Discovery Paroxysm.
An idea can start as a joke. It still has to survive a fair test. We build new ML architectures and run controls that could plausibly win. cDFM is the current main line. It keeps the field as state, shapes it with a learned aperture, and evolves it through physical diffraction. The iterations, interrupted runs, corrected baselines, and negative results stay in the record.
01Current line
cDFM means conditioned Diffraction Field Model. The intended architecture keeps a conditioned field as its primary state. A learned aperture shapes the field. Physical diffraction evolves it. The intensity of the field is the output; there is no learned decoder. One random field is carried through every scale. The architecture is now fixed. The MNIST stage is complete. CIFAR-10 is the current natural-image test.
Project
cDFM
Active
One random field is carried from the 4x4 source to the 32x32 output. At each scale a conditioned aperture shapes it and fixed diffraction evolves it. Nothing rebuilds the field between scales, and nothing decodes it at the end.
398,376Generator parameters
30,000Updates
32Batch size
02Current development
Each stage keeps its exact architecture and artifacts attached. Controls and seeds belong inside the run itself. Whatever survives decides the next stage.
01
Architecture definition
Persistent field as the primary state. Source field → learned aperture → physical diffraction → scale transition on the existing field → intensity readout. Fixed.
Done
02
MNIST evidence
Controls retrained from scratch at matched size and budget, over multiple seeds.
Done
03
CIFAR-10
The persistent architecture, natively at 32x32. In progress.
Active
04
Larger scale
Move to ImageNet or another larger-scale test when the evidence and available compute justify it.
Conditional
27 October 2026
The persistent architecture is fixed and the MNIST stage is complete. Current work is CIFAR-10.
03Past line
Atom Attention tested value-level atom slots beside local attention and exact-support cache paths. Pointer-backed v5-P became the strongest Atom variant under the monitoring protocol. The no-atom local+cache control made the mechanism result clear. The atoms added nothing. The line is closed. Its ablation record remains useful.
Project
Atom Attention
ClosedNegative result
Best validation. Lower is better. Mechanism attribution also depends on the no-atom controls.
| Variant | Best val | Read |
|---|---|---|
| GPT full | 1.2286 | Full attention reference |
| v5-P and v5c | 1.3518 | Pointer-backed atom variant |
| v5-P+ | 1.3659 | Stacked feature variant |
| No-atom local+cache | 1.4068 | Local and cache control |
| Linear attention | 2.2363 | Efficient attention reference |
04Open threads
(s, c, p) in {-1, 0, +1}
3-Parameter Ternary Reasoning proposes a small discrete control substrate for selection, commitment, contradiction, and revision. The Qwen sidecar reached experiments in May, but its causality and gradient-flow bugs were not fixed. There is no clean training result.
Prototype stalled
Targeted Rapid Refresh focuses a short high-learning-rate update on dense semantic reformulations of one narrow knowledge change. The protocol exists. The implementation and first experiment are still ahead.
Protocol
05Specialist model lines
Luminescence Series
LS-{name}-{n}-{size}[-preview]. LS is the Luminescence Series, LRML's main model series. Luminescence is low-temperature emission of light: small models emitting light. {n} is the generation and {size} the parameter count; -preview marks an early release. Example: LS-Inkprint-1-3B-preview. The earlier L0 releases keep their original names; every new model uses the LS format.
Inkprint
Writing style transfer. Given a few samples of someone's writing, it rewrites a text, or writes a new one from notes, so that it reads as if that person wrote it. It keeps the content and doesn't exaggerate their habits into a caricature. Our first serious public model.
In development
Main
The general model line. The current plan is to combine the strongest collected data without letting one specialty dominate the model's style.
Early plan
Luau
Roblox Luau specialist. The dataset has been scraped. Filtering and synthetic instruction generation come next for an internal 14B Luau model.
Data prep
PolyCore
Low-level C, C++, Rust, and assembly work. Clean C and C++ text data exists. Instruction data is the current gap.
Text dataset
Novelty
Conversation intelligence and coherence. The idea exists. Dataset work is still at zero.
Idea
Memex
English and Japanese information-accuracy Q&A. Wikimedia, OpenAlex, and OpenStax collection are the next steps.
Data prep
SagaForger
The storywriting umbrella. Specialist genre datasets feed the line. Tag-based MoE routing remains a later version target.
Planned
06Research chronology
December 2025
L0-Luau-1B and L0-Luau-4B were released. They were naively trained early trials, not serious work.
January 2026
The 3-Parameter Ternary Reasoning idea emerged.
February 2026
L0-PolyCore-4B was released. Like the Luau models, it was a naively trained early trial.
March 2026
Work started on value-level atom slots beside local attention and cache paths.
May 2026
The no-atom local+cache control showed that the atoms added nothing. The line was closed and its ablation record kept.
May 2026
The Qwen sidecar was tested, but its causality and gradient-flow bugs were not fixed. The experiments did not produce a clean result.
May 2026
The first WAD experiments combined learned field state, aperture, and spectral operators on MNIST.
September 2026
One random field carried through every scale, read out directly as intensity. The MNIST stage was completed.
October 2026
Work moved to natural images at native 32x32.
07About
LRML is run by Eino "Redd" Ota. J. "Brilliance" O. and "Quedol" contribute.
EO
Eino "Redd" Ota
Architecture research
JO
J. "Brilliance" O.
Contributor
Q
"Quedol"
Contributor