Documents and the planning text

How to read a planning document as a text and tie it to the ground.

The literature on documents gives Holagram a method for its first register more readily than for the other three. Anthropology and science studies treat a document as a thing with an author, an audience, a history and a route through an institution (Asdal and Reinertsen, Hull, Riles). Planning theory adds that plans are persuasive texts (Throgmorton), that they can be read as policy argument, decision opportunity and story (Mandelbaum), and that their promises are the part least often held to account (Abram and Weszkalnys, Mathur). Latour's account of circulating reference is the closest existing model for the Trace: a claim is tied to ground by a chain of steps that others can inspect. The same material threatens the project in three places. Plans are often promises or signposts, so a Test that sets text against street may ask the wrong question (Mastop and Faludi). Reading one PDF as a self-contained text ignores the process behind it. And the planning text may now be partly synthetic: by its authors' account PlanGPT is applied by Chinese planning bodies, and a 2026 test of four language models on two Greek masterplans found that prompt framing changed what they reported and that local professional bodies' positions were consistently missed. The thread found little that ties a planning document to a walked place by method; that gap is where the first study would be new. Four unconventional items (Latour's Aramis, Haacke, Rosi, the Atlas Group) show documents and fiction combined, with and without labels.

Sources

Ideas to take

Open each document study with a one-page document sheet before any passage is excerpted. It records where the document was made and for whom (site, implied author and reader), the tools and templates that produced it, what work it does in the process, who commented on it, what it was attached to or sent on to, and the file version and format.

Changes: Adds a step before the Trace register and stops the first study treating one PDF as a self-contained text. A Trace then carries a document identity (file, version, date, site) as well as a passage, and the sheet is a document in its own right with an author and a date.

Risk: Becomes a second research project on the document's history and eats the three weeks. Cap it at one page and one afternoon, and let it grow only if an entry needs it.

Use Mandelbaum's three readings (policy argument, decision opportunity, story) as modes of the Reading register. Each Reading names its mode. A strong or paranoid reading is allowed, but it is filed under story by default and has to survive a plain reading of the same passage first.

Changes: Gives the Reading register a vocabulary taken from planning practice and gives the Test a different default depending on the mode: an argument is tested by its premises, a decision opportunity by what was decided, a story by what the place shows.

Risk: The modes come from one US downtown plan and may fit a modern GIS-based Estonian or Finnish plan badly. They may also make readings look neater than they are.

Adopt a chain-of-reference rule for the Trace. An entry records the steps from a passage in the document to a thing on the street and back (passage, map layer, parcel or street, dated photograph with position), each step inspectable. An entry with a missing step is marked unchained and cannot be given the verdict 'held'. A Trace is defined as a thing someone has logged as evidence, with source, date and hand.

Changes: Turns the source rule into a testable procedure, supplies the vocabulary for what is lost at each step (where the lived place goes), and lets the diagram show where chains break. Pair it with a classification of claims as descriptive, normative or promissory, because only descriptive claims can be chained.

Risk: Planning documents are full of normative and promissory claims, for which there is no soil to sample. Applied rigidly the rule would either exclude most of the document or push readers to pick checkable claims because they are checkable, not because they matter.

Keep a promise log next to the failure log. Descriptive claims get a Test within the fieldwork block. Promissory claims go into a dated log with a statement of what 'kept' would look like, who made the promise, and when to look again; their verdict stays unchecked until then. Mathur's two columns, the paper life and the real life, can head the log.

Changes: Separates what a document says is so from what it says will be, matches the Test register to the kind of claim, and gives Holagram a way to use a document in a three-week study without pretending to evaluate its outcomes.

Risk: Few promises resolve within three weeks, so the log may fill with unchecked items and weaken the claim that the method tests anything. Mastop and Faludi also imply that a signpost-type plan may be performing as intended while failing a conformance Test.

Add a rule that the first Reading of any passage reads it along the grain: record the document's templates, mandated sections, repeated phrases, categories and signatures, and a strong reading must explain a sign better than the genre does. Use Brinkley and Stahmer's housing result as the worked example of a statutory form mimicking an emphasis.

Changes: Builds a safeguard against the self-confirming reading that Sellars's narrator falls into, using document genre as the plainest competing explanation. It gives the Adversary a first move.

Risk: Slows the work and may dull the paranoid register that is the point of the style. A statutory explanation can also be used to dismiss a sign that does matter.

Treat synthetic readers as logged instruments. Run the document through several models under different framings (minimal, impartial, sceptical, advocating), keep prompts and versions, and file the outputs as Traces of the model's behaviour, never of the plan or the place. Compare what a model reports with what the walk shows, and where a model misses local positions, use that as a map of what only a person on site holds. Use topic modelling against a peer corpus (plans under the same legal regime) for the absence question.

Changes: Gives the synthetic hands a bounded role in the Trace register without breaking the rule that they never supply Traces, and turns the known weakness of language models with local knowledge into a design input.

Risk: The rule needs rewording so that machine counts and model outputs are allowed as logged instruments under human responsibility. Models change between runs, and the Greek study calls its own evidence inconclusive.

Make the project's own documents subject to the same sheet. Each entry carries an author, date, audience and hands line; registers are kept visually distinct as Latour does in Aramis; and a public-roles-only rule keeps named private owners out of entries, with Haacke as the case for why.

Changes: Answers Riles's point that the researcher is also a document-maker, supports the condensed-novel format with a published model, and gives the care rules a concrete legal edge.

Risk: More apparatus than a three-week study can carry, and a public-roles-only rule may cut out the entries that matter most to a reader.

Give Fiction a label that survives copying: a made-on date, the in-world date, the hands, and a fiction stamp that travels with any excerpt. Rosi's opening caption and Raad's two dates are two precedents; Raad's archive also shows how a labelled fiction can still pass as record.

Changes: Turns 'labelled fiction, never evidence' from a rule into a format, and gives the outside reader something to check.

Risk: Stamps get cropped in circulation, and fiction written in the voice of a planning document can be mistaken for the document however it is labelled.

Counter-arguments

Gaps