Cities: reading a plan against the place

A plan says what a place is for and how it will change. The place goes on being used in ways the plan did not describe. This method reads the one against the other, and tests every reading at another level or by another agent.

What is already known

Planning has long read plans as texts. Mandelbaum reads a plan as a set of policy arguments, as a decision opportunity and as a story (Mandelbaum 1993). Throgmorton treats the plan as a future-oriented text written to persuade (Throgmorton 1996). Planning has also long asked how plans perform. Mastop and Faludi argue that a strategic plan should be judged by whether it guides later decisions, not by whether the built result conforms to it (Mastop and Faludi 1997). Flyvbjerg follows one plan from text to outcome and finds that power and interest explain the outcome better than the text does (Flyvbjerg 1998). Abram and Weszkalnys treat planning as a promise and ask what happens to the promise over time (Abram et al. 2013).

Holagram adds three things to this tradition: the four registers, the rule that no reading judges itself, and a place for labelled fiction and synthetic hands.

Levels, and what each admits as a Trace

Level	Name	Typical traces
5	Territory	Regional plans, statistics, statutes, maps at city scale
4	Institution	The plan text, terms of reference, consultation reports, minutes, officers’ reports
3	Site	Parcel and ownership records, permits, the plan’s own map layers
2	Street	Photographs and counts at fixed times, observation of use
1	Body	A time-stamped walking log: pace, stops, what was noticed (Piga et al. 2021; Pierce et al. 2015)

The text and the ground

The text is the planning document read as invisible literature: detailed plans, terms of reference, consultation reports, officers’ reports, permits and codes. Language models built for planning exist (Zhu et al. 2024), so authorship is a fair question to ask of any recent document. The ground is the place itself, walked at different hours, with the walking reported as the method: where, when, how long, what was looked at and what was recorded (Pierce et al. 2015).

A step between text and ground. Conformance studies compare policy text with permits and not with the place, and find implementation low (Talen 1996; Laurian et al. 2004). A useful middle step is the permit and the council decision under each plan, which gives two smaller tests, plan to permit and permit to place, that different hands can run. Fix the view of the plan in advance. Alexander and Faludi separate three views of a plan, each with its own idea of failure (Alexander and Faludi 1989). Say which applies and what departure would count as a break. When a planning office calls a general plan a framework that leaves approved building rights in force (City of Tallinn and ERR News 2023–2026), its test is whether the plan guides later plans, not whether plots match it.

Agents

Residents, users, officials, planners and models. A Finnish pilot built synthetic residents and experts inside a digital twin to test a participation platform, and presents them as a pilot aid that complements real input (Jauhiainen et al. 2026). Holagram does not use synthetic residents as evidence: synthetic agents argue tests and write labelled fiction, and never supply traces. A study of four models on two contested regeneration projects found that the framing of the prompt shifted which positions the models raised, and that positions held by local professional and academic bodies were consistently missed (Vartholomaios et al. 2026). That is where an Adversary will be weakest, and where a person should look.

Consultation records. The nearest published evidence on language models reading planning texts concerns consultation responses. In the United Kingdom a pilot with five planning authorities reported efficiency gains of about 90 per cent, tagging errors where policies overlapped, and no accuracy metrics (July 2026). The Greater London Authority authorised up to £25,000 for the tool to make the first analysis of representations on the draft London Plan, noting that it had not yet been tested on a document of that kind and that officers remain responsible for review (September 2026). In the latest Greater Cambridge consultation about 5 per cent of respondents opted out of AI summarising (MHCLG et al. 2025–2026). A Swiss multilingual evaluation reports about 90 per cent accuracy but about 70 per cent completeness on substance (Mathevet and Cailleux 2026), and a 2026 preprint treats disagreement between models as the signal and sends ambiguous items to a person (Najera et al. 2026). In Holagram, models write readings of responses and the trace stays the response itself, with its identifier and language. Divergence between models is data, and a sample goes to a person. If responses contain personal free text, ask the controller first.

A worked entry

Invented, to show the shape.

LevelInstitution (4), hands shown per register

TracehumanDetailed plan, section 3.2: “active frontage” appears eleven times, “residents” twice. Source and page recorded.

ReadinghumanThe plan is written for visitors. Residents are a side effect of the building programme.

Fictionsynthetic, edited by humanThe shops open at ten and close at eight. By nine the street is a set with its lights left on for an audience that has gone home. The people upstairs learn to live behind the lit window of an empty shop.

Testsynthetic and human, at street level (2)Counter-reading from the Adversary: the wording follows a template and says nothing about intent. Check: count occupied windows above the shops at 21:00 on two weekdays; ask the planning officer who wrote the section. Result: unchecked

Where it goes wrong

Intent read into a template. A count of words in a plan can reflect the statutory structure and not a priority. In a reading of 461 California general plans, housing topics sit apart from the rest, which the authors link to a statutory five-year update and not to a meaning (Brinkley and Stahmer 2021). Name the genre of the document before the first entry (Mandelbaum 1993).
A “broke” that misjudges the genre. If the document is a strategic plan, failing to match the ground may only show that a signpost was read as a blueprint (Mastop and Faludi 1997).
A pattern that is the observer’s. The failure of the narrator in Applied Ballardianism. The remedy is the test at another level or by another agent, and the outside reader.
The walker’s mood taken for the place. Body-level traces are the walker’s. Keep them time-stamped, and let the street-level counts be the check (Gehl and Svarre 2013).
A model of the place taken for the place. A three-dimensional model, a digital twin and a participation portal’s map are each made by someone, with a version and interests (Nochta and Oti-Sarpong 2024). Record which version each entry uses.
The office receives a test as a threat. Planning offices keep some tensions unspoken, and defensive routines are the expected reaction to being read (Mäntysalo et al. 2011). Show the office the test before publication and log its reply as an entry.
The place moves. Fix a snapshot date and keep a document log. A new document is a new entry, and an event that makes a reading moot is logged as such.

