The reference shelf

How we verify — the full method

Every claim on this site that carries a figure, a fact, or a quote is supposed to earn that place, not just occupy it. This is the shape of how that happens — the method, not the case files.

The loop

Publication here follows a fixed order, not a shortcut:

1. What we owe. Before anything is written, the questions a page is meant to answer are written down first. What a page owes its reader is decided before drafting starts, not backfilled after.
2. Drafting against sources. A draft is built directly against real, citable material — never written first and sourced later.
3. Backwards verification. This is the hard gate. Before anything ships, every load-bearing claim in the draft is pulled back out, one at a time, and checked independently against its original source. That check never relies on the draft’s own summary of the source, and never trusts a first pass at face value. If a source can’t be reached or confirmed, the claim is fixed, hedged, or removed — it does not ship unresolved.
4. A record that survives the source moving or disappearing later. When a source is checked, a copy of what was actually read is kept alongside the claim, so a claim’s backing doesn’t quietly rot if the original page later changes or vanishes.
5. Plain-language release. Only after that verification pass does a claim become eligible to appear in a public, plain-language page — never the other way around.
6. Corrections propagate everywhere, automatically. If a claim is later found wrong or disputed, every place on this site that repeats it gets checked and updated — found by a systematic sweep across the whole site, not by relying on anyone’s memory of where it was used.

Independence is part of the design

Checking is never done only by whoever drafted the claim. A second, independent pass re-verifies sourced material before it’s trusted for publication — the same discipline a careful editor would apply to any writer, regardless of who or what produced the first draft. Being “checked once” and being “verified” are treated as two different things on purpose.

What our status words actually mean

You’ll see a small, consistent set of words across this site. Here is exactly what each one promises:

• Sealed — every load-bearing claim in this page has been independently checked against its original source, with zero unresolved disagreements found. This is our highest bar, not a marketing word.
• In verification — drafted against real sources and going through the same independent check described above right now. Some pages carry one or two specific open items — a source that wouldn’t load, a figure not yet confirmed a second way — and we say so rather than rounding up.
• Draft — written, not yet through the verification pass. Treat anything marked draft as a starting point, not a finished claim.
• Proposed / coming — planned or scoped, not built or drafted yet.

We would rather show you an honest “in verification” than a premature “sealed.” A visible drop in how much of the site is sealed, if that ever happens because something failed a re-check, is treated as the system working, not something to hide.

What happens when something turns out to be wrong

The short version lives on the corrections door. The relevant point here is that a correction is never a quiet edit. Once a claim is confirmed wrong or no longer solid, every page repeating it is found and updated together, the change is dated, and the old version stays visible with its history — not deleted, replaced.

See the method applied, end to end

The clearest worked example on this site: the “check our work” page of Why can’t Toronto move? — a claim-by-claim claims register with its receipts, the first-party measurement the report rests on, the hostile questions we asked our own numbers before publishing, and the dated re-check schedule. That page is what the method above looks like when a full report goes through it. How the research itself is made, from question to library: the research process.

What this section is not

This is a description of method, not a log of specific findings, disputes, or any individual claim’s history — those live on the pages the claims actually appear on, and in the public corrections record. It also isn’t a description of the specific tools or internal workflow used to do any of this; that’s an implementation detail that can change without changing the promise. The promise is the point: a claim either traces to a real, checked source, or it says so on its face.

What our measurement sees — the complete accounting

Every page on this site loads one small, aggregate-only visit counter — Cloudflare Web Analytics. What it reports, pooled across every visitor: which pages get read, what site referred you here, your country, and your device/browser/OS class. Per Cloudflare’s own documentation, it does not collect or use visitors’ personal data, does not log query strings, and does not track individual people across sites — and we set no cookie and store nothing on your device ourselves; the counter is the only script we run. Raw beacon data is kept seven days, then aggregated down and the detail is gone. There is no dashboard anywhere that shows you as a person — only counts. If that ever stops being true, that’s a correction, filed the same way as any other: the corrections door.

The full invitation — corrections, ideas, hands

This is an open invitation, and we mean it as loudly as a quiet library can. Everything on this site is a working draft of something bigger, built in public — and the fastest way it gets better is you telling us something we don’t know yet.

