The Lakehead
Reconstruction
A complete genealogical digital twin of every person who lived in Thunder Bay, Fort William and Port Arthur before 1900 — rebuilt from census, newspaper and parish record, one verified node at a time.
Bring an entire vanished city back into the record.
Before it was Thunder Bay, the head of Lake Superior was two scrappy railway towns and a scatter of townships — Prince Arthur's Landing, Fort William, Shuniah, Neebing. Between the fur trade and the turn of the century, perhaps twenty-five thousand people passed through: railway navvies and Silver Islet miners, merchants and clergy, the families who stayed and the transients who left only a single line of ink behind.
Almost none of them have a page anywhere. The goal here is total: not the famous handful, but everyone — every name the census caught, every name the newspapers mentioned in passing, stitched into one connected, sourced, navigable record of a community.
We start from the census because it is the only document that tried to count everybody. From those rows we generate a stub for each person, resolve the same individual across decades into a single node, and only then turn to the newspapers — births, marriages, obituaries, business notices — to give each life its detail.
The output lives on Curriepedia, a genealogy wiki built over nearly a decade with Thunder Bay as its anchor. Every person becomes an article; every article links to its sources; the web of links is the reconstructed city.
A proof of concept for keeping the internet human.
The web is filling with AI agents that sign up, post, vote, apply and transact while wearing a human mask. The endgame is the infinite Sybil attack: a machine conjuring not thousands but millions of plausible fake identities, each able to drown out, outvote and defraud the real people underneath. Scanned IDs won't save us — a model that can forge a face can forge a document. What it cannot forge is a place in the human family tree: generations of real, cross-referenced relatives who also turn up in everyone else's records.
"A fake identity can forge a paper. It can't fabricate a place in the human family tree."
So we go looking for uncontaminated ore. Records written before the machines are the low-background steel of the human archive: just as the steel smelted before the first atomic tests is prized because it carries no nuclear fallout, documents set down before generative AI are provably human, untouched by it. A 19th-century census page is the purest such ore there is — handwritten, sworn, cross-checked by enumerators, and decades older than any tool that could counterfeit it.
One record on its own proves little; a single fibre snaps. Strength comes from cross-linking — the way lignin binds loose cellulose into the rigid structure of wood. Every relative, neighbour, employer and witness who corroborates a person is another bond, until the names stop being isolated silos and become a lattice no forger can reproduce. That lattice is the thing machines cannot fake.
Think of it as a 3D White Pages. Old directories were flat snapshots: a name, an address, a year. This record tries to follow the thread through time, linking each person to earlier and later evidence, and timestamping new additions into the blockchain of records. That depth matters because the fraud threat surface is now much larger: claims about reality need correlations with the past and dense connection networks before they can be trusted. The same machinery has a fortunate side effect: it preserves hyperlocal history at a level ordinary directories never could.
That is the bet behind Certified H — a proof-of-personhood badge anchored not in a government file but in the verifiable genealogical graph of humanity itself, with the private tree redacted behind zero-knowledge proofs. Prove you are a unique node descended from real, cross-linked people, and a service can confirm you're human with a single API call — without ever seeing who you are.
But a graph like that has to be built, from real records, with rigorous deduplication so that one human is exactly one node. The Lakehead Reconstruction is the first proving ground: take one bounded, well-documented historical population and reconstruct it completely and correctly. Get everyone, link everyone, double-count no one — and the same method scales outward to the living, and ultimately to an open-source, cross-referenced record of every human alive, or who ever lived. Read the bigger picture at onlyhumans.info.
Census first. Deduplicate. Then enrich.
Five phases, run as a loop. The census is the spine, deduplication is the integrity layer, and newspapers begin with obituaries once every name has a node to attach to. AI does the reading and the matching; humans verify and capture.
Take everything already digital
The full Canadian census for the district is already scanned and free at Library and Archives Canada; directories and some papers sit on Canadiana and the Internet Archive. We also ingest high-quality biographies that have already done careful local-history work, such as Brent Scollie's book on early people in Thunder Bay. We pull these first, so the spine costs zero hours at a microfilm reader.
Turn census rows into people
Census handwriting goes through a frontier vision model, which extracts each row — name, age, birthplace, religion, occupation, dwelling & family number — into a structured stub with its citation attached. Shared dwellings give us our first family edges for free.
One human, one node
The same person recurs across 1881, 1891 and 1901. We block candidates by name-sound, birth-decade and sex, then let the model adjudicate the hard pairs — "Jno. Smith, lab., 1891" vs "John Smith, teamster, 1901" — keeping every merge reversible with its evidence. The result is a deduplicated population graph, built before a single newspaper is touched.
Now open the newspapers
Only now do we scan, and newspapers start with obituaries rather than the whole paper. Printed pages are OCR'd cheaply, then mined for names, relatives, businesses and life events — each one matched back to a known node or proposed as a new one. Obituaries are the richest first pass: a single notice can summarise a whole life and name a dozen relatives before we broaden into other newspaper items.
Write it back to Curriepedia
For every resolved person, the model drafts a referenced article from the structured record and its linked sources, published to curriepedia.org. Everything new flows back through the resolver, so the graph grows without ever double-counting a human.
Built on Curriepedia.
Curriepedia is a MediaWiki genealogy archive that has documented Thunder Bay and its families for almost ten years — pioneer biographies, family lines, the places and businesses that made the town. It already knows how to be a graph: articles as nodes, references as edges, every claim tied to a source.
The Reconstruction pours an entire historical population into that structure at scale. Each volunteer's day of careful scanning becomes hundreds of permanent, cited, public pages — a record that outlives all of us. Explore it at curriepedia.org.
We're looking for student volunteers.
This is a volunteer research role, ideal for a curious, detail-oriented high-school student on summer or winter break. You'll work shoulder-to-shoulder with frontier AI models at the Brodie Resource Library — feeding them historical records, checking their reasoning, and watching how they read handwriting, resolve identities and build a knowledge graph — with one-on-one mentoring on how today's large language models actually work. No experience needed — just patience, accuracy, and curiosity about the people behind the names.
Real archival work
Operate microfilm scanners, capture clean page images, file them to the project archive, and check the AI's extracted names and dates against the original. You become the ground truth the whole graph is measured against.
Frontier AI, up close
Direct, one-on-one mentoring on how advanced large language models actually work — prompting, the vision models that read handwriting, entity resolution, retrieval and knowledge graphs — taught by someone who builds with these models every day. You sit right next to the systems doing the reading.
A reference that counts
A detailed letter of reference documenting your contribution and the real AI, data and digitization skills you built — plus a permanent named credit as a contributor on Curriepedia.
This is an unpaid volunteer position; volunteers under 18 will need a parent or guardian's consent. The Lakehead Reconstruction is an initiative of Fling.AI Solutions Inc. in support of Curriepedia. Work is performed on publicly accessible, pre-1900 (public-domain) records under the library's terms of use.