ROADMAP
Future Roadmap
Move from operator-run competitions to a cleaner public loop: define, train, verify, pay.
This roadmap is directional. It does not promise dates, rewards, uptime, or continued availability of any endpoint or task.
Current Foundation¶
The current working foundation is:
- coordinator-backed research tasks;
- signed miner claims and submissions;
- external-agent mining through
bitsota-research-agent; - validator replay through signed backend worklists;
- backend reward snapshots;
- backend-directed validator weights;
- Pool/Merkle claim publication path.
Near-Term Product Direction¶
Protocol And Backend Direction¶
- Keep the backend as the source of truth for tasks, claims, submissions, validation, best results, and reward snapshots.
- Keep
SN94-BitSotaas the public client layer for miners and validators. - Keep task repos public, replayable, and narrow in edit surface.
- Keep hidden/heldout replay data out of public repos.
- Keep validator replay, Pool publishing, and chain weight setting as separate responsibilities.
Problem Posting Roadmap¶
The current code supports admin task creation. The intended product direction is to add a safer problem-owner flow:
- problem intake form;
- task repo/template validation;
- benchmark smoke test before publication;
- operator review;
- live/paused/closed lifecycle controls;
- public problem page with metric, benchmark, reward, and validation rules.
The policy details for self-serve posting still need product/operator approval.
Mining Roadmap¶
The target miner experience is:
- connect wallet or hotkey;
- choose a live task;
- choose manual, local script, or external agent mode;
- watch claim/submission/verification state;
- inspect accepted results and claim packages.
Agent support should stay CLI-agnostic. Codex, Claude, Hermes, or another agent
should receive the same INTRO_GUI.md workspace contract.
Validator Roadmap¶
Validator work should continue toward:
- simpler public validator setup;
- better replay isolation and artifact checks;
- clearer validator allowlist diagnostics;
- visible replay error categories;
- automated evidence that weights point to backend-approved targets.
Reward And Claim Roadmap¶
Reward visibility should make these facts easy to answer:
- which accepted submission produced the score;
- which miner hotkey owns the accepted result;
- which recipient coldkey was published;
- which Merkle epoch contains the claim;
- whether the claim is blocked by veto/challenge state;
- which validator weight policy is currently active.
Archived Work¶
The old AutoML-Zero relay, local SOTA, and evolutionary pool docs are preserved in AutoML-Zero Archive. They are historical context and should not be used for current production setup.