engrammatic.

Engrammatic / Research notebook

LLM research.

Different arithmetic.
Same intelligence?

Experiments in how language models compute, retain information, and compose. We track what works, what breaks, and what the evidence can actually support.

Explore the experiments

The research log

Experiments

03 published records
Experiment 01All-64 held-out evaluation passed
Operator compilationSVD-guided learningLayer composition

Can a learned HDC mesh carry an entire language model?

Recompile Qwen’s transformer operators into population arithmetic, learn SVD-guided corrections, then test whether the replacements compose across all 64 blocks.

Read the experiment

Evidence snapshot ·

Text blocks replaced
64 / 64

Both held-out chains completed; no native transformer tail

Known-test agreement
95.90%

491 / 512 teacher choices · mean KL 0.00594

Fresh-test agreement
97.66%

125 / 128 choices · both predeclared gates pass

Experiment 02Fresh-test KL improves; quality target unmet
Endpoint predictionJEPA-inspired trainingFour-block fusion

Can one small mesh predict the endpoint of four Qwen blocks?

A JEPA-inspired distillation experiment: remove the inherited four-block base, learn a causal endpoint predictor, and measure compiled size against final token agreement.

Read the experiment

Evidence snapshot ·

Fresh-test output KL
4.62% lower

Versus starting mesh; 64 new passages; 60 native blocks remain

Larger compiled mesh payload
≈36.15 MB

No inherited teacher projection base; quality fails

Cloud spend / cap
$0 / $25

Local pilot; no paid instance launched

Experiment 03Capacity bottleneck measured; no new held-out result
Selective fusionCausal meshNetwork-wide scan

Which neighboring Qwen blocks are worth fusing?

Network-wide pair screening, frozen integer composition, then a diagnosis of output-subspace capacity, context features and regularized prediction.

Read the experiment

Evidence snapshot ·

Joint held-out teacher agreement
88.09%

Prior frozen joint integer test;60 blocks remain native

Error outside original output bases
82–89%

Squared endpoint error; post-hoc diagnosis, not token error

Cloud spend / cap
$0 / $25

Local diagnostics, fitting and native-suffix validation

How we work

A result is more than a number.

  1. 01

    State the hypothesis

    Define what changes, what stays native, and the result that would make the change useful.

  2. 02

    Freeze the test

    Bind the model, data, and implementation to an explicit protocol before held-out evaluation.

  3. 03

    Keep the boundaries

    Separate training, validation, and held-out evidence. Report failed gates alongside successful measurements.