ModelBreeder.com
Project source notesCurrent release evidence2.5 KB

ModelBreeder.com Theory Source Alignment Audit

Release note comparing public theory with uploaded reports and runtime source files, recording promoted concepts and rejected literal interpretations.

Download original MarkdownSHA-256 fc451b4c3fe65f53be4f35640a284387cd3d6b36978e140170ff0a12527e3786
Raw source report

This page renders the original supplied document for reference. It has not been fact-checked line by line. Use the curated learning guides for normalized terminology, maturity labels, implementation boundaries, and safety framing.

ModelBreeder.com Theory Source Alignment Audit

This source note records how the public ModelBreeder.com theory was compared with the latest uploaded reports and runtime source files for the v2.3.0 release.

  • The central unit remains a governed ecology of small, mutable, replaceable models rather than a single expanding monolith.
  • The digital Four Fs are kept in the engineering sense of Feed, Fork, Fight, and Flee, with no-op as a valid result when change does not repay lifecycle cost.
  • Code breeding and model breeding remain separated. Code breeding changes the execution machinery; model breeding changes learned behavior, adapters, merges, quantization, specialists, and routing policies.
  • The TinyRustLM runtime source files ground the local-first browser story: .slm headers, tensor directories, byte/BPE tokenizers, q8/q4 quantization, adapter packages, eval sidecars, diagnostics, and raw WASM exports are now treated as concrete implementation evidence.
  • The site now includes interactive tools for scorecards, population simulation, router policy, lineage viewing, and release-packet drafting.

Reframed instead of promoted literally

The improvement directive argues for removing safeguards and safety halts from the core platform. The site does not implement that literally. It promotes the constructive part: a lower-friction research lab and a more analytical dashboard. The deployable site keeps support boundaries, evaluation independence, no-op, rollback, and human review. Those are not warning clutter; they are the control plane that makes a model ecology usable.

Added site work

  • /theory/source-alignment-audit
  • /architecture/tinyrustlm-runtime
  • /architecture/slm-model-package-abi
  • /evolution-lab/epigenetic-steering
  • /evolution-lab/evolutionary-merge-search
  • /tools/apex-scorecard
  • /tools/population-simulator
  • /tools/router-policy-lab
  • /tools/lineage-dag-viewer
  • /tools/release-packet-builder
  • new blueprint pages for legal, coding, browser runtime, and personal research ecologies

Ongoing theory gap

The public site is now stronger as a curriculum and executable design simulator. It is still not a production model-breeding runtime; it does not ship model weights, adapters, training data, or a production evaluator. The next technical milestone is a tiny local example package with sample .slm model metadata, mock adapter deltas, frozen eval cases, lineage JSON, and release packets.

Curated synthesis

Guides drawing from this report

Theory4 min

Theory source alignment audit

A comparison of ModelBreeder.com public theory against the latest source reports and runtime files, showing what was promoted, reframed, or held as source evidence.