The working documents, preserved as found.

A computed inventory of the Git-tracked private working corpus behind a 35-page public website — every reproducible artefact classified by working area and by what happened to it at the publication boundary. The archive you are reading is the visible part; this is the shape of the whole.

The repository was an instrument panel, not the campaign. This is the real-world sequence — meetings, a receivership, media, a listing with a three-week fuse, an offer, a loss — with the documents each event produced set alongside it.

The publication boundary, described by the people who built it — the twelve never-publish categories, the fifteen-rule scanner that cannot approve anything, the rule-scoped waivers, the invariant that stops a preview claiming to be published, and the error every automated gate missed.

Thirty tagged claims — what the campaign thought it knew, how sure it was, and what would break if it was wrong. The most transferable thing the project built.

Eleven flagged phrases and three priority fixes: what an agent found when told to attack the project's own finance language for leakage risk.

Fourteen risks, one of which came true. The register that never got its 'materialised' row filled in.

The rulebook underneath every register in this archive — the five claim statuses, how a claim entered the record, what it took to close a research question, and the explicit list of things desk research could not settle.

A cross-stream audit of the project's own ignorance: the securities exposure, the unverified valuation and capex, the licensing pathway, the insurance nobody had scoped, and the volunteer capacity nobody had counted.

The first triage of the precedents: 'Remove until proven', 'Use only as broad inspiration', 'Verify before publication'.

The report that called the original operating model's economics 'spreadsheet astrology' and debunked its flagship precedent.

Fifty-seven precedent claims taken apart: 34 held, 5 were debunked, and two Cygnet hypotheticals had drifted into being read as Castlemaine facts.

A short history of the Commercial Hotel: who built it, what it was called, how it changed, and why a Victorian Georgian pub ended up on the heritage register.

Fifty-four documents loaded into a research tool that generated briefing papers, study guides, data tables, slide decks and three AI-hosted audio discussions about a community pub campaign. Useful for onboarding, dangerous as a source.

The whole viability question in one document — property, market, technical, legal and financial feasibility — written eight days before the deadline, with every load-bearing figure tagged FACT, REPORTED, ESTIMATE, CHOICE or STAGE 2, and a preliminary conclusion that refuses to conclude.

Every number the project ever used, with its source, its confidence label and what breaks if it is wrong. Eight hundred lines of it — and its own adversarial review flagged eleven phrasings as extractable, which were never fixed.

Where the money would have had to come from: community shares, community loans, philanthropy, debt and grants, sized against a purchase price nobody knew. Explicitly a scaffold, never a model.

Every grant, fund and social-finance route the project could find that a community pub might plausibly use — with eligibility read carefully enough to rule most of them out, and the blockers that made the rest premature.

A hospitality market assessment of a town of about 1,500 people: the other venues, the cider and wine businesses, festival and tourism seasonality, and whether the town can support a third pub run on community capital.

The same experiment on the tax and accounting side — GST, DGR status, income-tax treatment of a co-operative, the mutuality principle, and what any of it means for a community pub. A simulation, labelled as one on every page.

Which co-operative form, whether it can hold a liquor licence, where the securities boundary sits, and what an interim entity could lawfully do. The research that got as far as desk research can get, and stops precisely where legal advice has to start.

A short standing note tracking Tasmanian liquor-licensing reform, because the answer to 'can a co-operative hold the licence?' depended on law that was itself changing.

What it actually means to own a permanently registered heritage pub — which works need approval, which are exempt, what a Conservation Management Plan is for, and why nobody could scope the building work without getting inside.

For a month the group had a rumour and no listing. This is the systematic check — every commercial portal, every hotel broker, the insolvency notices register — that established what was actually knowable.

Ten days after hearing the pub might sell, the group bought a title search. It reframed the sale from a clean receivership disposal into a multi-party creditor problem — and it is the clearest example in the archive of desk research paying for itself.

An AI was asked to write the advice a careful Tasmanian solicitor might give to the project's brief — explicitly as a simulation, explicitly not to be relied on — so the group could find the gaps in its own questions before paying for an hour of the real thing.

A single register of every unresolved legal question across licensing, structure and property — the honest list of what the project did not know when it ran out of time.

The primary-source trail behind the licensing research: statutes, sections, regulator guidance and reform material, with current law kept separate from proposals.

