of 26 Oct, hotfixes with a fairness check), including for its share; “a change does not take away a right that was already published” and “nothing already earned is taken away” (replaced by 9.3); the 6-hour award decision; “we answer once an hour” (replaced by section 8); and “so it is never pruned” (replaced by 6.6 and 12.1).
Earlier versions: v2.8 · v2.7 · v2.6 · v2.5 · v2.4 · v2.3 · v2.2 · v2.1 · v2.0.
Every version and what changed in it: rules changelog.
Other pages: Miner FAQ · Competition rules (all tracks) · Live evaluation queue · Competitions.