Build status · measured 10 August 2026
The Approved Documents, indexed and quotable.
Verification of building plans against the Building Regulations for England. The whole corpus is Crown copyright under the Open Government Licence, which means a finding can show you the clause rather than just name it.
19
Approved Documents
24 PDFs, 1,633 pages
5
Instruments of legislation
876 provisions, as XML
1,594
Indexed passages
31 sources, searchable
100%
Freely reproducible
OGL v3, commercial use
One
What it can do
Six things, concretely. Everything here is built and working today.
Read a drawing set and say what it found
Extracts occupancy, construction, glazing, escape provision and energy measures from the PDFs you upload — and lists explicitly what it could not read off them.
Check a design part by part
Runs against any of the 19 Approved Documents, one part or all of them, returning a pass, fail, warning or needs-more-information against each clause it reaches.
Quote the clause, not just name it
Because the corpus is Crown copyright under the Open Government Licence, a finding can show the text of the provision it turns on rather than sending you to look it up.
Point at the sheet
Every finding names the drawing or detail its evidence came from, so it can be checked against the design rather than taken on trust.
Leave an auditable record
Dated, clause-referenced, with the unknowns written down. Part 2A asks a dutyholder to evidence reasonable steps — this is the shape of that evidence.
Answer from your own tools
An MCP server exposes projects, documents, regulation search and findings, so you can ask questions from Claude directly instead of working in the interface.
Two
What it cannot do
Worth reading before the rest. If any of these is a dealbreaker, nothing further down changes it.
It does not hold a duty
Under Part 2A and CDM 2015 the duty attaches to a person or organisation. A tool cannot carry it. What it removes is the document administration, not the responsibility.
It does not approve anything
Findings are advisory. They are not a building control decision, and not a substitute for a competent person's sign-off.
It does not guess
Where the drawings do not show enough to decide a clause, the answer is that more information is needed. A false pass is more damaging than a question.
England only
Wales has its own Approved Documents, Scotland uses Technical Handbooks and Northern Ireland Technical Booklets. Those are separate corpora and are not loaded.
It does not read badly scanned drawings
Legibility caps result quality. A downscaled sheet with unreadable dimension text produces questions rather than answers.
It does not track your editions yet
The corpus records which edition it holds and detects when gov.uk republishes one. Filtering retrieval by a project's application date is designed, not built.
Three
How it is made up
The corpus it reasons over, what may be reproduced from it, and how it was built.
Circle size is passages held, or citations received for a standard. Line thickness is how often one document cites another. Hover to trace, click to open the index. 22 standards are drawn individually; the remaining 255 are grouped into one node so the totals still add up.
Everything held
26 entries · 1,594 passages
Covering 31 source documents — a two-volume Approved Document is one entry here.
Choose one on the map or in the list to open its index.
Figures are indexed passages. The 277 British Standards cited across this corpus are not listed here because none of them is held — they appear on the map outside the boundary.
Every document, at the edition held
Every Approved Document currently published for England, at the edition live on 10 August 2026. Bar length is page count.
Legislation, as XML
Structured and section-addressable — better source material than any PDF.
- The Building Regulations 2010Schedule 1 — the functional requirements
- Amendment (England) Regulations 2023Part 2A — dutyholders and competence
- Building Act 1984The parent Act
- CDM Regulations 2015Health and safety in design
- Building Safety Act 2022Higher-risk regime and the BSR
Guidance
Covering both duty-holder roles, not just building regulations.
- HSE L153CDM 2015 guidance · 90p
- Manual to the Building RegulationsCode of practice · 62p
L153 is guidance, not an Approved Code of Practice — HSE decided CDM 2015 would not have one. Its text is reproducible; its figures are not.
What can be quoted, and what can only be cited
The distinction is enforced in the database and again on the server after the model responds — not by asking the model nicely.
Held and quotable
Approved Documents, the Building Regulations, CDM 2015, HSE L153 and the Manual. All Crown copyright under the Open Government Licence v3.0, which permits copying, distribution and commercial use with attribution.
A finding may show the clause text directly, up to a sensible length.
Cited by reference only
BS 9999, BS 9991, BS 7974, BS 7671, BS 5839. BSI’s terms prohibit storing extracts in a retrieval system, and a vector index is exactly that — so these are never ingested.
A finding names the clause and points you at your own licensed copy.