The same trail for the structure question: distributing versus non-distributing co-operatives, active membership, the securities boundary, and the interim-entity options.

Seven streams, their scopes, their outputs and the boundaries between them — the organisation the project designed before it had a legal entity or a single elected officer.

How meetings were called, how decisions got recorded, how conflicts were declared, and how an urgent decision could be made between meetings without anyone later arguing about whether it was made.

Five governance instruments written for a committee that did not exist yet: a first-meeting agenda, a conflict-of-interest declaration, a decision log, a risk register and a source register. Lift-and-reuse infrastructure for the next community group.

A community pub that could not be entered by everyone in the community would have failed on its own terms. The policy written to stop that happening, including the safeguarding obligations nobody enjoys writing.

Every use the building and site could be put to — core pub, kitchen, accommodation, events, garden, multipurpose space — with indicative revenue contribution and the constraints on each. Imagination mapped; viability deliberately not yet assessed.

The three ways a community-owned pub can be operated, compared on control, risk, capital, licence-holding and mission integrity — and why the choice could not be made until the legal structure was settled.

Award classifications, penalty rates, a minimum viable opening team and the wage bill underneath it. The least romantic document in the archive and one of the most important.

Four hundred lines on the unglamorous half of community organising: how an expression of interest was received, acknowledged, triaged, stored, consented and followed up, and how to do that without losing anyone or misusing their data.

From 'stay on the list' to 'join the board' — the deliberate design of a participation path, built so that people who could not contribute capital were not sorted into a lesser category.

What the community's expressed capacity looked like in aggregate five weeks after the town meeting — and the deduplication and consent lessons that came with it. A clean public copy; the respondent-level source stays private.

How to ask people with money for help without turning neighbours into targets: tiers of contribution beyond cash, public-record research only, and a hard rule against approaching anyone through a private relationship.

The practical half of the strategy: what to send, what to say, what to never say, and the disclosure rules for a group that had no legal entity and could not accept money.

Twenty-three decisions nobody had made yet, wired together by what blocked what. Published with the graph source and the register intact.

The steering committee's decision log as it exists in the repository — a governance template with an empty table. Published honestly as the mechanism that was designed, not as evidence of decisions recorded.

A cross-link of every internal document to its stream, its open questions and what needed a human — written on 18 May, marked DEPRECATED on 19 May, and never updated again. The thing that replaced it froze too.

The sprint plan written with eight days left on the clock. The dependency graph, the phase gates and the five ways it could fail — with the negotiating posture and the people taken out.

Twenty-four questions the project knew it had to answer before anyone could responsibly buy a pub. Most were still open at the end.

How the project broke a flawed 40-page proposal into seven bounded workstreams, each with a problem statement, tasks and acceptance criteria — and the board that tracked them, frozen at its last update.

The rulebook turned into a grep list: forbidden phrases, safe rewrites, and the compliant-usage carve-out that stopped the checklist from flagging its own safety language.

A binding, one-page standard for everyone speaking on behalf of the project: the phrases that were forbidden, the phrases to use instead, and what to say when someone asked a question the project could not answer.

Audience by audience — locals who drink there, locals who never did, the diaspora, the sceptics — what each needed to hear, where they'd hear it, and what the project was not allowed to say to any of them.

A heterogeneous fleet of AI systems did substantial research, drafting and review on this project. This is what each was actually used for, where they caught each other, where they were confidently wrong, and the constraint architecture that made them useful rather than dangerous.

Six representative agent contracts and task prompts, reproduced exactly as they were written. This is the layer that made a fleet of unreliable systems produce a defensible corpus — and it is much shorter and blunter than people expect.

The project built automation to answer "what is the current state?" and "has the property been listed?". It ran for three months, 91 checks, and never once reported a change — including on the day the pub was listed for sale. This is why.

Day by day, what changed and why — written as prose rather than commit messages. The single best record of how the project's thinking moved, including the days it talked itself out of something.

Seven streams of work broken into numbered tasks with dependencies, written from the first community meeting. A map of the whole job, most of which was never done.

A 22-page site with 13 rendered briefings, a strict content-security policy, three Worker deployments, a database, Zero Trust on two protected areas, and a build that refused to ship copy the project had banned. Described, not dumped — the runbooks stay private.

The working group's own account of why a community co-operative could not buy a pub from a mortgagee in possession — four reasons, none of them about enthusiasm — and what it would tell the next town that tries.