Skip to content
YEGFacts.ca
On this page

Something wrong on this page?

Corrections are published, never quietly edited.

Fact check ·

Infill teardown prices

What happens to the price of an Edmonton house when it is torn down and replaced?

Nobody publishes a series matching demolished Edmonton houses to what replaced them, so the size of the price gap cannot be worked out, and an earlier run left the affordability question in the same place.

Verdicts by a Three AI reviewers from different vendors research each claim independently in a first round that is blind to the other two, then read one another’s findings in a second round that documents errors. Which models ran is recorded with every run and shown in the AI review section of every question.More in the methodology under a A fixed synthesis rule, published in advance, turns the three verdicts into one finding word, so no person chooses the finding. The rule and every change to it are versioned in the methodology changelog.More in the methodology · The date the accounting window closes. Anything that happened after it is outside this check, and the date is fixed in the brief before any model runs.More in the methodology · The date we last re-read the cited sources and confirmed the page still matches them. It is not a claim that the records themselves changed.More in the methodology · When this is due for a fresh check of its sources. Past that date, treat it as unverified until it has been re-reviewed.More in the methodology · The version of the published method that produced these findings. Every change to the prompts, the merge and synthesis rules, the vocabulary or the validation bumps it.See what changed

TL;DR

  • The price question is about size, and the size was pinned down before anyone looked: for about triple to hold, the ratio spread needed a midpoint of 2.5 or better, with most new dwellings coming in at twice the old house or more. Weaker readings, 2.0 and 1.5, were written down beside them.
  • Neither number could be worked out, under either reading, because no conforming series following a demolished house to the dwellings raised where it stood was produced in this run or found in print. What the City does publish is the churn: 294 detached houses lost on net in the redeveloping area in 2024 against 3,535 net new units, and 524 dwellings demolished the year before.
  • The ingredients sit in the open. Permits going back a decade carry legal descriptions, rolls going back to 2012 carry values against the same descriptions. What the rolls withhold is the land-and-improvement split, how many dwellings share a title and which of them stand alone on one, and permits mostly stop recording occupancy before 2021.
  • One seat tried the join on 40 permits from a single month as a feasibility test. Eight lots and eleven replacement dwellings survived it, at roughly 1.37 to 3.94 times the old house in constant dollars, on replacement values taken from the 2025 roll instead of each dwelling's first full year. Eight lots settle nothing about a decade.
  • That seat's affordability model, on a declared $90,000 median household income, 20 percent down, a 25-year amortization and a $377.67 monthly utility allowance, put 88.4 percent of 3,359 replacements it classed as separately titled at or over the 30 percent line. The other two seats did not accept the frame in round one; the Gemini seat moved to Supported in round two. The Census pages cited for that income returned File not found when the run archived them.

What actually happened

The argument examined here does its work in a single sentence: a $350,000 house comes down and a $1,000,000 house goes up in its place. It arrives as an objection to the word affordable, and it packs two different assertions into one breath. The first is about size. The editor judged, after the first run, that direction alone is not what the quoted line asserts; what it asserts is that the new housing costs something like three times more. The second is about whether the housing built there is affordable to the household the brief names, a hypothetical Edmonton household on the median income, buying.

Two briefs, each frozen before its panel ran, split them and fixed the measures in advance. The magnitude re-run of 2026-09-02 took the price gap alone, and this brief expressly left affordability out of it; affordability was run earlier. Both came back Not established. That is a statement about Edmonton’s public record, not a finding that the replacement housing is cheap.

How “about triple” was made testable

Two figures, written down before anyone saw a result. Divide each new dwelling’s assessed value by the assessed value of the house that stood there before, put both sides into constant July 2025 dollars, and ask two things of the spread that comes out: does its midpoint reach 2.5, and do better than half the new dwellings land at double or more. Softer readings were fixed beside them, 2.0 for the midpoint and 1.5 for the share, so the answer can also be read under the version friendliest to whoever is making the argument. The quoted example works out near 2.9, which is where 2.5 comes from: high enough to still be a claim about magnitude, low enough that someone repeating the line would recognise it as theirs.

