skip to content
Reliquary

research · the published work

The protocol, in writing.

Reliquary publishes its work openly — research articles on the learning frontier, the protocol paper, and conceptual notes. The same documents drive the codebase and the chain.

01 · articles

Research, in long form.

Articles on how the network works and why it wins — the mechanism, the experiments, the results. Written for the site, not synced.

  1. technical report · september 2026

    2026-09-21

    A Market Mechanism for Rollout Selection in Group-Relative RL

    A 4B model trained for 12,852 updates on rollouts supplied entirely by independent operators: +31.1pp math pass@1 and +20.1pp code over Qwen3-4B-Base on exact-content-disjoint subsets, with transfer to AMC23, AIME, HumanEval+ and MBPP+. Single run, not replicated, and no matched centralized control — the report states what it could not measure.

    read the article →

  2. field note · behavioral diversity

    2026-07-23

    Eight rollouts are not eight votes

    A production window, the complete 0–8 frontier, 1,104 production trajectories in a frozen ten-window sample, and a brick-by-brick protocol for testing behavioral diversity against causal training utility.

    84 / 138

    groups at 2 / 8 reward-positive

    read the article →

  3. article · the market

    2026-06-16

    Reliquary: a market for the learning frontier

    How Subnet 81 turns prompt selection into competition among operators: miners bet compute on prompts at the policy's learning frontier, replacing DAPO's reactive filter with an ex-ante commitment — +14pp pass@1 over vanilla GRPO at identical step count (Qwen3-4B-Instruct, 300 steps, held-out math, 0.47 → 0.61, single run, not replicated).

    pass@1: 0.33 → 0.47 → 0.610.330.470.61+14pppass@1vs GRPO
    read the article →

  4. field note · the 2B run

    2026-08-05

    Reliquary 2B: training done

    Two environments, one system, two outcomes. Code gained accuracy while getting shorter; math learned to terminate. What the constraint decided, and what a post-trained checkpoint cannot tell us.

    read the article →

  5. update · incentive mechanism

    2026-09-14

    Paying for what we can actually consume

    Emission moves from a fixed amount distributed by rank to a price per unit of admissible data, and what the trainer cannot consume burns. The sensor, the controller, the task registry, the payment floor, and what we will measure.

    read the article →

02 · the protocol paper

The protocol paper.

draft · v0.1

committed

2026-05-06

source sha

78a43a2e…642b3566

path

reliquary-ledger/docs/paper/reliquary_protocol_paper.md

Reliquary: Proof-Carrying Inference for Decentralized AI Subnets

The protocol paper behind RL training on Subnet 81. Miners submit a compact sketch proof that validators re-execute in milliseconds to confirm the inference was computed with the declared model on the declared inputs. A nine-stage verifier pipeline checks schema, tokens, prompt construction, proof integrity, termination, environment, reward, log-probabilities, and output distribution. A mesh aggregates per-completion verdicts into stake-weighted consensus with outlier gating.

read the full paper →source on github

03 · research notes

Working notes, not yet preprints.

Living documents synced from the sibling repos. Each one is the canonical reference for one slice of the protocol; together they form the conceptual ground the paper stands on.

  1. note · concepts

    Why Reliquary, and what it defends against.

    The thesis: at this scale, which prompts you train on is the single largest published lever. Reliquary's design makes that lever cryptographically auditable.

    read →

  2. note · architecture

    The Forge: training architecture as a public surface.

    How verified selected groups become training-eligible, and how train-or-hold, optimizer, checkpoint, and evaluation evidence remain distinct.

    read →

  3. note · validating

    Validating: the nine-stage verifier.

    Schema → tokens → prompt → proof integrity → termination → environment → reward → log-probabilities → output distribution. Each stage names a specific class of attack it rules out.

    read →

  4. note · mining

    Mining: the deterministic-inference envelope.

    What it takes to produce a rollout the validator will accept. Determinism budget, sketch construction, the GRAIL commitment, and the failure modes (and what each one rejects).

    read →

  5. note · scoring

    Scoring: from verdicts to weight.

    How the validator's per-completion accept/reject decisions are aggregated into a stake-weighted median consensus, with explicit outlier detection and incentive shaping.

    read →

04 · citation

Cite the work.

The protocol paper is a living document; pin the source SHA you reviewed. BibTeX:

@misc{reliquary2026,
  title  = {Reliquary: Proof-Carrying Inference for Decentralized AI Subnets},
  author = {Reliquary Contributors},
  year   = 2026,
  note   = {Version v0.1, commit 78a43a2e},
  url    = {https://reliqua.ai/paper}
}