Corpus crosscheck — 3.1 structural recapture (2026-07-28)

Human review artifact for the regeneration that replaced the migrated-zip capture with a parser reading the official IFC-SG glossary pages.

How this was produced

1. python scripts/parse_glossary.py — parses the structural layer (bindings, subtypes, S/N, property, type, applies_to, unit, input_limited, examples) out of the cached official glossary pages. Carries the non-derivable fields forward from the previous parameters.json (discipline, authorities, source*, page_range, header_kind, binding pages, and the editorial annotations example_format_acmv / prefix / glossary_key / glossary_mandatory).
2. python scripts/resolve_enums.py — re-applies the value layer from _meta/enum_resolutions.json. The parser never writes allowed_values / reference_values.
3. scratchpad/validate_against_cop.py — cross-validates every parsed property against the CoP v3.1 PDF (sha256 1f9f7197…fbc2a, now recorded in manifest.json).
4. python scripts/validate.py → 0 errors, 0 warnings. python scripts/build.py twice → byte-identical output.

Anti-silent-drop contract. A dropped row is the failure that produced the previous capture's losses, so the parser checks for it three ways, two of them measured independently of the code doing the work:

• <tr> census — each table's expanded grid row count is compared against a plain count of <tr> tags in the source HTML, so a row lost during rowspan/colspan expansion cannot pass.
• S/N continuity — the official page numbers its own rows, so the parameters read off one table must carry S/N 1..n with no gaps. This catches a real data row misread as blank, as a repeated header, or as a footnote. It is the check that would have caught the previous capture: Lift's two surviving rows were S/N 2 and 3 of 8.
• Row accounting — every row below a header must be classified as exactly one of blank / repeated header / declarative sentinel / parameter.
• Blank-cell resolution — an empty (as opposed to -) requirement cell is a page defect and must be answered from the CoP PDF; an unanswered one aborts. See below. The guarded set includes Input Limitation, the column that decides whether a hard allowed_values constraint attaches at all — a blank there can no longer default to "not limited", and an unrecognised value aborts rather than falling through to false.
• No silent exemptions — a table whose S/N column cannot be established aborts instead of skipping the continuity check, and the Pollution Control listing table gets the same row reconciliation as the main tables.

All ten guards are verified against synthetic failing inputs on every change.

Those, plus any unparseable table or unrecognised Input Limitation value, abort the entire run with the component named. A partial or degraded file is never written. All three guards plus five others were verified against synthetic failing inputs (8/8).

Components added / removed

None. The official index lists 75 components; the corpus holds 74. The gap is fully explained and unchanged: Space Usage (Others) is not a separate component here — it is the page that publishes the 31-property IfcSpace / SPACE table that arc/space_usage already carries. The parser reads that component's structure from page d0a72875-f11a-4a53-9a1b-bfd84b9ea775 while keeping source_id on the parent Space Usage page (this is declared in arc/space_usage/parameters.json → glossary_pages).

Properties added (36 net)

Properties removed (1)

Binding changes

Multi-entity table banners (IFC Entity: IfcWall, IfcSlab) are split into one binding per entity, which is the convention the previous corpus already used for Fire Access Opening, Household Shelter, Pipes/Ducts and Prefabricated. Accessible Route is the only component where applying that convention uniformly changes the binding count.

Type / unit / input-limited changes

examples were rebuilt for every component. The previous capture read the CoP PDF's multi-column example cells row-wise and produced garbled lists — Space (Usage) EmergencyVoiceCommunicationSystem was ["Address System."] and is now ["1-way EVC System", "2-way EVC System", "Public Address System."]; Column SpliceDetail was ["Detail 3 1st Storey", "F2"] and is now ["Detail 3"]; Space (Area Requirements) AGF_DevelopmentUse was a 25-item interleave of four PDF columns and is now the official 25-value list in page order. Example cells are stored split on commas, which render.py re-joins with ", ".

