Multi-model check (Sonnet) — 2026-08-03

Every figure in baseline.json was measured on claude-opus-5. This tests whether the shipped instruction transfers to claude-sonnet-5, using the same ten tasks and the same three arms: no instruction, the 315-character rules, and the 176-character compression.

3 variants × 10 tasks × 3 replicates per model. Style only.

Result

model	variant	composite	grade	sentence-simplicity	directness	words	jargon%
opus	A — none	84.7	100.0	48.7	44.5	475.3	11.2
opus	Et — 176 chars	88.8	96.8	70.6	56.1	358.8	8.7
opus	E — 315 chars	88.5	98.8	65.7	53.8	278.7	8.5
sonnet	A — none	85.3	99.3	45.4	48.0	258.6	11.7
sonnet	Et — 176 chars	86.3	98.6	50.2	48.4	230.4	10.5
sonnet	E — 315 chars	87.3	99.9	52.2	47.2	226.5	10.2

The ranking transfers; the magnitude does not

effect against that model’s own control	words	jargon	sentence-simplicity
opus, E	−41%	−2.7 pt	+17.0
opus, Et	−25%	−2.5 pt	+21.9
sonnet, E	−12%	−1.5 pt	+6.8
sonnet, Et	−11%	−1.2 pt	+4.8

Direction holds on every axis for both models, so nothing in the rules is wrong on Sonnet. The effect is about a third the size.

Sonnet writes short without being asked. Its uninstructed output runs 258.6 words against Opus’s 475.3 — shorter than instructed Opus. There is far less headroom for a length instruction to work in, which accounts for most of the gap.

The instruction does not close the quality gap. Instructed Sonnet reaches 10.2% jargon and 52.2 sentence-simplicity; instructed Opus reaches 8.5% and 65.7. On sentence structure, instructed Sonnet is barely past uninstructed Opus. Applying the rules does not make one tier write like the other.

What follows for deployment

The character budget is tier-dependent. On Sonnet the 176-character form is indistinguishable from the 315-character one — 230.4 against 226.5 words, 10.5 against 10.2 jargon. The extra 139 characters buy nothing there. On Opus they buy 16 points of length reduction. Spend accordingly.

The verbosity complaint is substantially an Opus problem. A control of 475 words against Sonnet’s 259 says the rules are a corrective for a tier that over-writes, not a universal improvement.

What this does not establish

Three replicates, one run. The composite gaps on Sonnet (85.3 → 87.3) are small enough to be noise; the length and sentence-simplicity effects are not. Both models generated under a Claude Code system prompt, so some of Sonnet’s terseness may be that rather than the model. Haiku is untested, and on this trend would have even less headroom. No accuracy measurement was taken.