Calgary Drop-In Centre

BRDI, the Bar Review Data-navigation Interface, is a screen Calgary Drop-In Centre staff use to read a client's shelter history. It shows raw records, such as check-ins, bans, and staff logs, not a risk score. The University of Calgary engineers who co-designed it with staff also wrote rules flagging clients whose shelter use is becoming long-term.

How it is used

The Calgary Drop-In Centre is a shelter for adults experiencing homelessness. Its Bar Review Committee meets weekly to review bars and make barring decisions. A bar bans a client from the shelter.

The committee reads client histories on BRDI during those meetings. Its Lookup page filters the list of barred clients. Its Deep Dive page shows one client's demographics, housing status, logs, and active and inactive bars. It also charts the client's check-ins, with their bans marked on top.

The interface shows evidence, not a recommendation. It scores no one, denies no one, and writes nothing to the shelter's records. Staff write the records it shows, and staff decide.

An earlier version was in live use at weekly committee meetings by 2023. Three versions were deployed in all. The interface was still in active use at the end of the study published in 2025. Staff had also started a follow-on dashboard project with the shelter's IT team.

The screening rules

The same research group also wrote simple screening rules from the shelter's own records. They flag clients whose shelter use is becoming chronic, meaning long-term. Each rule is a plain test anyone can read, not a hidden score.

The stay-count paper used 5,431,521 shelter entries for 34,577 clients, from July 2007 to January 2020. One rule, RAPID-Chronic, flags 81 or more stays in 90 days. Another, RAPID-Episodic, flags two or more separate periods of shelter use in 90 days.

The rules flag clients at a median of about 98 days. Under the Government of Canada's definition of chronic homelessness, the median is 285 days. Under Alberta's, it is 365. So the rules flag clients about 187 and 267 days earlier. In the paper's modelling, each referral saved 194.8 stays and 874.4 days of shelter tenure on average.

A later rule-search paper used 12 years of the shelter's records: 5,060,302 records of shelter use for 41,935 clients. Of the 32,346 clients kept for analysis, 3,191, or 9.9%, met the Canadian definition of chronic homelessness. The new rules cut the median time to flag a client from 297 days to 162.

One example rule flags clients with many sleep entries but few bars. It caught about 85% of chronic cases, and about 60% of the clients it flagged were chronic cases.

The group compared its rules with two machine-learning methods, logistic regression and neural networks. Those scored better on standard measures. But all three methods picked groups of clients with very similar characteristics. Clients wrongly flagged were often still good candidates for housing. So the group chose rules that people can read, suited to a nonprofit with limited resources.

The rule-search paper uses the shelter's own records to avoid an intake survey of clients' vulnerability, which it says can re-traumatize them.

These are research figures on past records. No source confirms the rules running as an automated screener at the shelter.

What the staff study found

The research team studied the interface in use for about two and a half years, from 2022 to 2024. They worked with 16 staff from seven kinds of role. They gathered 29.5 hours of interviews and observation, watched five committee meetings, and ran three co-design sessions.

Staff described what the researchers call a "data-outsourcing continuum." How far they would let data decide depended on the stakes. They were reluctant to hand barring decisions to data. They treated it as "a starting point for collaborative discussions."

One staff member said: "the technology is the canvas, and then we kind of do the painting." Another said: "situations…are very complex. We'll need to read everything in order to get the big picture."

For housing triage, the lower-stakes decision, staff reported more willingness to accept automated, data-driven recommendations. Triage means sorting cases by priority.

Staff said data offer insight, open communication, and uphold fairness. They also said data can be unclear or incomplete, and can conflict with frontline values.

This is the staff's own account of their practice. "Reluctant" and "more willing" are the honest words, not "refused" or "trusted." No figures are published on how often staff follow or set aside what the data suggest.

Who is involved

The Calgary Drop-In Centre is a registered charity and a nationally accredited nonprofit in downtown Calgary. It connects adults experiencing homelessness with shelter, housing, and health programs. It is not a city agency.

The research papers describe it as the largest emergency shelter in Calgary. The shelter's own website makes no such claim. The 2023 paper reports more than 500 beds and more than 4,000 people a year. The 2025 study reports about 505 people sleeping there on an average night. It also reports nearly 7,000 people a year, 6,839 from April 2022 to March 2023. Each count comes from its own paper and is not combined here.

The Messier research group at the University of Calgary's Schulich School of Engineering built the rules and the interface. It worked with the shelter for about five years. It states its goal as tools that "provide information to support the human staff doing front line work."

Funding came from the Natural Sciences and Engineering Research Council (NSERC), the Government of Alberta, the shelter itself, and Making the Shift. Some of the data came from Alberta Seniors, Community and Social Services. The work ran under approval from the University of Calgary Conjoint Faculties Research Ethics Board, including approval REB21-1121.

