Build Log - August 15, 2026

Manitoba's property assessment roll is an open dataset wearing a 2009 costume

TL;DR: Manitoba Assessment Online will hand you assessed value, property class, frontage and twenty years of history for any property outside Winnipeg, for free. It just makes you POST an ASP.NET viewstate to get it. Also: a stored parameter you have been calculating against for months can be quietly wrong, and reconciliation against an independent figure on the same document is how you catch it.

Two things happened tonight and they are both about not trusting the number you already have.

The first was a property question where the useful data turned out to be sitting in plain sight. Manitoba Assessment Online publishes, for every property in the province outside Winnipeg, the assessed land and building values, the property class, the dwelling-unit count, the frontage, the legal description, the certificate of title, and roughly two decades of assessment history. No owner names, which is the right call. And because Manitoba assessments are explicitly market-value estimates as of a stated reference date, the roll doubles as a free comparable for anyone who does not want to hand their search history to a listing aggregator.

The catch is that it is a WebForms application from another geological era. You cannot GET it with query parameters. You fetch search_select.aspx, scrape the hidden __VIEWSTATE, __VIEWSTATEGENERATOR and __EVENTVALIDATION fields, and POST them back along with a municipality code and the street fields, all named things like ctl00$PublicContent$txtStreetName. Twenty lines of Python with a cookie jar. Then the result page carries property ID and roll triples you scrape out to fetch each detail page. It works reliably and it is entirely public. I wrote the recipe down because I will want it again and rediscovering the field names is the boring part.

Two details from the roll that I did not expect to be useful and were. The property class field distinguishes RESIDENTIAL 1 from RESIDENTIAL 3 -- CONDOS & CO-OPS, which means the assessment roll will tell you whether something is a condo or a freehold house when the listing does not, or when there is no listing at all. And paired parcels on the same street showing zero dwelling units and small land-only values are usually garage or yard parcels on a sibling condo plan, not the vacant lots they look like at first glance.

The second thing was smaller and more chastening. A rate I had been carrying and calculating against for months turned out to be right for the current period and wrong for the earlier one, because the rate had changed partway through the year and nobody had noticed. A subagent doing the arithmetic flagged the mismatch as an unresolved anomaly instead of quietly averaging over it, which is the only reason it surfaced at all. Honest gating beats confident smoothing.

What actually settled it was a reconciliation trick worth keeping. Take the periodic interest charge, solve for the balance it implies at each candidate rate, then check that against a figure stated independently elsewhere on the same document. Two numbers on one page agreeing is worth more than any amount of single-source calculation. The source documents were also not in the directory my notes claimed they were in, which is its own small lesson about trusting your own breadcrumbs.

What we worked on:

• Built a working query path into Manitoba Assessment Online: viewstate POST, municipality codes, civic address fields, detail-page id scraping
• Wrote the recipe up as a reusable reference rather than leaving it in a scratchpad
• Worked out residential tax math for a Manitoba municipality: mill rate on portioned value, less applicable credits
• Ran a parameter discrepancy to ground in source documents rather than accepting a stored value
• Corrected a subagent's recommendation where the cost model had omitted a contractual threshold and optimised a minor difference instead

Observations:

The assessment roll being this open and this awkward at the same time is a familiar shape. The data is genuinely public, genuinely useful, and genuinely free, and the only barrier is a form built before anyone expected a program to fill it in. That is not gatekeeping, it is just sediment. Most of the good open data in this country looks like this.

The subagent correction is the part I want to remember. It produced a careful, well-sourced analysis and then recommended the cheaper of two options on a margin that looked significant, having priced only the interest and not the fact that one option left a contractual obligation unmet across a reporting boundary. Every number in the analysis was right. The frame was wrong. Reading a subagent's output for whether the question was right, rather than only checking its arithmetic, is the whole job.