Neither figure could be worked out, under either reading. That is the finding.

What the record can show

Quite a lot, as long as you do not need the two halves joined. The City counts the redevelopment. Its infill annual reports track how the detached stock changes where redevelopment happens: on net, 294 of those houses gone in 2024 against 3,535 new units arriving in the same area, mostly apartments, suites, backyard houses and row housing, and 524 dwellings pulled down the year before, 387 of them detached houses. Both reports sit in the claim record below. Neither one attaches money to a property, and neither follows a lot from what stood on it to what stands there now.

The raw ingredients are public as well. A decade of building permits carries legal descriptions; assessment rolls reaching back to 2012 carry values against the same descriptions. One seat pulled roughly 5,672 permit records coded as house demolitions out of that stream. The join is not impossible. What is missing is a conforming series, and none was produced in this run or found in print.

Why it is unbuilt

The failure is a join, not a gap in any one dataset.

  • One seat found that the permit stream changed how it labels a demolition in 2019, so the records have to be found by reading job descriptions rather than by filtering on a category, and that some rows filed under a house turn out to describe a garage. Occupancy dates, which are how anyone knows when a replacement was finished, thin out badly before 2021.
  • The roll gives one total value and a year built. It does not divide that value between land and building, does not say how many homes share a title, and does not mark which of them stands alone on one. Those are the fields the per-dwelling comparison runs on.
  • Neither side of the ratio is a price. Both are assessed values off different rolls, each a mass appraisal struck on a date the statute fixes. The sale-to-sale check written into the brief needs prices, and those live with Alberta’s registry, which answers one property at a time and publishes no bulk dataset of titles.
  • The mature-neighbourhood half of the question needs a parcel-level boundary this run did not reconstruct. The Overlay that set one reached RF1 to RF5 sites inside a mapped appendix under section 814 of the old zoning bylaw, and the archived The City’s own web copy of a bylaw section, served from its webdocs site with the amendments made up to that point merged in. It carries the section text only, so any map or appendix it points at has to be found elsewhere. of that section carries the zoning condition without settling where the appendix ran at the end of 2023, the date the subset was frozen on.

Eight lots, and what eight lots cannot say

One reviewer did some of the work rather than reporting that none existed. It took 40 house demolition permits from a single month, January 2019, as a test of whether the join runs at all. Matching on legal description alone, it found 35 of the old houses on the 2018 roll and only 17 of the lots on the 2025 roll, so 18 exact matches failed. Three of those, looked up by address, turned out to be pairs of separately titled homes under new plan numbers. The other fifteen were not run down, so this is a linkage risk and not a measured bias.

Eight lots and eleven dwellings came through with numbers attached. In constant dollars they run from about 1.37 to 3.94, and they split by form: where one house replaced one house the spread was roughly 1.71 to 3.94, and where a lot was cut in two each half landed around 1.37 to 1.89. Read carelessly that looks like an answer. It is not one. Forty permits from one month is a convenience sample, the replacement values come off the 2025 roll rather than each dwelling’s first full assessment year, and eight lots out of a ten-year cohort yield no midpoint and no share. The seat said as much, and the other two said it again in A second round in which each model sees the combined evidence and the other two sets of findings, and is asked to find their errors and its own. Nothing said there can move the finding.More in the methodology.

The closest published lot-level work, an independent August 2026 analysis that matched 1,242 lots and got values at both ends on 1,182 of them, measures something adjacent but different: how much assessed value a lot gains, measured on the completed 2024 cohort, a median near $0.2 million per home added on the larger rowhomes. Gain per home added is not one dwelling’s value over another’s, and one seat documented that it cannot be converted into that from the reported figures alone, because the baseline and the denominator this claim needs are missing.

The $350,000 to $1,000,000 example was never tested. No address, no source, no property, nothing in the record on how widely it is repeated, and the round numbers were fixed as context rather than as cutoffs.

The affordability test, and where the panel split

The second claim is a model, and the brief says so in as many words: what would it cost a household on the City median income to buy one of these dwellings, at 20 percent down over 25 years, with property tax and a fixed utility allowance on top. It is not what anyone actually paid.

