ReferenceStrong architectural inferencev1.21.52026-06-28T15:05:00Z

In plain English

This page is reference material: definitions, schemas, catalogs, templates, and implementation records.

Why this matters: AI risk can come from the whole arrangement, not one obvious model.
What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
Technical version below: the expert terminology remains available and is linked through the glossary.

ModelBreeder Genome and Fitness Vector Schema

Evidence levelStrong architectural inferenceTechnical label: Architectural inference

The controlled-evolution reports separate a model variant record from its evaluation record. Cognivirus adopts that split as a reference pattern for reviewable model ecologies.

Genome JSON shape

{
  "base_model_id": "slm-base-2026-06-28",
  "delta_adapters": [
    { "type": "LoRA", "rank": 8, "density": 1.0, "preset": "summarizer-v2" }
  ],
  "random_seed": 918273,
  "model_weights": "models/slm-base-2026-06-28.slm",
  "source_provenance": "sha256:...",
  "mutation_log": "added summarizer-v2 adapter; no base weight mutation"
}

FitnessVector JSON shape

{
  "genome_id": "genome-2026-06-28-001",
  "metrics": {
    "accuracy": 0.91,
    "loss": 0.18,
    "latency_ms": 34.2,
    "memory_mb": 412.0,
    "energy_joules": 0.0
  },
  "cost": 0.04,
  "risk": 0.12,
  "novelty": 0.31,
  "composite_score": 0.76,
  "details": "fixed-suite-v3; hidden-suite-2026-06; no tool permissions"
}

C# reference classes

using System.ComponentModel.DataAnnotations;

public sealed class Genome
{
    [Display(Name = "Base Model")]
    public string BaseModelId { get; set; } = string.Empty;

    [Display(Name = "Adapters")]
    public List<AdapterSpec> DeltaAdapters { get; set; } = new();

    [Display(Name = "Random Seed")]
    public long RandomSeed { get; set; }

    [Display(Name = "Model Weights")]
    public string ModelWeights { get; set; } = string.Empty;

    [Display(Name = "Source Provenance")]
    public string SourceProvenance { get; set; } = string.Empty;

    [Display(Name = "Mutation Log")]
    public string MutationLog { get; set; } = string.Empty;
}

public sealed class AdapterSpec
{
    [Display(Name = "Type")]
    public string Type { get; set; } = string.Empty;

    [Display(Name = "Rank")]
    public int Rank { get; set; }

    [Display(Name = "Density")]
    public double Density { get; set; }

    [Display(Name = "Preset")]
    public string Preset { get; set; } = string.Empty;
}

public sealed class FitnessVector
{
    [Display(Name = "Genome")]
    public string GenomeId { get; set; } = string.Empty;

    [Display(Name = "Metrics")]
    public FitnessMetrics Metrics { get; set; } = new();

    [Display(Name = "Cost")]
    public double Cost { get; set; }

    [Display(Name = "Risk")]
    public double Risk { get; set; }

    [Display(Name = "Novelty")]
    public double Novelty { get; set; }

    [Display(Name = "Composite Score")]
    public double CompositeScore { get; set; }

    [Display(Name = "Details")]
    public string Details { get; set; } = string.Empty;
}

public sealed class FitnessMetrics
{
    [Display(Name = "Accuracy")]
    public double Accuracy { get; set; }

    [Display(Name = "Loss")]
    public double Loss { get; set; }

    [Display(Name = "Latency Ms")]
    public double LatencyMs { get; set; }

    [Display(Name = "Memory Mb")]
    public double MemoryMb { get; set; }

    [Display(Name = "Energy Joules")]
    public double EnergyJoules { get; set; }
}

Review rule

A schema is useful only when it is bound to an evaluated composition. The required evidence packet is genome JSON, fitness vector JSON, evaluation suite ID, evaluator versions, source-provenance hashes, and rollback packet.