Theory Advanced 3 minute read Updated 2026-06-28 UTC

Theory-source alignment matrix

A comparison of ModelBreeder.com theory against the supplied reports and Rust runtime files, with accepted, reframed, and rejected implementation decisions.

Research statusSource comparison and implementation decision record Publication statePublished Reviewed byMichael Kappel Source reports7

Direct answer

The source files broadly support the public ModelBreeder theory: adaptive AI should be built as small, replaceable, resource-bounded model ecologies rather than one monolithic parameter store. The strongest overlap is the Feed → Fork → Fight → Flee loop, the separation of code breeding from model breeding, the use of independent evaluators, the value of local tiny-model runtimes, and the need to preserve lineages, scorecards, and retirement decisions.

The source files also contain speculative and incompatible directives. In particular, instructions to remove kill-switches, safety boundaries, warnings, and governance are not accepted as operating policy for ModelBreeder.com. They are preserved in /docs as source evidence, but the public site implements the constructive research pattern: broader exploration inside a bounded lab, not unrestricted release.

Alignment matrix

Source themeSite statusImplementation decision
Small model ecologies over monolithsAcceptedThe homepage, Apex Multi Model, and architecture pages now frame the unit of optimization as a population.
Digital Four Fs: Feed, Fork, Fight, FleeAcceptedFoundation, theory, and lab pages use the loop as the core operating cycle.
No-op when gain does not exceed costAcceptedViability scoring, simulator tools, and release pages treat no-op as a protected outcome.
Code breeding vs model breedingAcceptedThe site keeps machinery evolution separate from parameter evolution.
Epigenetic steering vectorReframedAdded as a speculative parameter-efficient research pattern, not as a production claim.
Stochastic mating kernelReframedAdded as an experimental search operator with compatibility, budget, lineage, and evaluator boundaries.
Evolutionary model mergingAccepted with conditionsAdded merge-lab guidance, emphasizing compatibility, held-out evaluation, and fallback to distillation when schemas differ.
Proposer / Solver / Judge caste evolutionAccepted as lab patternAdded cooperative-caste page, while keeping judges independent from candidates.
Remove safeguards and kill-switchesRejected as site policyReplaced by bounded research-sandbox guidance and explicit support boundaries.
Rust/WASM local runtimeAccepted as implementation substrate evidenceAdded TinyRustLM bridge pages and raw source preservation under /docs/tinyrustlm-source/.

What changed because of this comparison

The comparison created four practical upgrades.

First, the site now contains implementation tools rather than only conceptual advice: an Apex scorecard, a population simulator, a lineage viewer, a router-policy lab, and a release-packet builder. These are browser-local teaching tools, not production evaluators.

Second, the site now surfaces source maturity more clearly. A source can be established practice, emerging practice, conceptual synthesis, speculative analogy, implementation note, or rejected directive. This protects credibility and helps readers understand which claims can be applied immediately.

Third, the Rust runtime files are now translated into the site architecture. The .slm model format, adapter-delta package checks, deterministic generation, quantized tensor paths, diagnostics, eval sidecars, and WASM ABI become concrete design anchors for a future local lab.

Fourth, the .uai memory now records this source-comparison decision. Future agents should not treat every uploaded report as direct operating truth. A report becomes current truth only when promoted into content, code, tests, release notes, or typed .uai memory.

Accepted synthesis

pseudocode
PROCEDURE compare_source_to_site(source_claim, current_site_contract)
    IF source_claim.supports_small_model_ecology
        PROMOTE_TO_CONTENT(source_claim)
    ELSE IF source_claim.is_speculative_but_useful
        LABEL_AS_RESEARCH_PATTERN(source_claim)
        ADD_BOUNDARIES_AND_EVIDENCE_NEEDS(source_claim)
    ELSE IF source_claim.conflicts_with_plain_php_or_governed_release
        PRESERVE_IN_DOCS(source_claim)
        RECORD_REJECTION_REASON(source_claim)
    END IF
END PROCEDURE

Current truth boundary

The public site is the curated teaching layer. /docs is long-term source memory. .uai is active project memory. Source files may inspire pages and tools, but source text does not override the package contract: plain PHP, no database, root-extractable, no WordPress, no machine-only manipulation text, no unbounded runtime autonomy, and no production promotion without independent evidence.

Source reports used for this guide

These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.