Split the sentences a listener cannot hold, and check both directions

[jepegit]

SENT-01 caps a sentence at 35 spoken words. 465 of the stress corpus's sentences are over it — the largest family of findings in the project, and the one the rule's own docstring says belongs to a rewriting stage.

The design rule says what that stage must do:

Long source sentences are split, not compressed.

Why a split is the one rewrite that can be checked

A summary of a sentence reads exactly like a split of it. The prose cannot tell you which you were given, so the check has to — and a split is lossless, which no other task here is:

direction	catches	can any other stage-6 task fail it?
every number in the source appears in the split	a summary, a dropped condition	no
every number in the split appears in the source	an invented value	yes
directions and names unchanged (GRD-03)	"higher" for "lower"	yes

That asymmetry is the entire basis for doing this to a paper's own words. A gloss and an anchor are new text and can only be asked whether they invented something; only a split can be asked whether it lost something.

Plus four cheaper checks: at least two sentences back, each under the cap, and no more than 1.5× growth — because a repeated subject costs words but half as much again is an expansion, and expansion is where facts get added.

The paper is never edited

The split is stored on Block.rewrites, keyed by the offsets of the sentence it replaces. The span still points at what the paper wrote — which is what GRD-02 checks the beat against, and what lets study.md show one against the other. Only what is spoken changes.

Two details that are easy to get wrong

Selection is measured on the verbalized text. "298 K" is two words written and four heard. Selecting on the written form misses the sentences that are long because of what they state, which are the ones most in need of splitting. The split is asked to get under 25 rather than 35 for the same reason: a sentence cut to exactly the cap in written words is over it once its numbers have been said.

The subject is repeated, not pronominalised. "X, which does Y" splits naturally into "X. It does Y" — a SENT-02 violation manufactured by the fix for SENT-01. The instruction says so, and the live output does it:

"Silicon has the highest theoretical specific energy density (4008 mAh/g for Li21Si5)… Silicon has yet to find use in commercial lithium-ion cells."

Capped at 40 sentences a build, and run last. Every other task scales with the number of concepts; this one scales with the paper, and a 98-page review offers over a thousand candidates. Running it last means the budget starves splits rather than the four analogies that carry the programme.

First live run

Haiku, one corpus paper, $0.09 for the whole stage (42 calls — the dry-run estimate was $0.09):

attempted: {'gloss': 12, 'why': 12, 'anchor': 4, 'analogy': 3, 'figure': 8, 'split': 10}
succeeded: {'gloss': 10, 'why': 7,  'anchor': 4,               'figure': 2, 'split': 8}
rejected : split — number '2.3': dropped by the split; a split may not lose a value
           split — number '0':   dropped by the split; a split may not lose a value

SENT-01 on that paper: 8 → 2, total warnings 26 → 12.

The gate caught two real failures on the first ten tries, and the sentences they came from are spoken as they stand. One of the two remaining SENT-01 findings is that rejected split — the system preferring a long true sentence to a short lossy one, which is the right preference. The other is a beat the model wrote itself; this stage splits the paper's sentences, not its own, and that gap is untouched here.

The deterministic path is unchanged

--local is still the default, over_long is pure, and the corpus is bit-identical without a model: 2 errors, 10/12 clean.

18 new tests in tests/unit/test_sentence_splits.py, scripted rather than recorded — they check the plumbing and the gate, neither of which should depend on what a model says today.

🤖 Generated with Claude Code