The XS Workshop

The seat, exploded

The house model is a ~32B mixture-of-experts: 39 MoE layers, 256 routed experts each, plus a shared expert every token visits. Below is the router's actual behavior — 25.6M routing decisions from the expert-usage census, one row per layer, one pixel per expert, brightness = share of that layer's traffic (log scale).

The router is heavily skewed: a hot set carries most of the traffic, and 103 experts (1.0% of all expert slots) never fired once across the census corpus. That skew is real capacity information — it is also a trap, which the next two sections are about.

Receipt: expert_usage_census, 160-window hard corpus. A prior version of this census had its layer rows reversed (a module-iteration-order bug) and was quoted in conversation as fact before the instrument caught it — rows here are in model.layers order, corrected 2026-08-28. We publish that because the correction is the method.

What we do with the quant

The served seat is our own 4-bit build — bf16 originals in, agent-trace calibration, outlier-layer protection, then a behavioural battery and an honesty gate that can refuse the build regardless of speed. The published controlled pair is the receipt: same bits, same perplexity (1.4802 vs 1.4806 NLL), and a monotone behavioural difference across a 96-variant probe battery. Calibration is a behavioural lever — a quality axis perplexity cannot see, which means calibration-free pipelines cannot see it either.

Receipts: the sbQ pair on Hugging Face (calibrated build = the served seat; generic-calibration control beside it) and the field note "Calibration is a behavioural lever." Release-qualified: the served build. Research-only: everything else in the sweep, and we label which is which.

The expert program, honestly

The skew above suggests two obvious moves, and we measured both. Promoting hot experts into an always-on shared lane looked like +8.7% — until the instrument was corrected and real routing measured it at −44.6% (best layer −21%, worst −65%): the payload had been measured on an operation real routing cannot run. Crushing the cold tail to 1–2 bits failed its pre-committed quality gate at 17–20× the kill threshold, with an independent honesty regression. Cold does not mean unimportant — measured. Both programs closed as negatives, and both negatives are published because a map of where the walls are is worth as much as a door.

Receipts: promotion-seam probe P3b (real router indices, slot-weighted); cold-crush gate verdict (NLL +0.028/+0.033 vs 0.0016 threshold; say_not_measured −0.09).

The ANE drafter — silicon nobody was using

During inference the Neural Engine sits idle. We reverse-engineered the drafter shape the ANE wants — stateless 64-token window, one fused launch, fp16, full-vocab head — trained it on the seat's own prediction stream, and ran it byte-exact against the seat's verifier. Draft call: 3.58ms under a saturating GPU burn (1.02× its solo time, while the same package on the contended GPU degrades 3.5×). Co-tenancy measured 92–98% in both directions. The drafter draws from an engine that was otherwise doing nothing.

Receipts: the D-ladder result docs (physics, existence proof, cycle-cost anatomy, four-cut surgery) on the omlx fork branch; one-command repro release in preparation. Chain-consistent retraining is running as this page deploys.

The ratchet — current banked truth

Serving speed is gated, not vibed: every change must beat the banked incumbent at the shipped configuration, n≥5, spread reported, thermal-gated, or it does not merge. This is the incumbent as of the last bank.

Model Laguna-XS-2.1-mlx-oq4e-agentic-ours · regime top_k=10 · ngram speculative OFF (shipped 2026-08-28) · snapshot 2026-08-29T17:55:23+00:00. Static snapshot regenerated at site deploy; medians are tok/s over n runs with MAD spread; the spec cell is copy-regime capability, not the served path.