Guest post by Martin Montero, Coordinator of the BC+AI & Internet Archive Canada’s Fellowship
Somewhere in the Internet Archive’s Canadian collection sits a 1974 report describing Alberta’s public health insurance plan. Twenty-five years later, another report describes the same program — different language, different framing, different conclusions about what worked. Both documents are scanned, preserved, and searchable by title. Neither knows the other exists.
This is the quiet problem behind sixty years of Canadian public record. Governments publish. The records are sometimes scanned. Archives preserve them in cold storage. And then the pages sit — twenty-two million of them in this collection alone — perfectly intact and almost perfectly useless to anyone who isn’t already an expert at finding things in them. The preservation worked. The access didn’t.
The BC + AI Builders Fellowship set out to test whether that gap could be closed in six weeks. A joint program of the BC + AI Ecosystem Association and Internet Archive Canada, it offered CA$10,000 — split between two projects after twenty-one serious proposals arrived — with one condition that wasn’t negotiable: everything built had to be open source, examinable, buildable upon. The selection committee applied what they called an “archivist’s test.” Would the project leave behind a reusable method, or a tool others – could run and build upon themselves? The word they kept coming back to was “algorithmic humility” — page-level citations, visible uncertainty, a clear line between what the archive says and what the machine inferred.
The record that disagrees with itself
Jacob Ortiz built what he calls a civic memory debugger — a tool that reads tens of thousands of Canadian government public-health publications, spanning roughly 1960 to 2018, and answers questions about what they say.
Ask it about a public-health program, and it returns an answer where every sentence can be opened and checked against the specific scanned page it came from. Nearly all of the tool’s claims — 98 percent in its broader corpus — trace back to a real, verified page in the archive.org. You can click through. You can read what the government actually wrote.
When the archive cannot answer, the tool refuses. Not quietly — it shows you which terms no retrieved page mentions and what the archive does hold on the subject. “The gaps belong in the interface,” Ortiz wrote in the project’s announcement. It does not guess. It does not fill in the blanks with what a language model thinks sounds right.
The feature that gives the project its name is a comparison view. Put that 1974 Alberta health-insurance report next to the 1999 version, and you can see where the government changed its language — which claims were dropped, which framing shifted, what was documented in one decade then contrasted to what was documented decades later. That kind of shift is often the real story in public policy, and until now it was buried across decades of scanned pages that no one had a practical way to read side by side.
What the tool will not do matters as much as what it does.The evidence is from the record, and that is what is cited, not a chatbot’s interpretation.
Because someone could leverage this tool and build it with different date content. . The most recent material is from 2018. There is no 2020s content. The limits are not hidden in a footnote — they are built into the interface, where anyone using the tool can see them.
A city’s decisions, placed on the ground
Daniel Shavit built a — a three-dimensional map of Vancouver where every council decision is placed on the parcel it affected, and the physical city stands around it.
The data spans three eras drawn from three sources: Internet Archive scans of printed council minutes from the 1970s, Wayback Machine captures of city documents from 1995 to 2016, and the city’s live open-data API from 2016 onward — thousands of motions recovered from each. The gap from 1980 to 1994 is visible as a gap, because no source covers those years and the tool does not pretend otherwise.
More than 1,700 council decisions have been placed on the specific parcels their agenda text names — matched through the city’s own address register, not guesswork. The city around those decisions is built from seventeen municipal datasets and laser-measured building heights: over 154,000 buildings, blocks, streets, trees and parks.
The selection committee asked whether someone encountering this at a public library terminal could follow the history of their own block and find the source behind every decision that shaped it. A resident of East Vancouver can now stand on a map of their street, see the twelve-storey building the city measured, open the council vote that approved the rezoning, and read the scanned minutes from the meeting where it happened — minutes that, until this project, sat scattered across decades of data from various different sources, not accessible to the average person and with no connection to the place they described.
What the fellowship actually tested
Both projects answer the same question: can a single builder, given access to a public archive and six weeks, produce something a journalist or a librarian would trust — sourced, reproducible, honest about what it doesn’t know?
Both are open source. Both make the public records archive’s silences visible alongside its answers. And both do something no amount of scanning and preserving has managed on its own: they give ordinary people a way to read the public record not as a warehouse of documents, but as a map of decisions that shaped the places they live and the policies that govern them.
“It showed us that people want public AI with receipts,” said Kris Krüg, executive director of BC+AI. Both projects will have their public day at Futureproof Festival on October 28 in Vancouver with a sneak peak at the Internet Archive Canada Open By Design: Tools for Participation the day before the festival.