| --- |
| title: DMCRA Elastic Centroid Lab |
| emoji: 🧪 |
| colorFrom: purple |
| colorTo: blue |
| sdk: static |
| app_file: index.html |
| pinned: false |
| license: mit |
| short_description: Elastic per-block level-set quantization, measured |
| --- |
| |
| # DMCRA Lab — elastic per-block level sets, measured on real weights |
|
|
| A client-side lab for the **Dynamic Multi-Centroid Routing Architecture** idea: instead of a fixed |
| ternary grid `{-1,0,+1}` on every weight, let the *codebook itself* be chosen per block from a nested |
| family (`{+-1}` → `{0,+-1}` → `{0,+-1,+-2}` → …) under a bit budget. It runs the real llama forward |
| pass over the real weight bytes of a real GGUF, in your browser, and reports what each choice costs. |
|
|
| No server, no GPU, no upload, no CDN, no build step. Every figure is measured on your machine. |
|
|
| ## What it measures |
|
|
| 1. **The level-set ladder** — `binary`, `quad`, `ternary`, `penta`, `hepta`, `oct`, `nona`, each run |
| through the real pass. Reported: the pass's own measured bits/weight, reconstruction RMSE, the mean |
| rank of the true next token, top-1, and Spearman rho of all 512 logits against the exact pass. |
| 2. **The multi-channel zero** — `{0}` vs the document's `{0/0/0}`, compared logit by logit for exact |
| equality. Reported whichever way it comes out. |
| 3. **Elastic allocation at equal bits** — a per-block plan that unclamps the most important blocks to |
| the 3-bit grid and clamps the least important to the 1-bit grid, held to exactly 2.0000 index bits |
| per weight, measured against uniform ternary at the same budget. The panel prints `UNDECIDED` when |
| the difference is smaller than the run can resolve. |
|
|
| ## Verdicts (Node suite, 89 scored positions) |
|
|
| | claim | measured | |
| |---|---| |
| | elastic unclamping beats uniform spending at equal bits | **not supported** — every keyed plan was worse (+46 to +147 rank positions, \|t\| 3.2-5.8) | |
| | the keys still matter | **supported** — keyed plans scored rho 0.52-0.55 against a key-free control's 0.06 | |
| | `{0/0/0}` is an information channel | **refuted exactly** — bit-identical logits, an extra index bit charged | |
| | the zero-free grid at equal bits (`quad`) | **supported** — rho 0.8659 vs ternary's 0.8359 at the same 2 index bits | |
| | per-block least-squares scales | **the largest effect found** — ternary rho 0.5248 → 0.8359 | |
|
|
| Per-claim detail, including what a static page structurally cannot test (memristive, photonic and |
| in-DRAM claims): `reports/dmcra-verdicts.md`. |
|
|
| ## Verified |
|
|
| 25 quantiser assertions (including that this project's ternary reproduces the sibling project's |
| already-verified `toTernary` exactly), 21 end-to-end forward-pass assertions, and 22 assertions |
| driving this exact page through a stubbed DOM — refusing an over-ceiling model, refusing an empty |
| prompt, and reporting an under-powered comparison as undecided rather than as a result. |
|
|
| Part of the AEUPH studio. |
|
|
