Question 1

How do miners earn in Reliquary?

Accepted Answer

Miners earn by submitting verified rollout groups that the validator selects. Each selected group occupies one fixed slot (regardless of completion length or throughput). Per-miner emission is calculated as an EMA score updated after each window and submitted on-chain once per epoch. A miner's share of the window pool depends on how many groups they contribute relative to other miners and the number of active environments and picks per window.

Question 2

What is the zone filter and why does it matter?

Accepted Answer

The zone filter requires σ ≥ 0.24 (the population standard deviation of the 16 rollout rewards in a group). Groups with zero variance (all rewards identical) carry no gradient signal under GRPO and are rejected. For binary math rewards, this admits every non-degenerate outcome (1–15 correct out of 16) while still rejecting trivial cases. This ex-ante prediction of high-variance groups is the core miner incentive: instead of generating broadly and filtering after, miners predict the frontier and generate only there.

Question 3

How does GRAIL verification prevent fabricated completions?

Accepted Answer

GRAIL sketches are compact linear commitments over sampled hidden-state activations from the model's forward pass for each completion. The validator recomputes the same forward pass with its own weights, draws the same random activation positions (seeded from the window's drand-derived randomness), and checks that both sketches agree within a calibrated numerical tolerance. Fabricated or copied rollouts diverge far beyond honest numerical drift, and sketches are bound to the exact token sequence and model revision, making cross-version attacks infeasible.

Question 4

What happens to prompts after they are selected?

Accepted Answer

Selected prompts enter a one-shot cooldown for 1,000,000 windows (effectively permanent within any realistic training run). The cooldown is tracked by both dataset index and canonical rendered-prompt digest, so duplicate content at another index cannot win again. This forces curriculum diversity and prevents the system from clustering on easy prompts as the policy improves.

Question 5

How is training data provenance verified?

Accepted Answer

Every rollout carries a signed GRAIL sketch commitment that binds it to the model checkpoint that generated it. The validator re-verifies the sketch with its own forward pass before any group can enter training. Admission also validates that the submitted checkpoint hash matches the current published revision, preventing stale models from contributing. The final training archive records selected groups with their exact rewards, GRAIL verification status, and window lineage.

Question 6

What is the difference between protocol v5 and v6?

Accepted Answer

Protocol v5 introduced reasoning-cue prompts (step-by-step instruction for math and code) and BFT masking (hiding forced-answer spans from gradient computation while keeping them in reward scoring). Protocol v6 replaces the timed auction window (100-second ceiling with adaptive early close) with a fill-closed stream: windows close when the configured number of picks are proven and sealed, making proof dispatch continuous and selection FIFO rather than deferred to a seal-time auction.

Reliquary

Reliquary

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup