Document begins

The corpus map

The public website had 22 pages plus 13 rendered research briefings. Behind it sat a Git-tracked working corpus of 138 text artefacts in two private directories, built by a handful of people and a set of AI agents working under a written evidence discipline. At the archive-preparation snapshot (18 August 2026, 101 days after the first commit) the repository stood at 605 commits, 72 issues and 54 pull requests; those figures describe that moment, not the repository as it stands when you read this, and they include the work of building this archive. The publication checkout also contained 27 ignored/untracked local text files. They were real campaign working material — and, as the section on local-only material below records, some of them are the provenance sources of published archive pages — but because they were never committed they cannot be independently reproduced from Git, so they are excluded from the verified totals rather than smuggled into a machine-verified claim.

This page is the map of that corpus. It exists because an archive of a dozen well-chosen documents tells you the group was thoughtful; it does not tell you how much there was. The interesting fact about this campaign is not any individual document. It is that a town of about 1,500 people generated a civic-acquisition operating system — epistemics, finance, legal research, property work, governance architecture, community intake, communications discipline, technical infrastructure and an agent fleet — in fourteen weeks, and then lost anyway.

How to read the dispositions

Every Git-tracked artefact in internal/ and private/ falls into exactly one of three states at the publication boundary.

Nothing here is a live document, and nothing in internal/ or private/ is exposed directly. The originals stay where they were, unedited, as the evidentiary record — a record that is mostly Git-preserved and partly local-only, as described below. Publication always meant authoring a new public copy and putting it through the scan-then-human-read boundary described in How this fossil was published.

The corpus by working area

What is withheld, and why

Withheld material is not allowed to simply disappear from this map. Every tracked withheld file must resolve to a stated class and reason before the build passes; generic catch-all reasons remain for otherwise-unclassified working notes, so the mechanism guarantees a reason string for every file, not a bespoke considered reason for each one. The classes below are what remains private, described as precisely as they can be described without undoing the withholding.

The 63 tracked withholds cluster into four honest categories, and it is worth being blunt about which is which:

• Genuinely protective — the named supporter prospect research, the who-does-what map by named person, and the respondent-level intake data. Three files, about living people who did not sign up to be in a public archive. They stay private permanently.
• Negotiation posture — the sprint-week approach material (the letter drafts to the vendor’s agent, the donor approach pack, the evidence pack, the open due-diligence questions and the sprint README), the authorised receiver-contact script, the acquisition-specific half of the lawyer brief and the adviser-engagement checklist. The withheld material preserves draft variants, donor and vendor approach sequencing and negotiation posture. A conditional expression of interest was ultimately lodged on 2 July; this archive withholds the internal drafting and strategy around that contact, not the fact that the contact occurred.
• Copyright-adjacent — two text indexes of the reference-photograph collection (the photographs themselves were never committed and are described under “What the counts do not show”). Small, and here mainly so the class is not silently empty.
• Duplicative or superseded — by far the largest group: the 18 withheld near-identical agent role contracts, the agent directory index, the 12 generated study aids, the epic and spec detail, the governance templates already derived into the steering pack, early strategy notes, working research trackers, copy workshopping, an internal privacy notice superseded by the live one, and a handful of short working notes. This material is not sensitive; most of it is already in the archive, restated more usefully in a derived page. Withholding it is an editorial call, not a protective one, and it is the category where a future reader is most entitled to be sceptical of us.

Nothing in the tracked withheld set is withheld for security. The live-system runbooks that are tracked were derived into the technical infrastructure page; the two retired-endpoint source files are implementation files outside the document count.

Local-only material outside the verified census

The 27 ignored or untracked text files in the publication checkout are real campaign working material, but they are not part of the 138 Git-verifiable corpus and the Git-based census cannot independently reproduce or vouch for them. They include the per-person dependency map, two API guides for third-party services, four single-purpose agent prompts, per-agent operating notes, raw source dumps behind two early documents, a scraped vendor-documentation cache, four research JSON files, the local claim register, and a few superseded plans and notes. Some of them are the provenance sources of published derivatives: two of the task prompts reproduced verbatim on the constraint architecture page, the two API guides behind the technical infrastructure page, and the per-agent operating notes behind the agent ecology page all live only in that working checkout. So the evidentiary record behind this archive is partly Git-preserved and partly local-only. That is a less flattering thing to say than “the whole corpus is machine-accounted-for”, and it is the more interesting one: it marks the boundary of reproducibility itself, and it was found by the release gate rather than declared in advance.

What the counts do not show

Three further things sit outside the file count and are worth naming, because they are a large fraction of the actual work.

Binary artefacts, none of them ever committed to Git. Four purchased title-registry documents (two URDS reports, two council certificates), six archival and contemporary photographs of the building and the town, five generated slide decks and three generated audio briefings. (The two flashcard sets from the same tool are text files and are inside the counts above, not here.) Roughly 150 MB, almost all of it the generated study material described below. None is published: the registry documents and photographs for copyright and privacy, the generated material because it is derivative of documents that are published.

The public site itself. 22 public pages and 13 rendered briefings, built from 85 tracked documents under docs/, 41 scripts and 22 Cloudflare Functions. That is the part that was always public, and it is still at bottom.pub. The archive is not a substitute for it.

Everything that happened in rooms. Meetings, doorstep conversations, phone calls to advisers, the community meeting, the conversations that produced the working group. The repository is the instrument panel, not the campaign — see What actually happened for the events these documents were made in response to.

A caution about the shape of this map

Roughly half the corpus is published or derived. The distribution across areas is uneven, for reasons that are themselves part of the record.

Feasibility and finance is fully covered; the acquisition sprint barely is. Every artefact in the finance area is published, because the group had already written it to be checked. Almost nothing from the sprint week is, because that is where the negotiation posture lives — the proposal draft, the donor pack, the approach sequencing. The critical path is published and the mechanism is described; what the group would have said to whom, and in what order, is not.

Legal, licensing and property is one of the best-covered areas, and most of its tracked material is published or derived. The pattern there is that the research — statutes, regulator guidance, licensing pathway, heritage constraints, the title search method — goes out, and the posture around it does not. The one document that needed full re-authoring rather than redaction was the title search, because it identified individual creditors and speculated about their finances.

Agents, prompts and specs remains the largest single area and the most withheld. That is a deliberate editorial position rather than a protective one. Publishing 19 near-identical role contracts would demonstrate volume and teach nothing; the agent ecology page derives what those contracts were for, how they were constrained and where they failed, and the constraint architecture page publishes six representative constraint artefacts — the base contract, shared prompt fragments, one role overlay and task prompts, not six of the nineteen role contracts. Readers who think that is a dodge are entitled to; the raw contracts are not secret so much as repetitive.

Evidence is heavily withheld, but much of what is missing is derivative. Twelve of the thirteen withheld files in the tracked evidence area are generated study aids — reports, tables and flashcards — produced from documents that are themselves published here; the thirteenth is a short case-study verification working note that was not separately published. Publishing a summary beside its source adds a less reliable version of the same thing. The NotebookLM experiment describes what they were and why.

Epistemics is over-represented in the published set, because that material was designed from the start to be defensible in public. That is a real property of the work, but it also flatters us: it is easier to publish the part you wrote carefully.

The honest summary is this: a small group in a small town built far more machinery than anyone outside it ever saw, most of the machinery was ordinary and much of it was never used, the group lost anyway, and the reasoning is still worth reading. That is the whole claim of this archive.