| mechanism under test | per-block level-set selection. A weight matrix is cut into 32-weight blocks and each block is assigned a grid from a nested family — {±1} → {0,±1} → {0,±1,±2} → {0..±3} — so the codebook, not just the scale, is chosen per block under a bit budget. |
|---|---|
| the “multi-channel zero” | {0/0/0} is three zeros. A decoder that maps a channel to the number zero maps all three to zero, so this page puts both grids through the same pass and reports whether the logits come out bit-identical — and what the widened alphabet is charged for it. Both rows are printed either way. |
| elasticity at equal bits | measured here, at a bit budget held equal by construction, against uniform ternary. The verdict printed in that panel is produced by your run; the page does not assert a winner in its copy. |
| what is NOT measured | memristive crossbar arrays, photonic waveguides, “zero-latency” propagation and in-DRAM routing. This is a static page running JavaScript in a tab: it has no such datapath and makes no claim about one. Latency figures here are the cost of the mechanism in software. |
| quality instruments | top-1 of the true next token with its standard error, mean rank of the true token, and Spearman ρ of the 512 logits against the exact pass. Both latter are invariant to logit scale, so no quantiser can buy a better score by shrinking its logits. |
| why not perplexity | on this fixture the exact pass scores worse than a uniform distribution in mean log-likelihood while being far above chance on top-1. A metric that uniform beats rewards degradation, so it is printed as a warning and never used to rank. |
| what a level costs | ceil(log2 levels) index bits per weight, plus one 4-byte scale per block. The measured bits/weight column is reported by the forward pass itself, and asserted against this arithmetic in the test suite. |
| every figure | computed in your browser from the GGUF's real weight bytes, or labelled with the suite that measured it. Nothing is a simulated or definition-derived number. |
| field | value |
|---|---|
| state | press “Stream + parse” |
The pass needs the weight tensors, so the whole file is read, not just the metadata block. Larger models are refused with their byte count and a reason rather than appearing to hang.
| # | token | logit | p |
|---|---|---|---|
| run the pass to see the model's own next-token distribution | |||
| grid | bits/weight | RMSE | mean rank | top-1 | rho vs exact |
|---|---|---|---|---|---|
| measure the ladder to fill this table — the exact pass is measured first, because every rho is against it | |||||
Mean rank is the position of the true next token in the model's own ordering; a uniform guess averages 256.5 on this vocabulary, lower is better. rho is the rank correlation of all 512 logits against the exact pass. Top-1 is shown with its standard error in the log, because at these position counts it cannot resolve small differences.
| configuration | bits/weight | logits identical to {0}? | mean rank | top-1 |
|---|---|---|---|---|
| the claim is that widening the zero into three channels is an information channel; the test is whether a decoder can tell them apart | ||||
Compared bit-for-bit across every position and every vocabulary entry — not a tolerance, not a similarity.
| plan | bits/weight | mean rank | rho vs exact | rank diff vs ternary |
|---|---|---|---|---|
| uniform ternary is the baseline at the same average bits; a positive rank difference means the plan is worse | ||||
The plan promotes the most important blocks to the 3-bit grid and demotes the least important to the 1-bit grid until the promoted and demoted parameter masses balance, which holds the average to 2.0000 index bits. Both plans are shown against the same uniform-ternary baseline.