Engineering checksStandard-library fixtures
A final record is not a finished run
Nine isolated completion-guard tests passed. They validate the record checks, not a new training run or a working model.
Read the noteLab notes
The build, written down as it is measured.
Notes from a one-person research lab: measured baseline results, bounded engineering checks and openly labelled proposals. Mind-1B remains an architecture target; no 1B assistant has been trained here.
Engineering checksStandard-library fixtures
A quality signal still needs a cost review
The paired-report fixture accepts a narrow quality signal while explicitly refusing to claim efficiency or authorize training.
Read the noteEngineering checksStandard-library fixtures
Name the optimizer scope before comparing it
Eleven metadata-only tests check which parameters a proposed optimizer experiment would receive, without executing the optimizer.
Read the note
The ladder
S2, passed
The run that died at step 320
The clean-room core passed its first real-data gate, the surprise-memory ablation returned an honest null inside seed noise, and a deterministic NaN exposed a trained-decay overflow that would have killed every future run. Total bill: about $13.
Read the note →
Hardware
Calibration, passed
One hour on a B300, and the plan got a fifth cheaper
Measured 45.1% MFU and 135,600 tokens/s on a 1B-shape reference repriced the 300B-token campaign from 692 assumed to 538 measured GPU-hours. Plus 2.6B content-addressed training tokens and a clean-room recurrent core with a surprise-gated write ablation.
Read the note →
Sleep and memory
Gate M2, passed
Sleep, and the 81% that survived
In a synthetic continual-memory check, cleanup retained 81.4% mean sign recall after eight cycles, against 34.4% for a naive raw-magnitude retention control.
Read the note →
Surprise economy
Prototype
Seven percent of the FLOPs, and no error to show for the rest
On a generated redundant stream, the delta-matvec prototype used 7.7% of dense MACs with a 4.14e-6 relative RMS residual. This is a mechanism check, not a language-model benchmark.
Read the note →
Thesis
Founding note
Why we are not building a transformer
The transformer is a magnificent, uniform machine, and the uniformity is the tax. It runs every region at once, grows its memory without bound, and multiplies floats it could skip. A brain does none of those things. Here is the case for a different shape.
Read the note →Follow the run
One email when a gate passes or fails. Written by the lab, read by humans.