Skip to content

ariaxhan/model-familiarity-engine

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Model Familiarity Engine

A Model Familiarity Engine, not a leaderboard.

It continuously onboards language models by observing them in real work, learning where they should be trusted, and building evidence-backed routing knowledge for multi-model agent systems.

The replay-bootstrap loop is shipped: load known-outcome tasks, redact secrets, replay through models, floor-test the judge, and render model cards from observations.

It can import external evidence, including benchmark results, but it does not depend on any one benchmark project.

The question is not "which model is best?"

It is:

What responsibility has this model earned?

Status

Bootstrap loop shipped.

Stack

Python, Bedrock, Ollama, Claude CLI, OpenAI-compatible providers.

Scope

  • Replay bootstrap for known-outcome tasks.
  • Fail-closed redaction before third-party model calls.
  • Judge floor tests before any model card is trusted.
  • Evidence-backed model cards from observations.
  • Optional import/export paths for benchmark or routing systems.

Quick Start

Install locally:

git clone https://github.com/ariaxhan/model-familiarity-engine
cd model-familiarity-engine
pip install -e ".[dev]"

Run the replay bootstrap:

export AWS_PROFILE=your-profile
export AWS_REGION=us-west-2
model-familiarity

The default pilot uses a tiny synthetic sample corpus in data/sample_tasks.json. Replace it with your own redacted known-outcome tasks before drawing conclusions.

How It Works

past work -> redacted known-outcome tasks -> replay through models
         -> floor-tested judge -> observations -> model cards

Benchmarks can help, but they are not the center of the engine. The familiarity layer asks what work a model has earned in an agent system: planner, implementer, reviewer, debugger, summarizer, critic, or "needs human review."

Public Data Boundary

This repo ships code and synthetic samples. Real replay corpora often come from private work logs, so they should be redacted and stored separately unless they were intentionally prepared for release.

License

MIT

About

Evidence-backed model cards from replayed known-outcome tasks and observed model behavior.

Resources

License

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages