BLUF as a label, and four rules for explanations — run 5, 2026-08-08

Two findings. Naming a technique instructs at least as well as spelling it out, and compresses harder. The four explanation rules cost 19% more words for a vocabulary gain prosemeter can barely see — which is the outcome that was predicted in advance, and the reason they still need a human verdict.

Four arms, six tasks (T1–T6), five replicates, n=30 per arm, 120 answers. claude-opus-5, scored by prosemeter 0.4.2 against the chat profile.

arm	instruction
A	none — control
E	the five shipped rules
F	E plus four explanation rules
L	E with rule 1 replaced by “Use BLUF: bottom line up front”

Results

arm	words (sd)	vs control	jargon%	sentence-simplicity	grade-band	clarity	composite
A	655.5 (83)	—	10.7	52.7	100.0	76.6	84.8
E	454.5 (62)	−30.7%	8.9	68.2	99.2	89.1	89.6
F	540.0 (104)	−17.6%	8.0	67.2	99.3	86.7	89.1
L	406.7 (81)	−38.0%	8.5	67.4	98.2	90.1	89.8

The composite column is there to be dismissed. E, F, and L span 89.1–89.8 against within-arm spreads of 12 to 16 points, so no composite difference in this run is interpretable. Every finding below comes from the dimension columns.

L — the label carries the sentence, and then some

Replacing “Lead with the answer, then explain” with “Use BLUF: bottom line up front” did not weaken the instruction. It shortened the output by a further 10.5%.

L cuts 38.0% off the control against E’s 30.7%. The 47.8-word gap between them is about 2.7 times its standard error, so it is unlikely to be sampling noise — though it is one model on one day.

On everything else the two are indistinguishable:

	E	L
jargon%	8.9	8.5
sentence-simplicity	68.2	67.4
grade-band	99.2	98.2
clarity	89.1	90.1

Only rule 1 differed; rules 2–5 were held constant. So the length difference is attributable to the swap rather than to a rewritten instruction.

The gloss after the colon was deliberate. A bare acronym would have confounded “does the model know the term” with “does a label instruct as well as an instruction”, and only the second is interesting.

What this does not show. That labels beat sentences in general. One label was tested, for a technique with a widely-known name and a one-line gloss. A label for something the model has no prior for is a different experiment.

F — the prediction held, including the part that matters

variants.md recorded before the run: “Predicted to score worse, not better… words should rise toward the control while jargonPct holds near E.”

Both happened. F adds 85.5 words over E (+18.8%) and gives back a third of E’s length cut. Its jargon is the lowest of any arm at 8.0 — 0.9 points under E, against a control of 10.7.

Sentence-simplicity is flat across all three instructed arms (67.2–68.2), which is expected: the four rules operate on document structure, and sentence-simplicity counts hard sentences.

The composite cannot separate F from E, and that is the finding, not a failure of the run. The four rules exist because a reply scoring 85 on this profile drew “I have no idea what you’re saying”, and the rewrite that landed scored 84. This run reproduces that blind spot on 30 fresh answers rather than on one anecdote: the arm built to fix a legibility failure moves the composite by −0.5.

So F cannot be adopted or rejected on this evidence. What the scores establish is its price — 19% more words — and that the price does not buy a vocabulary regression. Whether it buys legibility is a human judgment, and the answers are on disk to make that judgment against.

The control moved 50%, and that is a harness confound

The control arm ran 655.5 words. Run 3’s control, on the identical six tasks and the same model name, ran 437.9.

Relative effects survive it intact. E cuts 30.7% here against 32.0% in run 3. Control jargon is 10.7 in both runs, to the decimal. So the shift is specifically in length and it moved every arm together.

The likely cause is the generation harness, not the model: these answers were written by Claude Code subagents holding tools and writing files, where earlier runs used a different setup. That is a difference in generation conditions, and it means absolute numbers from run 5 are not comparable to run 2, 3, or 4. Within-run comparisons are unaffected, which is the entire reason eval/README.md insists on a control arm in every run.

compare.mjs fired a false regression, and the control is what exposed it

words   baseline 290.1   observed 454.5   +164.4   REGRESSED (> 360)

The gate is correct on its own terms and wrong about the world. Its tolerance is an absolute word count calibrated against a 437.9-word control; when the whole population shifts up 50%, an unchanged instruction trips it.

No tolerance was widened in response. Widening it would destroy what the gate is for — catching a drift back toward uninstructed length — and this run gives no evidence the instruction stopped working. The fix, if one is wanted, is for the gate to compare against a within-run control instead of a stored absolute, which is a design change that needs its own justification.

Recorded here so the next run that trips this recognises it rather than re-deriving it.

Reproducing

pnpm build
node eval/score.mjs   eval/out-run5
node eval/compare.mjs eval/out-run5 E

Answers are gitignored — see “Why outputs are not committed” in eval/README.md.

What this run does not establish

• Nothing about factual accuracy. No reviewer pass was run, per the standing finding that instruction wording moves length and vocabulary 2–3x and moves correctness not at all.
• Nothing about whether F reads better, which is the only question F was written to answer.
• Nothing about other models. Ranking has transferred to Sonnet 5 before; magnitude has not.