Refresh the fleet from its published sites, and say ten everywhere - #89
realmarcin wants to merge 3 commits into
Conversation
Four cards had drifted away from the Mechs they describe: TraitMech 630 → 696, CellStructureMech 360 → 421 (and 306 → 367 GO-grounded), CommunityMech 374 → 379, ProteinTraitsMech 408,978 → 429,291. Each number was read off the live site and then re-derived independently before it went in. Two of those sites lie in their own HTML, so the figure a scraper sees is not the figure a visitor sees. ProteinTraitsMech ships 408,978 records and 29 sources as static markup and overwrites both at runtime from data/facets.json, which serves 429,291 and 34. MediaIngredientMech ships ellipsis placeholders filled the same way. _fleet/README.md now says to read the rendered number. CultureMech's card no longer takes its Mech's headline tile. That tile reads 10,657 "recipes" and has never matched any data layer: when it was hand-written on 2026-06-30 the corpus was already 15,878 normalized and 6,103 merged, and nothing regenerates it. The card states 6,286 canonical media instead — what CultureMech's own pages/index.html publishes, and what the fleet census counts, so the two halves of the page now agree. The card note says why this one card is sourced differently. Six timestamped backups under data/ingredients/mapped/backups/ were being counted as MediaIngredientMech records, and their prefixes double-counted in the heatmap, so the census read 2,957 where the Mech says 2,951 (#88). A `**` glob cannot exclude a subtree, so roots.py grew an EXCLUDE_DIRS map rather than enumerating mapped/ and unmapped/ explicitly — enumerating would silently miss a new record directory, which is the failure record_paths() exists to prevent. The fleet has had ten members since TaxonMech joined, but the site still said nine, and resources.md listed eight and gave sections to five. Every enumeration now names all ten and links to /mechs/; resources.md gains sections for CellStructureMech, HabitatMech, AntibioticMech, NaturalProductMech and TaxonMech. communitymech.md still claimed 304 communities and pointed both its browser and embedding links at the landing page. Census reread on 2026-09-19; card totals rechecked against the live sites on 2026-09-20. Known gaps filed rather than fixed here: the prefix regex omits TOGO, CultureMech's second namespace at 2,833 occurrences (#84); the CommunityMech and TraitMech checkouts lag their remotes (#85); upstream site defects (#86); TaxonMech still has no census row (#87). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An adversarial pass over the branch raised 27 findings; 7 survived independent verification. All seven are addressed here. The normalized-recipe count was off by one across the site, and this branch made that visible by stating the right figure on the card. Six places said 15,877; the corpus holds 15,878, which CultureMech's own merge_stats.json records as its input_recipes and its README states in a CI-checked block. Corrected in culturemech.md, resources.md and research.md rather than reverting the card. "Has never matched any data layer" was too strong a claim about CultureMech's 10,657 tile, and wrong. It is the March 2026 normalized count, recorded as input_recipes in the Mech's tracked merge_stats_2026.json, and was already four months stale when the tile was written. Saying which layer it came from is both true and more useful than calling it baseless. The NaturalProductMech card listed NPAtlas among its sources, which contradicts the resources.md section added two commits ago, the guidance file, and the Mech's own README — NPAtlas is CC BY-NC and explicitly cannot be seeded into a CC BY 4.0 corpus. It is a cross-reference target, which is why it stays in `vocab` and leaves `sources`; NPClassifier and AntibioticMech, both genuinely adopted, take its place, restoring the nine the README names. Two of the new resources.md sections stated partial coverage as universal. AntibioticMech does not record a mode of action for every antimicrobial: 454 of 2,939 have one and 282 a molecular target. NaturalProductMech carries bioactivities on 176 of 3,115 records, not all, and only 805 of its 3,407 producer claims rest on evidence that addressed the organism. The TaxonMech bullet linked pages/straininfo.html, which 404s, and its card claimed browsers for StrainInfo deposits and AllTheBacteria assemblies that the site does not publish. TaxonMech publishes browse, type-strains and sources; the bullet now points at type-strains. Every culturebotai.github.io link the page carries was then fetched, and all resolve. record_paths() had no test, so the exclusion added for #88 could regress silently. Four cover it: backups are excluded, a sibling directory whose name merely starts with "backups" is not, an exclusion that empties a corpus still raises rather than returning zero, and every excluded Mech is one the census reads — which catches an entry left behind by a rename. Filed rather than fixed: enumerations outside the X-Mech pages that are still short of ten, the site-wide footer among them (#90), and two Quick Navigation anchors that miss the emoji leading-hyphen id (#91). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Review passReviewed the diff adversarially across five dimensions (numbers, ten-Mech consistency, markup/links, the Confirmed and fixed
