A clause is not a concept

[jepegit]

38 of the corpus's 482 concepts were clause fragments, and the programme asked questions about them:

"What did they report for figure shows?"

One paper had figure shows, shows the XRD pattern and scan shows the xrd among its 27 concepts. Another had see Figure and see Table. Also becomes amorphous, containing species, consistent with the presence, peaks characteristic, capacity associated, against sodium metal counter, discharged to mv, above mv.

Both guards were written in the past tense

PHRASE_BOUNDARY_VERBS is 35 entries and every one is a past participle — measured, observed, shown, reported… So a phrase opening or closing on a finite verb went straight through.

CLAUSE_LIKE, whose own docstring says "a finite verb makes a phrase a clause rather than a term", had the auxiliaries and nothing else. A verb in the middle was therefore invisible to both: the boundary check cannot reach it, and CLAUSE_LIKE did not know the word. Hence scan shows the xrd.

The lead family is deliberately absent from CLAUSE_LIKE. In a battery corpus "lead" is the metal, and a rule that loses lead acid battery to catch "leads to capacity fade" is a bad trade. Tested.

Corpus

	before	after
clause fragments	38	8
concepts	482	442
SEG-03	27	23
errors	2	2

A segment reported as introducing five new terms, when two of them were the same noun phrase cut in different places, was the working-memory rule being broken by arithmetic rather than by content.

The eight that remain are reduced relatives — electrodes cycled with LiPF6, electrode cycled in EC. They are at least comprehensible as terms, and the judgement separating them from cycled electrode surface is not one a word list can make.

What this does not fix, and why I stopped

43 pairs where one concept is a sub-phrase of another: metal counter beside metal counter electrodes beside sodium metal counter electrodes.

_drop_redundant already decides these on the ratio of their mention counts (drop the shorter at ≥0.8, the longer at <0.35). Measuring every surviving pair in the corpus says the ratio cannot separate them:

0.75  'thin film' (12) inside 'thin film electrodes' (9)      <- both real terms
0.75  'specific surface' (8) inside 'specific surface area' (6) <- both real
0.75  'metal counter' (8) inside 'metal counter electrodes' (6) <- a truncation

What distinguishes them is that a truncation ends on a modifier rather than a head noun — metal counter, a-Si thin, Si hollow, TOF-SIMS depth — which is a judgement about English that a regex does not have. I tried two window-level discounts before measuring this; one was too weak (it compared against a single containing phrase instead of their union) and the positional one cut Irisarri from 32 concepts to 5, taking irreversible capacity and graphitization degree with it. Reverted both. Left alone rather than guessed at — this is the shape of thing docs/PLAN-ai.md assigns to the critic role.

🤖 Generated with Claude Code