Human review of the PROPOSED records

[realmarcin]

ribosome, bacterial-type flagellum, carboxysome and S-layer were LLM-drafted from the cited reviews and are PROPOSED. Before promoting any to REVIEWED, a human should: (1) open every cited DOI and confirm each notes paraphrase; (2) confirm the physical-property values and their sources (ribosome mass is the weakest); (3) assign InterPro groundings to the flagellum components (open CURATION_TODO in the record); (4) decide whether the ribosome's two paralogue-set components should be split (open discussion in the record); (5) check the flagellum stoichiometry claim (~20,000 flagellin subunits) is actually in Macnab 2003.

[realmarcin]

Retitled: it said four seed records. There are now 16, and every one is PROPOSED. The curators appearing in curation_history are claude, codex and three scripts (uniprot_sl, fetch_commons_image, complex_portal) — no person has read any record.

The number will keep drifting, so the title no longer carries one. Scope is unchanged: promotion to REVIEWED is a human decision, and nothing in the tooling can close this.

[realmarcin]

Progress on the machine-checkable parts of this list, so the human review starts
from a smaller and better-defined problem.

Item 5 — "check the flagellum stoichiometry claim (~20,000 flagellin subunits)
is actually in Macnab 2003": it is not. That review is abstract-only in PubMed
and the figure does not appear in the reachable text.

Item 2 — "confirm the physical-property values and their sources (ribosome mass
is the weakest)": the weakness is broader than the mass. All six quantities on
the flagellum and ribosome rest on abstract-only sources in which none of the
numbers appears. Full inventory and options in the issue filed alongside this
comment.

Item 1 — "open every cited DOI and confirm each notes paraphrase" is now
partly mechanised. just evidence-audit reports what is reachable per reference;
just evidence-verify --strict checks every verbatim snippet against its
source. That does not check paraphrases — a notes field cannot be verified
mechanically, which is the argument for snippets — but it narrows the manual pass
to the paraphrases that matter.

The method has already found a real error of exactly the kind item 1 is looking
for: #161, where an evidence note attributed to Shi et al. 2018 a statement about
RodA that the paper does not contain, because RodA's role was established five
years after the other cited review was published.

Items 3 and 4 remain untouched and are genuinely human: assigning InterPro
groundings to the flagellum components (the record's own CURATION_TODO says
they were left blank rather than guessed), and deciding whether the ribosome's
two paralogue-set components should be split.

Promotion to REVIEWED remains a human decision and nothing here changes that.