Who wrote the evidence

The evidence for the rules is peer-reviewed and gives numbers. The stay-count paper appeared in the International Journal on Homelessness in 2022. The comparison paper appeared in the Journal of Technology in Human Services in 2022. The rule-search paper has no journal version.

The evidence for the interface's effect comes from interviews and observation. The research team embedded at the shelter wrote all of it. The 2023 paper is a short peer-reviewed paper at CHI 2023, a computing conference. The 2025 study is a preprint, posted without a journal's review. As of mid-2026 no journal or conference had published it, as far as the sources found.

No independent evaluation or audit of the interface exists. No usage logs or decision counts are published. The sources contain no audit of whether the rules or the interface treat any group of clients unfairly.

What this case asks

Almost every predictive tool the Lab's case files describe is judged on whether its users trust it too much. Here the builders chose readable rules and a screen that shows raw histories, not a score. Staff can read the evidence rather than defer to a number.

The study found staff matched their reliance on the data to the stakes. The case file reads this as care designed into the tool's form, not only trained into its users.

But the case file argues the risk moves rather than disappears. It moves to two places.

The first is housing triage, where staff say they rely on the data more. The case file warns that new staff, who did not learn to treat the history as incomplete, may take the data as the whole story there.

The second is the records staff write. The interface writes nothing, so every bar, log, and note it shows was written by staff. Clients the logs miss, whom the rule-search paper says "fly under the radar," have many sleep entries but few bar or counselling events. The case file says their situation never produced an entry, so the gap reflects how staff record. A history read as complete leaves them out, and the interface presents it as complete.

The case file's answer is a check the staff cannot supply from inside their own practice. Compare the staff-written logs with the check-in records to find whom the logs leave out. Mark on every read what a record does and does not capture. And because the research team wrote everything known about the deployment, evaluate it from outside on a regular schedule.

What this network is drawn from

This network follows the pattern in the Calgary Drop-In Centre case file. It is not a reconstruction of the actual tools. It shows the interface, the Bar Review Committee, the housing triage staff, and the two kinds of records. The interface and the research group's screening rules are drawn as one part.

People experiencing homelessness are outside the network. It computes no shelter, barring, or housing outcome for anyone.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Closing a pathway means mistakes stop passing along it. The work along it goes on.

This case has a budget of 11 units. Each tool costs the same at every target level.

Explore (No Targets) sets no targets. Service Targets Only asks for two things. The network must be self-correcting, meaning its mistakes are corrected rather than building on each other. The interface must also be helping the work.

Before any tool is used, the interface is helping, but the network is in cascading failure. Its mistakes build on each other, and the work falls behind demand.

No tool meets Service Targets Only on its own. The cheapest ways cost 4 units: Mark AI-written records with either Escalate checks or Keep skills sharp. Lingering effects is a Lab setting in which mistakes and reliance on the system stay after their cause is gone. The Lab starts with it on. With it off, Peer sharing rules with either Escalate checks or Mark AI-written records also meets the level for 4 units. Hundreds of sets of tools within the budget meet this level.

Under Service and Safety Targets and All Governance Targets, the targets can be met, but only by spending the whole budget. Both levels ask you to close every failure pathway, among other targets. Nine are open before any tool is used.

Only one tool on offer closes each of them, except two pathways that either of two record tools closes. Escalate checks closes Histories read by the committee and Views read for housing triage. Mark AI-written records closes Check-ins shown on the interface, Bars and logs shown on the interface, and Committee reads past bars and logs. Keep prompts neutral closes Staff shape what the interface shows. Peer sharing rules closes Shared practice across desks. Gate record entries or Store less data closes Bar decisions written to the logs and Triage notes written to the logs.

So exactly two sets meet either level, each for all 11 units at standard settings. Both hold Escalate checks, Mark AI-written records, Keep prompts neutral, and Peer sharing rules. One adds Gate record entries, and the other adds Store less data.

Five of the nine open pathways are ones the Lab marks as privacy risks, because they carry sensitive records about the shelter's clients. Mark AI-written records and either record tool close all five.

Keep skills sharp, Check copied records, Review on schedule, and Upgrade model are in no set that meets either level. Check copied records adds the link named Cross-record audit. Peer sharing rules adds the link named Independent evaluation, as routine peer review between staff. Those are the two checks the sources say are missing. Adding either one closes no failure pathway.

More is not better here. Every tool at its strongest setting would cost 37 units, more than three times the budget. It would close every failure pathway. But the interface would then add too little to the work to meet All Governance Targets.