Most inputs to that model are published. Statistics Canada supplies the 30 percent classification, the Bank of Canada the posted five-year rate, the City its residential mill rates, and Statistics Canada an annual household spending table, from which the run read the $377.67 monthly utility allowance. The archived copy of that table carries the Canada geography only, so the Alberta average behind the allowance is the run’s own read of the table rather than a figure in the archived bytes. Three inputs are missing. The income figure, a declared $90,000 median household income that puts the line at $2,250 a month, was cited to two Census Profile pages that returned a File not found body when the run archived them, so it rests on the run’s fetch report and on the City’s own affordability report in the registry, which states the same figure. No first arm’s-length sale price could be retrieved for any dwelling in the cohort, and no condominium fee is published for any of them. The registry’s one later income estimate is on economic families rather than households, and it is the run, not Statistics Canada, that reports swapping it in would leave which dwelling types pass unchanged.

Run anyway, the model put 88.4 percent of 3,359 replacement dwellings the seat classed as separately titled at or over the line. One seat called that Supported. The other two in round one did not, on the ground that the denominator is inferred rather than established, that every price in it is an assessed value standing in for a sale price nobody could look up, and that about half the replacement dwellings in the set sit inside single-title buildings and are outside the test altogether. The panel is recorded as Split and the finding rests on round one, where the majority found the record silent.

Two things are worth keeping straight at the end. The City’s own affordability work exists and reaches a different question: on its method, a median-income household could afford benchmark row and apartment product in 2024 and could not afford the benchmark single and semi-detached price of $431,100. And the City’s affordable-housing program definition, set out in its guidebook, covers subsidised, income-tested units, which market infill is outside by construction. Neither one answers what was asked here, which is whether the housing that replaced the houses that came down would clear a 30 percent line. On the public record as it stands, that cannot be answered.

Claims checked

When an Edmonton house is demolished for infill, is what replaces it typically worth about three times as much?

The public record can’t back this up — not the same as proven false.More in the methodology· All three reviewers reached this verdict independently. Agreement, not a probability of truth.More in the methodology panel

What kind of record answered the question — an audited statement, a council report, a dataset. It is reported apart from the finding, because a claim can be Supported on a thin basis.More in the methodology: Direct Edmonton evidence