Blank cells on the live page, filled from the CoP PDF

A requirement cell holding - is the page saying "not applicable". A cell that is empty is a rendering defect, and the two must not be conflated: an earlier pass of this regeneration wrote type: null for Column Stirrups / StirrupsType because it treated them the same. The parser now separates them, and every empty cell must be answered by a CoP-sourced entry (LIVE_CELL_GAPS in scripts/parse_glossary.py, each quoting its PDF row) or the run aborts — a future recrawl cannot quietly open a new hole.

Six cells corpus-wide are blank. Two carry real values in the PDF; the other four are blank in the PDF too (-), so they resolve to "omitted" and change nothing:

Normalisations applied (the complete list)

Every one of these is a spelling/rendering defect where the CoP v3.1 PDF is unambiguous. Nothing else was changed silently.

lfc → Ifc (lower-case L rendered for capital I) — 21 IFC-entity banners across 15 components: Beam, Damper, Distribution Chamber, Interceptor, Pile, Sanitary Appliances, Space (Usage), Staircase, Tank, Type Bedding for Pipe, Valve, Wall, Waste Terminal, Water Meter, Window.

Deliberately not normalised — published spellings preserved verbatim because both sources agree: AI-DioxinsAndFurans (hyphen, Pollution Control), BarrierFreeAccessbility (Lift), and the Space (Area Requirements) example values Assisted Living Facilitys, Service Appartment II, Electrial Substation, plus the duplicated Mechanical Ventilation in Space (Usage) VentilationMode.

Footnote markers stripped from property names

Markers are removed from the stored property; the original cell is kept in property_raw. The footnote text itself has no corpus field, so it is recorded here:

A * on a value (e.g. Input Limitation = No*) is the page's "parameter is populated from the dimensions of BIM elements modelled" note. It appears under Column, Slab, Wall, Beam, Borehole, Footing / Pilecap, Pile and Household Shelter and does not change the parsed value.

Cross-validation against the CoP v3.1 PDF

scratchpad/validate_against_cop.py. Page offset verified empirically before use: corpus page_range is Section-4-relative and PDF page = value + 246 — 722 of 723 parsed properties land inside their own page range at +246, while the next-best offset scores 0.66.

• 0 of 723 parsed properties are absent from the PDF.
• 1 parsed property is in the PDF but outside its component's page range: Tank / CompactionRatio, which the PDF carries on p.335 (Refuse Chute). Consistent with the REFUSEHANDLINGEQUIPMENT subtype move above.
• 0 unexplained "in the PDF but not parsed" candidates. All 81 raw candidates (across 31 components) were ruled out by inspecting their PDF context: Property Type values (Length, Volume, Area) — Length alone accounts for 26 of them — prose (Purpose, Public, Private, Vehicular, Duty, Accessway, Egress), a figure caption (Stirrups), values of other properties (TradeEffluent), worked-example tables (Space (Area Requirements) p.360), overlapping page ranges (Space Usage 114–171 spans the Space Area Requirements pages), and one explicit exclusion — CoP p.292 states that PileModelFactor / ShaftR4DesignFactor / EndBearingR4DesignFactor do not apply to Footing / Pilecap.
• Entity/subtype disagreements: exactly the ones listed in Binding changes above.
• Property Type compared for 526 parameters. Name-presence alone proved too weak — it passed both the blank-cell defect above and the Distribution Chamber type change — so the check now asserts our type against the type token the PDF prints in that property's row. Result: 0 parameters typed null where the PDF states a type, and 2 genuine mismatches, both already recorded above (Distribution Chamber TopLevel / InvertLevel, ours Text vs PDF Real). Two properties have no type token beside their name in the PDF and are explained: Space (Area Requirements) ACN_ConnectivityType (the PDF spells the name ACN_ConncectivityType) and Tank CompactionRatio (the PDF prints it on the Refuse Chute page, outside Tank's page range — and types it Real, the conflict noted above). Pollution Control's 58 AI_/TED_ rows are exempt by name: their two-column listing table states no type on either source.