RefutedThe other 20 were dismissed on verification — most were pre-existing and untouched by this diff, or faults in generated Checks20 tests pass, One thing #90 raises that is worth a thought beyond the fix: nothing mechanically catches an enumeration falling behind. A test asserting that any page naming two or more Mechs names all ten would have caught the footer, |
The sweep stopped at the X-Mech pages. Three enumerations elsewhere were still short (#90), the most visible by far being the site-wide footer, which named five Mechs and linked no hub — on every page of the site. It now names all ten and ends with a link to /mechs/, which also gives the two task-organized tool lists (kg-microbe.md, research.md) a route to the hub without distorting them: those group tools by what a reader is trying to do, so "all ten under For Culture Optimization" would be wrong, and they are left alone deliberately. The separator before MicroGrowAgents, which marks it as the one entry that is not a Mech, needed a margin of its own. The anchors around it carry .55rem leading and .1rem trailing, so an unstyled separator sits 5.3px from the link on its left and 12.6px from the one on its right, reading as part of the hub link rather than as a divider. Measured in a headless browser against the real stylesheet: .45rem evens it to 12.5px on both sides. microgrowagents.md's Related Tools named five Mechs and now names ten. The bibliography entry in publications.md enumerates the organization's repositories, so it names all ten and the orchestrator rather than the same five, and links the hub. Both Quick Navigation anchors for emoji headings were dead (#91). The repo already documents that an emoji heading renders with a leading hyphen in its id, and the deployed HTML confirms it: `## 🔬 Advanced Research Tools` emits `-advanced-research-tools`, where the link said `#advanced-research-tools`. `## 🛠️ Developer Resources` is the awkward one. Its emoji carries a variation selector, and GitHub Pages keeps that codepoint while dropping the glyph, so the live id is U+FE0F followed by `-developer-resources` — an anchor beginning with an invisible character. Dropping the selector from the heading should make it the ordinary `-developer-resources`, matching what the 🔬 heading already emits. That one is an inference from the parallel case rather than something this branch can prove: a locally installed kramdown reports different ids, so it needs checking against the deployed page after merge. Seven of the eight anchors now resolve against the live HTML; the eighth is the one whose heading this commit changes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
d77f902 to
826235d
Compare
1 participant
Stacked on #83 — base is
fix/mech-palette-and-records-label, because both branches regeneratemechs.md. Review #83 first, or read this diff alone.Cards had drifted from the Mechs they describe
data/facets.jsonEvery changed figure was read off the live site, then re-derived independently — against
data/facets.json, the local checkout, and a second fetch — before it went in. The records tile sums the cards by construction, so it moved to 1,075,244.Two sites lie in their own HTML
ProteinTraitsMech ships
408,978 recordsand29 sourcesas static markup and overwrites both at runtime fromdata/facets.json, which serves 429,291 and 34. MediaIngredientMech ships literal ellipsis placeholders filled the same way. A scraper and a visitor see different numbers._fleet/README.mdnow says to read the rendered figure, and to fetch thepages/orapp/URL directly — those landing redirects are client-side meta refreshes thatcurl -Ldoes not follow.CultureMech's card no longer takes its Mech's headline tile
That tile reads 10,657 "recipes". It is hand-typed, was written once on 2026-06-30, and has never matched any data layer — at that commit the corpus was already 15,878 normalized / 6,103 merged. Nothing regenerates it.
The card now states 6,286 canonical media, which is what CultureMech's own
pages/index.htmlpublishes and what the fleet census counts — so the card half and the census half of the page finally agree. The card note says why this one card is sourced differently. This reverses the choice made in e626d77; the evidence that it was never right is new.A census defect, fixed
Six timestamped backups under
data/ingredients/mapped/backups/were counted as MediaIngredientMech records, and their prefixes double-counted in the heatmap: the census read 2,957 where the Mech says 2,951 (2,679 mapped + 272 unmapped). Fixed by anEXCLUDE_DIRSmap inroots.py— a**glob cannot exclude a subtree, and enumeratingmapped/andunmapped/explicitly would silently miss a new record directory, which is the failurerecord_paths()exists to prevent. Full pipeline rerun after. (#88)Ten, everywhere
The fleet has had ten members since TaxonMech joined, but the site still said nine and
resources.mdlisted eight while giving sections to five./mechs/resources.mdgains sections for CellStructureMech, HabitatMech, AntibioticMech, NaturalProductMech and TaxonMechindex.mdnow carries a number for all ten and links the claw orchestratorcommunitymech.mdclaimed 304 communities and pointed both its browser and embedding links at the landing pageFiled, not fixed here
Checks
assemble_page.py --check,refresh_manifest.py --check(10 members), and all 16 tests pass. Jekyll is not installed locally, so rendering is unverified beyond the assembler's own placeholder and Liquid guards.Note:
CultureBotAI/CLAUDE.mdsits above the repository root and is untracked, so corrections made to it there are not in this diff.🤖 Generated with Claude Code