Epago

<div align="center">

<img src="assets/github-header.png" alt="EPAGO — the frontier open deep-research model, forged by competition" width="100%">

### The world's frontier open deep-research model, forged by adversarial competition

*A Bittensor subnet where the best model wears the crown — provably, replayably, with no one to trust. Across all of science.*

![tests](https://img.shields.io/badge/tests-452%20passing-2a9d3a?style=flat-square)
![python](https://img.shields.io/badge/python-3.11%2B-4584b6?style=flat-square)
![license](https://img.shields.io/badge/license-MIT-8a6a14?style=flat-square)
![status](https://img.shields.io/badge/status-testnet%20ready-d8a63a?style=flat-square)

[**Whitepaper**](docs/WHITEPAPER.md) · [**Handbook**](docs/subnet.md) · [**Mechanism spec**](docs/DESIGN.md) · [**Validate**](docs/VALIDATING.md) · [**Mine**](docs/MINING.md) · [**Rounds**](docs/ROUNDS.md) · [**Dashboard**](docs/DASHBOARD.md) · [**FAQ**](docs/FAQ.md)

</div>

---

## What we're building

Epago is building the **world's frontier open deep-research model** — small enough to
run and fine-tune on modest hardware, provably better every round. **Across all of
scientific literature.**

The bet rests on one observation: **deep research is a procedure, not a knowledge
store.** The facts live in the documents being read, not in the weights. What decides
quality is the procedure — decompose the question, search, open sources, read,
cross-check conflicting studies, attribute every claim. Procedures distill into small
models; memorized world-knowledge does not. In grounded research, the recall a large
parameter count buys is supplied by the corpus instead, so parameters stop being the
deciding variable.

That is demonstrated, not hypothesized. The genesis base model,
**Tongyi-DeepResearch-30B-A3B**, released by Alibaba under Apache-2.0 — a **30.5B mixture-of-experts with ~3.3B
active parameters per token**, small enough to fine-tune and serve on a single consumer
GPU — outperforms OpenAI o3 and DeepSeek-V3.1 (671B) on **5 of 7** agentic
deep-research benchmarks in its authors' published results: HLE 32.9 (o3 24.9,
DS-V3.1 29.8) · FRAMES 90.6 (84.0 / 83.7) · xbench-DeepSearch 75.0 (67.0 / 71.0) ·
WebWalkerQA 72.2 · GAIA 70.9 · BrowseComp-ZH 46.7. **o3 still leads BrowseComp, 49.7
vs 43.4.** Those are the *base model's* published numbers — the floor Epago starts
from, not a result Epago has earned; nothing here is Epago's own until a challenger is
crowned. And the claim is narrow by construction: better **at deep research, per unit
of compute** — never general intelligence.

So the open frontier of deep research is already small — and **static**. A lab ships a
checkpoint and moves on. Nothing makes it keep improving, and nothing proves the next
version is genuinely better rather than benchmark-tuned. **Epago is that mechanism.**
There is no lab, no training team, and no owner API. Anyone on earth can fine-tune the
reigning champion — *the king* — and submit a challenger. Independent validators run
statistical **duels** between challenger and king on research tasks neither has ever
seen. Beat the king with 99.9% confidence across a stake-weighted quorum of validators,
and the crown — and its emissions — are yours until someone does the same to you. The
model structurally cannot stop improving, the scores structurally cannot be faked, and
every verdict is replayable by anyone from public data — the protocol exactly, the
model scores to within a measured noise floor the mechanism was designed around.

The headroom is real and we publish it: on Epago's own deliberately hard exam the base
model measures **33.9%** with full agentic tooling; handing it perfect retrieval lifts
it only to **56.1%**, and taking the corpus away collapses it to the **17.3%** guessing
floor. The exam is validated, not assumed — the same harness change that raised it by
24 points also raised an external deep-research benchmark, and a same-size model
without research training ranks below the base on both. That 34-to-56 gap is not a
footnote — it is exactly what the competition exists to close, and episode completion
(most lost points die at the turn, clock, or context budget mid-research) is the first
obvious target for any miner.

**The scope is all of scientific literature.** The task templates are field-neutral by
construction — blank a reported value, describe a study instead of naming it, compare a
quantity across papers — so what a task needs is a claim that traces to a source and an
answer that is mechanically checkable, not a clinical trial. The current chain
generation, `EPAGO-DR-30B`, pins a **50,420-paper corpus spanning ~135 fields in all four
OpenAlex domains** (Life 1,450 · Physical 1,449 · Health 1,449 · Social 1,444) and reads
it through the `SCI4` task release — hard to find, easy to check: constrained search,
cross-study comparison, and computed evidence whose answers exist verbatim in no document. The work is worth
attacking everywhere it happens: a fabricated number is an integrity failure in any
field and a patient-safety event in some, evidence synthesis is slow and expensive
(published estimates put systematic reviews alone near 29,000 a year at roughly 1.72
scientist-years each), and many of the buyers who need it most are regulated parties who
cannot send documents to a hosted API at all — the structural opening for an open,
self-hostable model.

## Architecture

```mermaid
flowchart LR
    subgraph MINERS["⚒ Miners"]
        MM["Miners<br/>train challengers"]
    end

    subgraph CHAIN["⛓ Bittensor chain"]
        RC["Timelock commit-reveal<br/>e2 challenges · er1 rounds<br/>ev3 verdicts · ep1 pools · ek1 king"]
        W["Weights → emissions<br/>(Yuma consensus)"]
    end

    subgraph VAL["🛡 Every validator (identical box)"]
        TG["Task generator<br/>seeded by block hash"]
        PP["Private task pool<br/>(published at rotation)"]
        DUEL["⚔ Paired duel engine<br/>king vs challenger · GPU"]