Value layer

resolve_enums.py alone could not re-apply 46 value lists that the migrated corpus carried inline, because they had never been declared in _meta/enum_resolutions.json. 44 were re-declared there verbatim — 43 sourced from _reference/structural_enums.json (each asserted equal to its enum group before writing) and 1 = Site Boundary BroadLandUse (32 values, CoP p.349). The existing resolver now reproduces them. The remaining 2 are intentional, not losses:

• System SystemType — now reference_values instead of allowed_values (see above).
• Roof ConstructionMethod — same treatment, applied late; see the correction below.

Correction — Roof ConstructionMethod keeps its vocabulary

An exhaustive old-vs-new audit of every value list in the corpus found one property that had lost its vocabulary outright: Roof ConstructionMethod carried the 7 structural-enum values as allowed_values before, and briefly carried nothing after.

Roof is the same structural case as System SystemType — the live page and CoP p.339 both mark Input Limitation No while the example still reads "Refer to list*" — but the two were handled inconsistently: System kept its vocabulary as a soft reference_values, Roof kept none. Per the project's own calibration (official vocabulary on a free-input parameter is retained as a suggestion, not a constraint), _meta/enum_resolutions.json now carries a reference_resolutions entry for Roof ConstructionMethod with the 7 values copied verbatim from _reference/structural_enums.json and asserted equal to what corpus v3.1 carried. All three Roof bindings now expose them as reference_values.

The audit is now clean: 0 properties lost a vocabulary, and exactly one value-layer change remains corpus-wide — Roof ConstructionMethod allowed:7 → reference:7, which is this intended re-calibration.

A stale entry was also removed: System / SystemType still appeared in the hard resolved list alongside its documented soft entry. It was inert only because resolve_enums.py gates the hard branch on input_limited; had a later recrawl flipped that flag, the parameter would have silently regained a hard constraint against intent.

Input-limited parameters still awaiting a resolution

The other 11 new properties (the DA1- / WorkingLoad_ / MinEmbedment* block, NominalCapacity, CompactionRatio) are free-input on both sources and need no resolution.

Component README.md / notes.md regenerated

Both files are shipped raw to dist/ and inlined into llms-full.txt, so after the corpus regeneration they contradicted the very data they describe — arc/lift/README.md and the llms-full.txt copy of it both still read "Parameters | 2" against a parameters.json holding 8. They are now generated by scripts/generate_component_docs.py, which runs after resolve_enums.py (notes cite allowed_values) and asserts, per component, that the stated counts and the parameter-table row count match parameters.json — a contradiction now fails the run instead of shipping.

75 files regenerated across 47 components: 45 README.md + 30 notes.md. 72 of them belong to the 44 components whose parameters.json this regeneration changed. The other three are the settled rule decisions below.

Editorial prose promoted to data. Four documents held hand-written passages. They now live verbatim in _meta/editorial_notes.json and are emitted from there, so the generator holds no per-file exceptions and "the corpus is generated, never hand-edited" stays literally true. All four were verified byte-identical to the originals after regeneration:

Corrections made while regenerating

• arc/site_boundary — BroadLandUse re-filed. It sat under "Input-limited parameters resolved against structural enums" although its allowed_values_source is CoP v3.1 p349 (source table), not a structural enum. Parameters are now filed by their actual source, so it moves to "Other input-limited parameters" — consistent with Ceiling / FireRating, which has always been filed that way. A presentation fix; the values in parameters.json are unchanged.
• arc/pollution_control/README.md — Type column fixed. A pre-existing bug in the original generator rendered a null type as the literal string None, so all 58 AI_/TED_ rows read | None |. An absent value now renders as an empty cell.
• List ordering is first-appearance / source order, mirroring the official table rather than alphabetising it. Two files were alphabetised by the old generator and change under the rule: arc/green_verges/notes.md and arc/ramp/notes.md (Ramp's list now reads in S/N order 3–8, as the page prints it).