> ## Documentation Index
> Fetch the complete documentation index at: https://docs.theminos.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Architecture

> How the Minos subnet operates — miners, validators, and the scoring loop

## Subnet Overview

Minos (SN107) is a **merit-based competition** for genomic variant calling on the Bittensor network. The system has two roles:

<CardGroup cols={2}>
  <Card title="Miners" icon="pickaxe">
    Share their hyperparameter configuration for running variant-calling tools (GATK, DeepVariant, FreeBayes, or BCFtools). Compete on **call accuracy**.
  </Card>

  <Card title="Validators" icon="shield-check">
    Distribute tasks, benchmark results against truth sets, and assign **scores that determine ALPHA rewards**.
  </Card>
</CardGroup>

***

## The Scoring Loop

<Steps>
  <Step title="Task Distribution">
    A validator selects a genomic region with a known truth set. It sends the miner aligned reads (BAM) and a reference genome for that region.
  </Step>

  <Step title="Variant Calling">
    The miner runs their chosen variant caller with their tuned hyperparameters. The output they share with the validator is their **hyperparameter config**.
  </Step>

  <Step title="Benchmarking">
    The validator uses the miner's config to generate a VCF, then compares it against the truth set using standard metrics:

    * **Precision** — what fraction of called variants are real
    * **Recall** — what fraction of real variants were found
    * **F1-score** — harmonic mean of precision and recall
  </Step>

  <Step title="Reward">
    Miners are ranked by accuracy. Higher accuracy earns more ALPHA. The system is **purely meritocratic** — only call quality matters.
  </Step>
</Steps>

***

## Why Hyperparameters Matter

Every variant caller has dozens of tunable parameters that control sensitivity, specificity, and how the algorithm processes data. The default settings are general-purpose — **they are not optimized for any specific benchmark.**

Miners who understand and tune these parameters gain a competitive edge:

| Lever                            | Effect                                                         |
| -------------------------------- | -------------------------------------------------------------- |
| Lower quality thresholds         | More sensitive — finds more variants but risks false positives |
| Higher confidence thresholds     | More precise — fewer false calls but may miss real variants    |
| Adjusted priors                  | Control the model's assumptions about variant frequency        |
| Variant-type-specific parameters | Tune sensitivity per variant class                             |
| Contamination / PCR models       | Model potential data artifacts                                 |

The [Miner Guides](/guides/gatk) document every parameter for each supported tool.

***

## Supported Tools at a Glance

| Tool            | Approach                                             | Docker Image                          |
| --------------- | ---------------------------------------------------- | ------------------------------------- |
| **GATK**        | Local de novo assembly + Pair-HMM genotyping         | `broadinstitute/gatk:4.5.0.0`         |
| **DeepVariant** | CNN image classification on pileup tensors           | `google/deepvariant:1.5.0`            |
| **FreeBayes**   | Bayesian haplotype evaluation with toggleable priors | `staphb/freebayes:1.3.7`              |
| **BCFtools**    | Pileup-based genotype likelihoods + Bayesian caller  | `quay.io/biocontainers/bcftools:1.20` |