11 checked categories
It must retrieve, or refuse.
The failure mode of every AI writing tool is the same. Asked for a policy reference, a model will produce something that looks exactly like a policy reference. In planning that is not a rounding error, because applications are decided on paragraph numbers.
The rule
For every checked category the drafter must retrieve from a verified source or refuse to write. A citation that cannot be resolved against its source at the moment it is written is a hard error. Generation fails, nothing is saved, and the client is not charged.
It is written into the codebase as a numbered rule, it applies to every draft on both paths, and there is no way to switch it off.
Checked
The 11 categories
Each one has a source of record, and a specific thing that is checked against it.
01
Statutory text
legislation.gov.ukThe Act, section or Class exists, and its wording matches the retrieved text.
02
Framework paragraphs and guidance
gov.uk policyThe paragraph or guidance reference exists in the version currently in force.
03
Listed building citations
National Heritage List for EnglandThe list entry resolves to a live record, and the grade and the name both match.
04
Case law
Find Case Law, The National ArchivesThe neutral citation resolves to a judgment in the index.
05
Flood zone and climate allowances
Environment AgencyThe zone comes from the flood map for planning. Rainfall and river-flow allowance taxonomies are never crossed.
06
Comparable applications
PlanItEvery comparable references an application actually retrieved for this site's radius.
07
Local plan policy
The council's own adopted planThe policy number resolves for this authority, and what the draft says about it is grounded in the retrieved text.
08
Validation requirements
The council's own published listGenerated from that authority's list, never from a national baseline.
09
Supplementary planning documents
The ingested corpusEvery citation traces back to a document actually retrieved.
10
Conservation area appraisals
The ingested corpusEvery citation draws from an appraisal actually retrieved.
11
Design guides
The ingested corpusEvery citation traces back to a guide actually retrieved.
Not just existence
Checking a policy number is not enough.
A model can cite a real Policy 13 and then paraphrase what it plausibly says. That is fabrication in a subtler form, and it is harder to catch.
So every characterisation of a local policy goes through a second test. If the draft block-quotes a span that is actually present in the source, it passes with no further work. Otherwise a strict judge decides whether every substantive assertion the draft makes about the source is supported by the source text. General planning knowledge is explicitly irrelevant to that judgement. Only the retrieved text counts.
An unsupported claim fails the draft. A judge that cannot be reached fails the draft. It never quietly passes.
There is one deliberate middle case. Where the source is real and was retrieved, but the judge could not support the draft’s description of it, the draft is not thrown away. It becomes a query to the chartered planner marked as a matter of judgement, and it is logged alongside the verified retrieval. That is the right place for it: it is a professional disagreement about characterisation, not a fabricated source.
Errors caught
Things that are wrong and look right
Each of these is a specific mistake the checks are built to catch, and each has been made by something.
The wrong Act
Sections 66 and 72, the heritage duties, live in the Planning (Listed Buildings and Conservation Areas) Act 1990. Not the Town and Country Planning Act 1990 of the same year. A heritage duty cited against the wrong Act is a fabrication, and the correct mapping is encoded rather than trusted.
Crossed climate taxonomies
Rainfall allowances use one set of band names over one set of epochs. River flow allowances use different names over different epochs. Higher Central is a river-flow band only, and a rainfall figure quoted under it is rejected.
A number that drifts between sections
One canonical framework paragraph per topic, fixed. For flood risk: the sequential test is one paragraph, the exception test another, the site-specific assessment a third, and they never swap between sections of the same document.
A policy number cited bare
Every sentence citing a policy must, in that same sentence, state the requirement it is cited for. A reference with no proposition attached is decoration.
Guidance references recalled from memory
Only a small set of guidance topics is verifiable, and a slightly wrong reference fails the whole draft. So guidance is cited by topic name instead.
A search that failed, reported as a search that found nothing
Planning history has three separate states: found records, searched and found none, and could not search. The third used to be flattened into the second, which put a false statement into a signed document.
Recorded
Every check is written down, whether it passed or not.
Which is what makes any of this checkable rather than merely claimed.
Every citation, verified or refused, writes a row: which source was consulted, what was cited, the URL it was resolved against, whether it passed, why not if it did not, and the timestamp. Every verified passage is additionally stored with its own hash, so a citation can be re-verified after the fact against the exact text that was retrieved that day.
When a source changes wording, a new version is written and the old one is demoted. Nothing is overwritten, so the record of what a policy said on the day your document was drafted survives the policy being rewritten.
That trail is what an AI Use Disclosure is built from, and it is why the disclosure can name the actual version of the framework in force on the day rather than a generic statement that AI was involved.
Also enforced
Never draft for the wrong site
The typed address is geocoded and compared to the drawn boundary. Too far apart and the draft is refused, because every constraint and every figure would be for somewhere else.
Never draft an unpaid case
Payment is asserted at the single point every drafting path passes through, rather than in each of the six things that can start one.
Never draft forever
Three attempts per document, counted regardless of outcome and never reset by a success. Only an administrator can clear it, because clearing it commits real money.
Never invent a figure
A figure is emitted only when its real image was produced. An unavailable image means the figure is dropped, never replaced by a placeholder box.
Watched
And the sources are watched for change.
A verified citation is only true until the source moves.
The national corpus, the framework, the guidance, the permitted development order, the use classes order and the fee schedule, is checked every day. Local plans are checked on a weekly cycle, taken oldest first rather than all at once. When the text changes, the difference is examined and a judgement is made about whether it matters. Most changes do not: a council’s content system re-rendering the same document is not news. A council renumbering its policies is, because every draft citing the old numbers is now wrong.
One category cannot be caught by watching content at all. Some sources are replaced every year at a new address rather than revised in place. The old document keeps returning perfectly healthy responses forever, so a content check reports it as fine precisely because it is stale. Those are checked against the calendar instead, and they are counted separately from the healthy ones.
Start here
Then a chartered planner reads all of it anyway.
The checks decide what may be written. A professional decides whether it is any good, and puts their name to that decision.