Reliquary

<p align="center">
  <a href="https://www.reliqua.ai">
    <img
      src="docs/assets/readme/reliquary-hero.svg"
      alt="Reliquary — verified frontier search for decentralized model training"
      width="100%"
    />
  </a>
</p>

<p align="center">
  <strong>Verified frontier search for decentralized language-model training.</strong><br />
  Sixteen trajectories per prompt. Validator-authoritative rewards. Public evidence at every boundary.
</p>

<p align="center">
  <a href="https://github.com/reliquadotai/reliquary/actions/workflows/validator-tests.yml">
    <img alt="Validator tests" src="https://github.com/reliquadotai/reliquary/actions/workflows/validator-tests.yml/badge.svg?branch=main" />
  </a>
  <a href="https://github.com/reliquadotai/reliquary/actions/workflows/cross-box-determinism.yml">
    <img alt="Grader determinism" src="https://github.com/reliquadotai/reliquary/actions/workflows/cross-box-determinism.yml/badge.svg?branch=main" />
  </a>
  <a href="https://github.com/reliquadotai/reliquary/actions/workflows/docker-image.yml">
    <img alt="Validator image" src="https://github.com/reliquadotai/reliquary/actions/workflows/docker-image.yml/badge.svg?branch=main" />
  </a>
  <a href="LICENSE.md">
    <img alt="MIT License" src="https://img.shields.io/badge/license-MIT-e87a3e" />
  </a>
</p>

<p align="center">
  <a href="https://www.reliqua.ai/dashboard">Live dashboard</a>
  ·
  <a href="https://www.reliqua.ai/research">Research</a>
  ·
  <a href="https://huggingface.co/Qwen/Qwen3-4B-Base">Base model</a>
  ·
  <a href="https://github.com/orgs/reliquadotai/packages/container/package/reliquary-validator">Container</a>
  ·
  <a href="docs/mining.md">Mine</a>
  ·
  <a href="docs/validating.md">Validate</a>
</p>

---

Reliquary is the open protocol and reference implementation for decentralized
GRPO training on Bittensor subnet 81. Independent miners spend their own GPU
compute searching for prompts at the policy's learning frontier. The trainer
admits, ranks, verifies, rewards, and—when the safety gates permit—trains on the
best groups.

The key incentive shift is simple: miners are not paid for producing the most
rollouts. They compete to contribute verified rollout groups the trainer can
use.

## V1.0 miner migration

Adapt your existing miner deployment; no separate starter package is required.
Stop the old miner before starting V1 and preserve your registered hotkey,
miner state directory and Hugging Face cache.

Use the pinned miner image (the controller has additional server-only fixes):

```sh
docker pull ghcr.io/reliquadotai/reliquary-validator:sha-3b3af1d-logic@sha256:7bb847499605308b1ad7b21f5430b1662eae506deacb6899d682bc0c11ce2742
```

Set these variables **inside your existing miner service or container**:

```dotenv
RELIQUARY_PROTOCOL_PROFILE=qwen3-4b-base-dapo-reliquary-v1
RELIQUARY_EXPERIMENTAL_FILL_CLOSED_ENABLED=1
RELIQUARY_TRAINING_RUN_ID=qwen3-4b-base-dapo-reliquary-v1-20260910
```

Keep your existing GPU selection, wallet mounts and persistent state/cache
paths. Inside that configured environment, adapt your existing miner command:

```sh
reliquary mine \
  --network finney --netuid 81 \
  --wallet-name YOUR_EXISTING_WALLET \
  --hotkey YOUR_REGISTERED_HOTKEY \
  --wallet-path YOUR_EXISTING_WALLET_DIRECTORY \
  --validator-url http://62.238.81.36:8000 \
  --environments openmathinstruct,opencodeinstruct,reliquary_logic_v2
```

Choose any nonempty subset of **Math** (`openmathinstruct`), **Code**
(`opencodeinstruct`) and **Logic** (`reliquary_logic_v2`). Running all three is
optional. Use one miner process per hotkey/state directory.

Before starting, follow the current operator notice and check readiness:

```sh
curl --fail --silent --show-error --max-time 10 http://62.238.81.36:8000/readyz
curl --fail --silent --show-error --max-time 10 http://62.238.81.36:8000/state
```

A failed readiness check or HTTP 503 means keep the miner paused. This README
is configuration guidance, not a live GO notice. The initial V1 checkpoint is
**N1812**, repository `ReliquaryForge/qwen3-4b-base-dapo-v4`, revision
`6c3f02be8e720d3ccfd1bd320de8ea7f864a04e6`. Follow the validator's advertised
successor checkpoints automatically; do not permanently pin N1812 or delete
checkpoint identity records to bypass a mismatch.

V1 uses **fill-closed windows**. Follow the deadlines and phase in `/state`:
windows may close underfilled, and an upload acknowledgement is not a final
proof, selection or payment verdict. No extra controller bounded-service flag
is needed on miners. See [Mining](docs/mining.md) for the submission lifecycle.

Logic uses a pinned integration with the **Prime Intellect Verifiers format
for supported environment contracts**. This supports further integrations;
it does not certify every Prime environment.

## Legacy auction profile loop

The fixed-window values below describe the legacy auction profiles. For V1
configuration and fill-closed timing, use the migration instructions above and
the validator's advertised runtime state.

<p align="center">
  <img
    src="docs/assets/readme/protocol-loop.svg"
    alt="Reliquary protocol loop: mine, grade, rank, prove, accumulate, train, and publish"
    width="100%"
  />
</p>

1. **Collect.** Math and Code each accept up to 96 productive candidates. The
   100-second profile value is a hard ceiling; a GPU-aware quiet/drain check may
   close from 60 seconds onward, after the primary 64-candidate population.
   Every group contains exactly 16 rollouts.
2. **Grade.** The validator recomputes Math rewards and executes Code cases in
   its sandbox. Groups below the reward-variance gate are rejected.
3. **Rank.** In-zone groups rank by
   `std(rewards) × (1 − mean(rewards))`; equal values use a capped throughput
   bucket, then a post-seal drand tie-break. Validator-observed arrival is used
   once, as the throughput denominator, not again as a second speed preference.
4. **Prove.** Economic proof runs top-down only for candida