This matters less in England than it would elsewhere. Approved Document B is the mainstream prescriptive route and it is free; the British Standards are the alternative routes. Attribution is a condition of the licence wherever that text is shown, not a courtesy.
How the corpus was built
Five stages, from a public URL to a citation in a report.
- 01Acquire
Asset URLs resolve through the gov.uk content API at run time. gov.uk mints a new URL whenever a document is republished, so a hardcoded list silently serves a superseded edition — and resolving live makes a changed URL a new-edition alarm.
- 02Extract
All 26 PDFs extract cleanly, including Approved Document D, whose published asset is a locked PDF. Legislation skips extraction entirely: it arrives as structured XML with permanent, section-level identifiers.
- 03Chunk
Split at real structural boundaries — Section, Appendix, and the Schedule 1 requirement boxes — with paragraph numbers carried through as the citation. This is the part that had to be rebuilt; see below.
- 04Index
Hybrid retrieval: keyword and vector legs fused by rank. Keyword alone handles clause numbers and exact terms well, which is why the corpus is useful before embeddings are generated.
- 05Verify
Two stages. Read the drawings and state the facts, including what could not be read; then check those facts against retrieved clauses, one part at a time.
The part that had to be rebuilt
An Approved Document is not a standard
The indexer was written for South African standards, where 4.2.1 opens a section and is followed by a short title. In an Approved Document 2.21 opens a body paragraph and is followed by prose. Because extraction hard-wraps, the first line of every paragraph looked exactly like a heading.
Before
477
false headings detected in Approved Document B, Volume 2
- — every paragraph became its own chunk
- — each labelled with a sentence fragment as its title
- — one span swallowed pages 14 to 151 in a single passage
After
20 / 20
sections found, in document order, with no overlap
- — split at Section, Appendix and requirement boundaries
- — paragraph numbers carried through as the citation
- — page furniture and watermarks stripped before indexing
# a passage as it is now indexed
clause_ref = 2.10–2.15
heading = Approved Document B Volume 2 > Section 2: Design for horizontal escape
pages = 27–30
“2.10 Alternative escape routes should satisfy one of the following criteria…”
The reference is a paragraph range because that is how a professional cites an Approved Document — “ADB Volume 2, paragraphs 2.10 to 2.15” — not a chunk index that means nothing outside this system.
What runs where
The regulations and your drawings are different kinds of thing, and they are handled differently. On-premise deployment is available.
Ships with the product
The corpus
All 31 documents and 1,594 passages travel as a single 1.2 MB file, restored in seconds. No downloads on install, no extraction, and every deployment holds an identical, versioned edition.
It is public data under the Open Government Licence. There is nothing confidential in it and no licence to negotiate.
Stays on your infrastructure
Everything about your projects
Drawings, projects, findings, users and your encrypted API key. The application, database, file storage and regulation search all run on your own hardware. No third party holds your project data.
Each company’s key is scoped to that company, and two firms on one server never see each other’s work.
Leaves the building
The pages sent for analysis
The drawing pages under review go to the model, under your own API key and billed to your own account. We never see that traffic or that bill.
If vector search is switched on, short phrases built from the extracted plan facts also reach the embedding provider. The drawings themselves do not. It is off by default.
This is not air-gapped, and we will not say otherwise
The analysis is a frontier-model call. There is no local model in the system, and substituting one would read drawings materially worse — vision over dense A1 sheets is exactly where the large models earn their place, and that is the entire task. Everything around the model already runs on your own infrastructure. If nothing whatsoever may leave your network, this tool cannot do this job today, and you should know that before you start rather than after.
What on-premise costs you
An isolated install drifts. Approved Document B currently carries 2026 and 2029 amendments, and BS 9991:2024 becomes mandatory for new applications on 30 September 2026. Corpus releases are versioned and shipped for exactly this reason — the alternative is learning that the guidance moved from a building control officer.
Four
What you need to do to make it work
Seven steps, start to finish. Each one notes where it tends to go wrong.
Add your API key
~2 minutes, once per companySettings → Verification key. Paste an Anthropic key and save.
- Saving runs a one-token test call before the key is stored, so an unusable key is caught here rather than part-way through a paid run.
- The key is encrypted before storage and never returned to a browser again — only a hint like sk-ant-api0…4Xk2.
- Analysis is billed to your own Anthropic account, not ours. Set a monthly spend cap in the same console while you are there.
Each company's key is scoped to that company. Two firms on the same server never see each other's keys, projects or findings.