What we’re asking for

All of it is welcome. Any of these makes the site better the day it arrives:

• Corrections — a claim that’s wrong, a source that doesn’t hold, a figure that’s out of date. Include the claim (its ID if you have it — every figure on this site shows one on hover), what’s wrong, and the better source if you have one. No better source? Send it anyway; “this doesn’t hold” with a reason is plenty. You don’t need credentials. The evidence needs credentials.
• Improvements — a clearer way to say something, a page that’s confusing, a path through the site that dead-ends. If it made you squint, tell us.
• Ideas — something we should build, cover, ask, or try. The project grows by exactly this route.
• Resources — a dataset, a report, a person’s published work, a tool, a precedent from another city. Receipts are our whole currency; bring us more.
• Hands — if you want to help build, verify, translate, or organize, say so. There is real work here and it is genuinely open.

Send it

Two doors, both watched: the Harvest — the site’s front door for ideas, offers, and what you’d fix — and the mailbox, for corrections directly: unknownsoldier@unitetolove.ca. The one thing we really need is which page you’re writing about, so please paste the page’s address in.

What happens to your note

It gets read — a person reads every note; nothing vanishes into an unwatched inbox. It gets verified — we check it against sources, with the same discipline we apply to our own claims, and then one of three things happens in the open: we fix it (the claim changes everywhere it appears, the old version stays visible with a dated correction), we show our work (if the claim holds, we publish why, with sources, so the next person with your doubt finds the answer waiting), or we downgrade it (sometimes the honest answer is “we’re no longer sure,” and the claim’s status changes to say exactly that). And you get credited, if you want credit — tell us the name to use, or stay anonymous; both are honored.

What never happens: silent edits, quiet deletions, or a claim that goes on being cited after it broke. Fixing things fast and logging the fix in the open isn’t damage control here — it’s the product working as designed.

The record to date

Corrections are logged, dated, and left visible on the pages they touched. As the record grows, this section becomes its running index — every correction filed through the door, what changed, and who caught it (with their permission). An empty index today means the door is young, not that the site is flawless; somewhere on this site right now there’s an error nobody has found yet. There always is. We built the whole system around that sentence staying permanently true.

If it’s serious

If you believe you’ve found something that doesn’t just weaken a claim but points at real wrongdoing by real institutions — stop. Don’t send it here first. Serious allegations deserve journalists, auditors, or authorities with the power and the duty to investigate them properly; we are neither, and we don’t play at being either. This library documents what’s published and verifiable. It is not a tip line.

A standing note on tone

Blunt is fine. You don’t have to wrap a correction in compliments — a correct stranger is worth a hundred agreeable ones, and we’d rather build with the first group. Bring what you’ve got; we’ll bring the gratitude and the follow-through. That’s the trade. This page isn’t fine print. It’s a help-wanted sign, and it’s always up.

The whole story — who runs this, how, and what it actually cost

Updated July 28, 2026. An earlier version of this account held a visible placeholder, noting we hadn’t yet decided whether to name the specific AI systems and vendors used. That call is made, in three words: be completely honest. What follows is one story, told straight — the person, the controls, the money, and the month. All figures in Canadian dollars.

Start with the person, because everything else hangs off him. One human being directs, orchestrates, and controls all of it. To borrow from a certain gravel-voiced movie dad: he has a very particular set of skills. Skills acquired over a very long career. Not those skills — his are three-plus decades of living on the internet: finding things, reading things, checking things, building things, and refusing to log off. If a forgotten municipal report exists, he will find it, and he will cite it.

And the machines answer to him — not loosely, but by written constitution. Every instruction he gives is captured word-for-word in a permanent log, and an AI session’s first duty is to read those words back correctly before acting on them. Every ruling he makes becomes a numbered decision on a permanent record. Every substantive edit preserves a snapshot of the version before it, so the full history of every document stays inspectable forever. Every load-bearing claim needs a receipt. Every page faces adversarial review — hostile-reader personas hunting for the ways it could mislead — before it’s called ready. Automated checks sweep every batch of work for leaks and drift. And nothing goes public without his hand on the button. The machines here are brilliant and fast, and they are on a leash held by someone who has been on the internet longer than some of their training data.

Think of it this way: if a human researcher with a brilliant reading speed and an occasional habit of misremembering worked here, you’d want every one of their claims checked against sources before publication. That is literally our process. The species of the researcher was never the load-bearing part.

