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: conditioned state evolution through learned field operators. The iterations, interrupted runs, corrected baselines, and negative results stay in the record.
01Current line
cDFM treats generation as conditioned state evolution. A learned field passes through a wave stage. An aperture shapes it. Diffraction evolves it. Projection turns the resulting state into output. The displayed run is the Stage A promoted multiscale anchor.
Project
cDFM
Active
Scale topology was the tested axis. WAD, sharing, fusion, aperture, and the objective stayed fixed across the Stage A comparison.
86.3%Displayed-grid accuracy
0.023228Best validation x0 MSE
558,050Parameters
3,000Optimizer steps
192,000Images seen
10 of 10Class rows alive
02Live architecture search
The July runpack changes one architecture axis at a time. Every curated row gets 3000 steps after contract checks. Validation x0 MSE is logged as a reconstruction metric. Visual artifacts are reviewed separately.
A
Scale topology
Sequential 4, 16, 64 became the fixed multiscale anchor. Best validation x0 MSE: 0.023228. Final displayed-grid accuracy: 86.3%.
Complete
B
Operator sharing
Independent per scale: 0.024884. Shared across scales: 0.028341.
Complete
C
State fusion
Gated residual: 0.024870. Separate readout resumed after a user pause and completed at 0.026939.
Complete
D
Aperture role
Three of five rows are complete. Conditioned: 0.024860. State-conditioned: 0.024879. Static learned: 0.025959.
In progress
07 July 2026
The 300-step promotion gate was removed. Sequential 4, 16, 64 looked incomplete at 300 steps. Its regenerated final grid reached 86.3% frozen-classifier accuracy at step 3000. Step-300 outputs now stay in the learning record without deciding which full runs execute.
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
L-series
L0 marks the first experiment. L1 is a scaled iteration. L2 and onward are the next version, upgrade, or iteration.
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 L0-Luau-14B.
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.
January 2026
The 3-Parameter Ternary Reasoning idea emerged.
February 2026
L0-PolyCore-4B was released.
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 used learned wave, aperture, and diffraction operators on MNIST.
July 2026
The cDFM runpack now executes full curated rows with mandatory controls and artifact preservation. Stage D is in progress. TRR remains at protocol stage.
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