Observe
Collect immutable real-data anchors, benchmark failures, and provenance-rich interaction traces. Generated samples are always labeled as generated.
Maple Brain Lab · Research Protocol
Recursive learning without recursive collapse
DOGMA and Hermon DNA may generate curricula, critiques, hard negatives, and bounded architecture proposals. They may not silently train on their own world model. Improvement is a lineage of candidates judged by independent, frozen evidence.
Why the guardrails matter
Self-generated data can erase the tails of reality
Recursive training is useful only when generated material remains attached to real observations and external tests. Research on model collapse shows that indiscriminate generational training can lose rare events and progressively distort the learned distribution. Maple therefore treats synthetic data as a proposal source, never as ground truth.
Closed Loop
Failure becomes curriculum; evidence decides promotion
Diagnose
Cluster errors by capability, calibration, invariance, contamination, and distribution shift. A cycle begins from measured failure, not elapsed time.
Propose
Generate curriculum changes, hard negatives, objective weights, or bounded architecture mutations. Every proposal declares a hypothesis and compute budget.
Train candidate
Train in isolation from production with versioned data manifests. Preserve real data in every cycle and cap synthetic-data contribution.
Challenge
Run frozen held-out suites, reverse-complement controls, rare-case retention, forgetting tests, calibration, and matched-compute baselines.
Archive
Retain a scientifically useful stepping stone when it is best in a declared behavior cell, even if it is not safe to deploy.
Promote or reject
Promote only on global objective improvement with every hard gate passing. Keep rollback and feed diagnosed failures into the next cycle.
Non-Negotiable Contract
A model can propose its next lesson. It cannot approve itself.
- A candidate cannot write its own passing grade.
- Frozen real-data anchors remain present across every generation.
- Synthetic examples carry generator, prompt, model, and filter provenance.
- Evaluation sets are isolated from curriculum generation and training.
- Rare and difficult cases are tracked explicitly to detect tail collapse.
- Architecture changes compete under matched tokens, parameters, and compute.
- Promotion requires reproducibility across seeds and a retained rollback checkpoint.
- Production weights never mutate online; learning produces a new candidate lineage.
Real-data sampling floor65%
A cycle may use synthetic proposals, but audited real anchors retain the majority of sampling mass.
Fixed probe seeds2
Every DNA, explanation, and self-critique probe must pass under seeds 17 and 29.
Variables changed per ordinary cycle1
Bounded single-variable mutation keeps failure analysis possible on limited compute.
Critical regressions allowed0
Archive membership preserves research diversity; deployment still requires every hard gate.
Promotion Gate
Six receipts for every accepted generation
Capability delta
The candidate improves the targeted frozen suite with confidence intervals or repeated-seed evidence.
No critical regression
Prior capabilities, rare classes, calibration, safety, and long-context behavior stay within preregistered bounds.
Data integrity
No train/eval overlap, unverifiable sample, hidden generator, or unlicensed source enters the accepted manifest.
Compute fairness
Architecture claims use matched tokens, parameters, wall-clock, memory, and accelerator budget where applicable.
Reproducibility
Another seed or worker reproduces the direction of improvement and produces a complete lineage record.
Rollback
The previous accepted checkpoint, data manifest, optimizer state, and serving route remain recoverable.
Different Recursions
DOGMA evolves architecture; Hermon DNA evolves representation and evidence
DOGMA
Bounded architecture evolution
Candidates may alter tetra scales, state dimension, regulation sparsity, mixer kernels, memory size, or objective weights inside a declared search space. Matched-compute ablations decide whether the mutation survives.
failure → hypothesis → bounded mutation → matched baseline → promote/rejectHermon DNA
Evidence and curriculum evolution
Candidates may mine uncertain examples, generate matched hard negatives, rebalance underperforming tasks, or add retrieved evidence. Frozen biological splits and calibration gates decide promotion.
error cluster → hard cases → encoder candidate → isolated suite → promote/rejectMethods
Self-training ideas and their limits
STaRNeurIPS 2022
Iterative bootstrapping from generated rationales, with correctness filtering.
Self-Play Fine-TuningICML 2024
Iterative self-play learning anchored to human demonstrations.
Model collapse under recursive dataNature 2024
Evidence that indiscriminate recursive training on generated data loses distribution tails and degrades models.
Population Based TrainingDeepMind 2017
A population trains while periodically exploiting stronger members and exploring bounded hyperparameter mutations.
MAP-ElitesArtificial Life 2015
Quality-diversity search retains the strongest candidate in each behavior niche instead of optimizing one scalar alone.
Darwin Gödel MachinearXiv 2025
A research system proposes self-modifications and keeps changes only when an empirical evaluator verifies improvement.
AlphaEvolveDeepMind 2025
An evolutionary coding system combines model proposals with automated evaluators; DOGMA borrows the evaluator-first discipline, not its scale claims.