Now the money, because it is the most instructive part. This person had not paid for a single AI tool in over two years. Then, on June 27, 2026, a friend lent him a Claude account — and bought him his own month of it as a gift. Everything on this site began in the thirty-one days that followed. The complete paid-AI claims register, as of July 28:

Committed in month one: about CAD $1,030, of which $140 was the gift. But here is the number that matters more, because most of those payments landed late in the month: pro-rate each subscription by the days actually used by July 28 — and the AI actually consumed came to roughly CAD $330. A few hundred dollars. The rest was capacity for the month ahead, already bought. Everything you can read on this site — every page, every receipt, every guide — cost a few hundred dollars of machine time, directed by one unpaid human. This claims register updates as the story does; the dates on it tell you how current it is.

Around the paid seats, a rotating free bench did the bulk reading and cross-checking — a dozen or so free tiers from every major lab, each linked below with the date we last checked its terms, because free tiers change monthly and a stale claim about somebody else’s pricing is still a stale claim. Plus one already-owned gaming PC running free open-weight models around the clock as a tireless local research assistant. The paid seats do the judgment-heavy work: drafting, verification passes, orchestration. Kimi is the deliberately different outside eye — a second family of models, so no single vendor’s blind spots become ours. The cheap and free tiers haul the bulk. Every tool’s output passes through the same receipts discipline — the vendor was never the load-bearing part either.

(Free tiers as checked July 28, 2026 — they change often; each link is the living source): Qwen (free chat app, Qwen3.7-Max/Plus previews), Z.ai (GLM-4.7-Flash), DeepSeek (V4-family chat), ChatGPT (GPT-5.5 Instant), Gemini (3.6 Flash, with a daily taste of 3.1 Pro), Perplexity (Sonar), Google’s Antigravity preview (rate-limited Gemini 3.1 Pro, Claude Sonnet/Opus 4.6, GPT-OSS 120B), Vibe by Mistral (Medium 3.5), Meta AI (Llama 4), Grok (light free tier), Suno (v4.5-all, for music), and Microsoft Copilot (GPT-5.1 Smart mode).

Could we have done this without AI? A version of it — smaller, slower, and honestly, years too late to matter. A handful of unpaid neighbours cannot read a hundred thousand pages of municipal reports, audits, and studies. With these tools, they can — and the whole point of this project is that the tools of institutional-scale knowledge finally fit in a community’s hands. It would be strange to build a movement about that and be embarrassed by it. We’re not embarrassed. We’re careful — which is different, and harder, and documented.

What did thirty-one days and a few hundred consumed dollars produce? A civic research library of 666 live pages — backgrounders, briefs, cards, and day-one implementation pages across every major issue, each carrying its verification status on its face. A living atlas of 765 free-help entries across 54 categories, from crisis lines to free skating. A plain-language how-to-vote guide checked against the election regulation itself — a check that caught the City’s own published ID list omitting a document the regulation accepts, now flagged publicly. A candidate’s handbook for anyone deciding whether to stand in October. The verification machinery under all of it. And since that count was taken: the conversation itself opened — the door where a neighbour signs up, the standing offer that brings the author to any room in the city that invites him, and a poster pack any neighbour can print at home, every line of it traceable back to this same library. If you like numbers, the auto-generated tally that day: 5,950 files holding just over a million lines of text and code, 3,386 research and strategy documents, more than 150,000 rows and entries of structured data, 41,167 distinct external sources consulted, 93 numbered decisions, 112 logged working sessions, 74 written work orders between AI sessions — counted by a script precisely so the legend can’t outgrow the claims register, and re-run at launch.

The honest asterisks, before anyone else finds them: a friend’s gift started it; the free tiers are subsidized by their vendors; the gaming PC was already owned; quality varies by page and every page says so on its face — that is what the statuses are for. And none of it counts the human’s own time, which was enormous and unpaid.

You’ll meet people who say AI involvement discredits all of this automatically. Ask them the only question that matters: is the claim true, and how would we know? We’ve made that checkable. Most of what you’re asked to believe every day — by institutions with far bigger budgets than ours — has not.

Here is the message inside the numbers. One skilled person — unpaid, anonymous, directing machines with a written constitution and a few hundred dollars of compute — can now stand up a receipted civic research library at a scale that used to need an institute and a budget with commas in it. The machines didn’t bring the values, the questions, or the stubbornness. They made one stubborn person’s month go institutionally far. Now imagine a thousand of us, in every city, at this standard, with receipts. That is the actual invitation of this page.

Judge the work. Check the receipts. It’s the same invitation as ever. It doesn’t change with the typist. The door is on the home page. It takes thirty seconds.