Care, and a declared interest

Attribute statements to offices and dates, not to persons. Write tests about documents. Declare in the entry header any interest the author’s organisation has: if it builds participation software, or works for a city, say so. Use models and prompts that are published. A live political decision is contested ground: strong readings are allowed, labelled, and held back until tested.

Objections the literature raises

“Plans are signposts, not predictions.” Setting text against place can misread the genre (Alexander and Faludi 1989; Mastop and Faludi 1997). The answer is to fix the view in advance, but a planning scholar may say that is plan evaluation under a new name.
“Conformance can be measured from permits without a walk.” The field has called for implementation studies since the 1990s (Talen 1996; Laurian et al. 2004). Holagram has to show what lived experience, the four registers and the failure log add beyond a conformance study, and to cite that work.
“The gap is in what officials do with responses.” A better reading of a consultation record changes nothing if the input has no route to the decision (Kahila-Tani et al. 2019). Institution-level entries therefore trace that route, or the method will be taken for an analytics tool.
“Planning officers know the sites.” They hold negotiations that never reach public documents, and public documents are the cleaned output. Three weeks of outside reading will not tell them what they do not already know. The answer has to be a result they can check and not a stance.
“Government tools are good enough to use for analysis.” Barring models from traces looks over-cautious to those who run them. The reply rests on completeness, tagging errors where policies overlap, missing published accuracy figures (MHCLG et al. 2025–2026; Mathevet and Cailleux 2026), and the fact that the benchmark is itself a human summary.
“Simulated residents test plans cheaply at scale.” Holagram’s rule forbids the one thing that scales (Zhou et al. 2024). The reply is that the metrics are the simulators’ own and that a twin is a document with interests (Nochta and Oti-Sarpong 2024), which will not satisfy someone who takes lived experience to be anecdote.

First study: one place, one document, three weeks

Choose a place that can be walked in an afternoon and one public document that plans or describes it: a detailed plan, terms of reference, a consultation report. Public material only; no client material.

Week	Work	Produces
1 · Document	Name the document’s genre. Read it as invisible literature. Pick Traces, write Readings, draft Fiction with the synthetic agent.	About ten entries at the institution level.
2 · Ground	Four walks, at different hours. Two with a synthetic companion describing photographs, two without. Time-stamp everything.	Traces at street and body level. New readings, some against week 1.
3 · Tests	Send readings to the people who wrote the document or use the place. Run the Adversary on each. Close entries as held, broke or unchecked.	The diagram, the failure log, the short case, a conventional write-up of about 2,000 words.

What would show it works. These are starting values, to adjust:

At least a third of the readings are tested at another level, and some of them break. If none break, the tests are too soft.
The conventional write-up can be written from the log alone.
The Fiction register raised at least one question about the place that Trace and Reading had not.

The study claims a case and not a pattern: one place over three weeks is thin for comparison (Pierce et al. 2015). It is also the first test of the synthetic hands protocol.

Candidate places. Facts about places change. Check each before starting, and fix a snapshot date. A first study is easier where there is a stream of documents, a place that can be walked in an afternoon and little at stake; Linnahall is the richest candidate and the most contested.

Linnahall, Tallinn. A Soviet-era waterfront hall with several texts, in different genres, about one building and its surroundings: a detailed plan from 2017, a 2025 proposal for a new plan, an investor presentation in March 2026, and press reports of the mayor’s statement in favour of demolition on 18 March 2026 and the National Heritage Board’s reply that the building is protected and must be preserved. By 9 July 2026 the city was reported to be seeking EU funds for 2028 to 2034 for a project of EUR 150 to 200 million, with a strategy centre due to report in the autumn and the city government deciding afterwards (ERR News 2017–2026). The 2017 and 2025 items were not re-checked.
Tallinn’s new general plan. As reported by ERR News, the city government sent the initiation of a citywide general plan to the council in May 2026, with completion expected in 2029. An ideas survey in Estonian, Russian and English closed on 31 August with more than 1,700 submissions reported, and the planning office calls the plan a framework that leaves approved building rights in force (City of Tallinn and ERR News 2023–2026). It gives a live stream of documents at territory and institution level.
Lasnamäe and Ülemiste. Lasnamäe is a Soviet-era panel-housing district, and work on Baltic housing estates gives its history (Hess and Tammaru 2019). Ülemiste, by the airport and the rail terminal, had an area plan initiated in May 2024 for more than 13 hectares around the shopping centre (City of Tallinn and ERR News 2010–2026). Plans are being implemented in both, so the step between text and ground has material to work on.
Helsinki. The City Plan of 2016 is in force (City of Helsinki 2016–2026). Malminkenttä, the former Malmi airfield, is an optional second place once a first study is done; its dates were not re-checked.

Candidate places are named as examples of what a first study needs, not as a proposal to study any institution or person.