TRANSFORMS / rowH8
Borrow an orthonormal signal transform to redistribute concentrated energy before block quantization.
UNDER ACTIVE DEVELOPMENTCROSS-DOMAIN ATOMS / RUST / v0.2.3
WEIGHT CODEC RESEARCH / AQ-001ATOM
ATOM
QUANTIZER.
The fascinating part is not the bit width. It is the collision: signal transforms, audio-style repair, information theory, vector geometry, and integrity mechanisms become bounded atoms inside one adaptive codec.
ENTER THE CROSS-DOMAIN COLLIDER ↓DUAL
GATEAQ
Q2Q4Q8GATEAQ
CROSS-DOMAIN COMPOSITIONDSP · KLGEOMETRY · VERIFY
STATUSRESEARCH SOFTWARE
CURRENT CONTAINEROQ02 · EXPERIMENTAL
LATEST SAFE SNAPSHOTDUAL-GATE v0.2.2
NEXT PROOFFULL v0.2.3 MODEL RUN
THE ARCHITECTURAL HOOKDomains that were
Domains that were
never supposed to meet.
Click an origin field to see exactly what Atom Quantizer borrows, how it becomes a bounded atom, where it enters the codec, and whether it is live or still experimental.
CROSS-DOMAIN TRANSFER / SOURCE-BACKEDLIVE PATH
ORIGIN DOMAINFAST TRANSFORMS
BORROW THE INVARIANT
BOUNDED ATOMROWH8
BORROWED PRINCIPLEORTHONORMAL ROTATION
Preserve L2 norm while redistributing concentrated energy across an eight-wide row window.
ROLE AFTER COMPOSITIONINSIDE THE QUANTIZER
An outlier stops holding one block scale hostage; the same transform returns the decoded row to weight space.
SOURCE SURFACE · main.rs + stacks.rsANALOGY MUST SURVIVE MEASUREMENT
Cross-domain does not mean analogy by decoration. Every borrowed idea needs a narrow contract, a legal position in the chain, a measurable cost, and evidence against a baseline.
INTERACTIVE ARCHITECTUREThree candidates.
Three candidates.
No free pass.
This teaching surface makes the acceptance logic visible. Its values are deterministic illustrations—not measurements from a real model.
CONCEPTUAL SIGNAL LAB / NOT A MODEL BENCHMARK
Make the picker decide.
A representative teaching surface where Q2 drifts and Q4 usually clears.
LIVE INTERACTION
ORIGINAL / 64 WEIGHTSATOM-COMPOSED / Q4
ROWH8→Q4→RF REPAIR→PRESERVE→VERIFY
CANDIDATE WALKFIRST PASS WINS
Q2
COSINE0.9857
MAX KL0.714
REJECTQ4
COSINE0.9947
MAX KL0.079
PASSQ8
COSINE0.9994
MAX KL0.006
PASSSELECTED TIERQ4BOTH GATES CLEARED
CROSS-DOMAIN COMPOSITION / v0.2.3Borrowed separately.
Borrowed separately.
Measured together.
Each field contributes one bounded behavior. The final decision is made on the reconstructed weight-space result, after those behaviors interact.
QUANT / Q2→Q8
Walk the bit-width candidates from aggressive to safe instead of forcing one format across every tensor.
DSP + GEOMETRY
Decode through finite-field low-pass repair, then restore each eligible row’s original L2 norm.
INFO + VECTOR GATE
Accept a candidate only when KL distribution drift and cosine direction drift both clear their limits.
INTEGRITY / mix_u32
Borrow avalanche diffusion to tag packed codes and verify every OQ02 record without claiming cryptography.
RECORDED EVIDENCE / HONESTLY BOUNDEDThe spectacular number
The spectacular number
was the wrong number.
The historic KL-only picker reported 10.66× compression, then produced incoherent inference. That failure stays public. The cosine gate was restored, and the last inference-validated baseline became the less dramatic—but honest—2.51× snapshot.
MODELQWEN3.5-0.8B
VALIDATED PIPELINEv0.2.2 DUAL GATE
SOURCE GGUF → OQ2.51×1516.7 MB → 603.7 MB
STRATEGY MIXQ2 0 · Q4 198 · Q8 122
INFERENCE CHECKCOHERENT IN LM STUDIO
Recorded R&D snapshot, not a release guarantee. v0.2.3 adds rowH8 and row-norm preservation, but its full model-level benchmark and downstream validation have not yet replaced this baseline.
BEFORE ALPHAWhat is still
What is still
not solved.
“Under development” is a technical boundary, not decorative copy.
Real activations
The current Gaussian proxy cannot replace imatrix or calibration-corpus evidence from real layer inputs.
Model-level proof
Perplexity, representative tasks, model hashes, and a reproducible v0.2.3 run must precede performance claims.
Format + license
OQ02 compatibility is not frozen, and a project license must be selected before a release.