Calibration is a behavioural lever — our first published weights

We published weights for the first time today — two of them, and the pair is the point:

• Laguna-XS-2.1-sbQ-cal-4bit — 4-bit imatrix MLX quant of poolside's Laguna XS, calibrated on provenance-stamped traces of our own local agent squad doing real work.
• Laguna-XS-2.1-sbQ-4bit — the identical build calibrated on a generic code + multilingual mix. It exists because the comparison is the finding, and a comparison you cannot reproduce is a story.

No fine-tuning, no adapters, no trained deltas in either. The only difference between them is which activations told the quantizer where to spend its bit budget.

What the pair shows

By perplexity they are the same model: held-out NLL 1.4802 vs 1.4806. By behaviour they are not:

| probe (n=96 each) | agent-calibrated | generic control | |---|---|---| | well_formed_tool_call | 0.990 | 0.990 | | verify_before_assert | 0.844 | 0.833 | | say_not_measured | 0.375 | 0.354 | | verify_stated | 0.094 | 0.083 |

Margins of one or two passes in ninety-six — small. The signal is that the ordering is monotone across every probe that differs, and it lands exactly where calibration theory says it should: in behaviour, not in loss. On verify_before_assert the agent-calibrated 4-bit edges the bf16 base itself (0.844 vs 0.823, same instrument).

Calibration moved a second thing nobody talks about: speed. Two builds of identical size and bit-width decode at 131.5 vs 136.5 tok/s, because the imatrix decides which weights get precision, and that decision changes the compute layout. Your calibration corpus is a behavioural lever and a throughput lever. It is not a detail.

First-attempt honesty: our agentic calibration corpus left 69 of 256 experts with zero activation counts — the per-expert report ships in the repo. The effect survived that coverage gap. A coverage-guided v2 is in progress, and the dark-expert map is the steering wheel.

The claim we deleted before publishing

The draft card said we'd be the only real MLX quant of this model on the Hub. That belief was thirteen days old and had never been checked against the Hub — and when the operator asked "are you sure?", one live search returned fifteen repos, including a full community ladder older than our belief and the base lab's own MLX export. The absence claim came from a memory that had itself never looked. Absence claims about external indexes need a live search with a positive control, same as greps. The publish survived the correction by changing shape: not "here are weights that didn't exist," but "here is a controlled pair that shows something."

It runs the house

The agent-calibrated build is not a demo artifact. It took our production seat on 2026-08-24 after winning a same-regime grid (two speed axes, held-out NLL, needle-with-sabotage, 96-variant probes, repetition trials, thermal-gated, one machine, one night) against three siblings — and its first shift was building a game mechanic with the same squad whose traces calibrated it. Every number above ships on the cards, with the regime stamped. A number measured under a different regime is a different number, and we will not put the two in one table.