Proof-carrying frontier market
100-second windows · difficulty auction · GRAIL verification
This is not a calendar of promises. It is the sequence of proof required to move from a live frontier market, through replayable public evidence and behavioral-diversity research, to portable RL infrastructure.
The roadmap moves from the live Subnet 81 training substrate, to replayable public proof, to evidence-gated behavioral-diversity research, and finally to a proposed open training market.
100s
qualifying window
production cadence
8×
rollouts per group
fixed reward geometry
4
successful steps
default publish cadence
0%
behavioral-diversity influence
observation only today
evidence language
Live means running now. Measured means a bounded receipt exists. Building means a versioned artifact is in progress. Research means no production influence. Direction means no committed date.
Subnet 81 already runs the qualifying circuit: miners search the learning frontier, generate 8 rollouts per group, GRAIL verifies the work, and healthy balanced batches can move Qwen3.5-2B. Selection is eligibility—not proof that an optimizer step happened.
100-second windows · difficulty auction · GRAIL verification
ReliquaryForge/qwen3.5-2b-reliquary-v3 · every four successful balanced optimizer steps, with earlier publication on behavior-policy drift
300 steps · Qwen3-4B-Instruct · one controlled run
Qwen3-4B-Instruct · 300 steps. A receipt to repeat, not a production guarantee.
gate to advance
inspect the receipts
A public systems walkthrough should let anyone follow one window from generation through an explicit train-or-hold decision. Reliquary is preparing that evidence path for a possible future community showcase; no invitation, episode, schedule, or endorsement is claimed.
Dashboard, validator state, public archive, model, and source remain independently inspectable.
16 selected groups · 128 trajectories · 0 of 16 accumulated · checkpoint ceiling held.
Redacted receipt, current-data preflight, failure fallback, terminology boundary, and exact claim ledger.
explicit train-or-hold path
Selection completed. Training remained unattempted at checkpoint 53 because the checkpoint ceiling held. The hold is the receipt—not missing progress.
Historical aggregate receipt. Identities and task payloads are intentionally absent from this presentation surface; use the linked source for provenance.
gate to advance
inspect the receipts
Correctness comes first. Canonical content identity and private utility observation are deployed, but behavioral novelty has zero influence on rank, selection, rewards, training, or miner payloads. It earns influence only through a preregistered comparison against the production baseline. This research is separate from Bittensor’s Novelty Search community call.
Difficulty auction, GRAIL proof, content identity, and alias cooldown.
Private utility telemetry writes; the production contract remains difficulty-only.
Observable behavior distance, isolated archives, and fixed-budget evaluation remain hypotheses.
Gates 04–06 are research. They cannot alter rank, rewards, selection, or training before a preregistered comparison survives.
gate to advance
inspect the receipts
After the canonical loop and its research gates survive repetition, open the market to outside models and environments. A client should receive portable, verified training signal without trusting a hidden rollout service. Multi-tenant economics come last—not before isolation, replay, and abuse resistance.
Client model + environments in; verified rollout bundle + provenance out.
Sandboxed graders, resource caps, no network, and no collision with the canonical run.
Portable proofs, cross-run reputation, and multi-tenant economics only after the safety model holds.
gate to advance
inspect the receipts
operating rule
Every claim needs an object another team can inspect: source, proof, replay, checkpoint, or evaluation. The roadmap moves when the evidence does.
inspect the research record →