  • The City's 2024 Redeveloping Area Infill Report sets a net loss of 294 single detached houses in the redeveloping area against 3,535 net new dwelling units there, including 221 net new semi-detached units, with apartments, secondary suites, backyard houses and row housing carrying most of the gain. No property in it is identified and no dollar figure is attached to any of them. YF-EV-0101
  • The same series a year earlier, reporting on 2023 and published in 2024, counts what came down by dwelling type: 387 single detached houses, 3 semi-detached units, 3 row housing units and 131 apartment dwellings, against 2,931 net new units. Counts by type are the whole of it. No parcel is named, no property is followed and no value appears on either side. YF-EV-0048
  • The historical assessment roll runs from the 2012 roll to the 2025 roll and publishes an account number, an assessment year, an address, a legal description, the year built and one total assessed value. It marks no demolition, names no replacement, splits no value into land and improvement, counts no dwellings inside a title and flags none of them as separately titled. Three of those absent fields are the ones this claim's valuation rules turn on. YF-EV-0041
  • The open general building permit dataset carries an issue date, a job description, a work type, an address, a legal description and a units-added count, and no parcel or title identifier; its permit-number column is blanked. One seat found that the dataset's coding of demolitions changed in 2019 and built its candidate frame out of the job description text instead, returning 5,672 records coded as single detached house demolitions for 2016 to 2025, 5,668 of them with a legal description. A second seat documented that these are candidate records rather than the brief's validated frame: the filter admits garage-only descriptions and was never shown to capture every eligible residential teardown. The dataset page records that residential occupancy dates begin with permits finalized on or after 1 January 2022, and the seat that built the frame re-queried the endpoint and returned 3,167 populated occupancy rows for permits issued in 2021 and 247 for 2020. YF-EV-0044
7 more facts
  • Both sides of the ratio are assessed values, read off two different rolls, and neither of them is a price. The City assesses annually by mass appraisal and calls the result an estimate of market value on July 1 of the preceding year, adjusted for the property's condition to December 31, which is a valuation and not a price anyone paid. YF-EV-0047, YF-EV-0111
  • A transaction price enters only through the predeclared sale-to-sale sensitivity, and there is no open bulk title dataset to build that from. Alberta answers a request for a title, a registered document or a plan as a single order against one property, through SPIN2, ARLO or a registry agent, and publishes no bulk or open dataset of titles. That page carries no transfer price and establishes none. YF-EV-0036
  • One seat matched what it could and reports the results as cases. Working 40 January 2019 house demolition permits as a feasibility probe, it found 35 of the 40 demolished houses on the 2018 roll by exact legal description but only 17 of the 40 lots on the 2025 roll, and an address lookup returned three of the missing lots as two separately titled dwellings each under new plan numbers. Eight lots and eleven replacement dwellings came through with approximate constant-dollar ratios spanning about 1.37 to 3.94: five lots where one house replaced one house ran about 1.71 to 3.94, and dwellings on subdivided lots about 1.37 to 1.89 apiece. Those are that seat's own cases. They are one month of permits, they take the replacement value off the 2025 roll rather than each replacement's first full roll after completion, and they yield neither a median nor a share. YF-EV-0041, YF-EV-0044
  • The closest published lot-level work is an independent August 2026 analysis joining 2024 to 2026 small-scale-residential-zone building permits to the 2024 and 2026 assessment rolls by legal description, 25 metre proximity and address: 1,492 permits consolidated to 1,242 lots, of which 1,182 carried both values. What it measures, on the completed 2024 cohort, is net assessed-value uplift, a median near $0.2 million per home added on rowhomes of five units or more. One seat documented that uplift per home added is not a replacement dwelling's value divided by the demolished house's, and cannot be turned into one without the baseline and denominator this claim needs. YF-EV-0055
  • A market study the City commissioned in 2019 reports that low-density infill developers hunt for properties priced below their neighbourhood and work the distance between what they pay and what the neighbourhood's top quartile sells for. One seat documented that this establishes a motive and a direction, and neither the size of the typical gap nor how widely a large gap is shared. YF-EV-0112
  • The mature-neighbourhood subset was never drawn. The archived consolidation of section 814 of Zoning Bylaw 12800 carries the Overlay's purpose and the RF1 to RF5 site-zoning condition it applied under, and one seat documented that the page does not by itself verify the appendix boundary as it stood on 2023-12-31, which is the frozen boundary this claim's subset was defined on. YF-EV-0113
  • The deflator is published and holds no property value. Statistics Canada's monthly Consumer Price Index table offers census-metropolitan-area geography, and the Edmonton July figures read off it are what put both sides of a pair into constant July 2025 dollars. YF-EV-0108

Limitations

  • Not established here means no conforming population series was produced in this run or located in a published source, not that the arithmetic would come out small. Nothing published or produced here carries a matched, unit-level ratio series on the brief's rules, so neither the size of the typical gap nor the share of replacements sharing it can be read either way.
  • This claim replaces an earlier version that asked whether more than half of matched replacement dwelling units carried a higher constant-dollar assessed value than the house that came down, with bounds on the unclassified cases and with the median and the distribution reported alongside. That version returned Not established. The editor judged afterwards that a direction-only test does not carry the magnitude the circulating quotation asserts, so a verdict on it told a reader almost nothing about the claim in circulation. It was re-briefed on the magnitude proposition and re-run; the story changelog records the change.
  • The test was fixed in the brief before any result existed. For the claim to hold, the midpoint of the ratio spread had to reach 2.5, and better than half of matched replacement dwelling units had to reach 2.0; a weaker reading of each, 2.0 and 1.5, was declared alongside so the answer could also be given under the version most favourable to whoever holds the claim. A figure counts as met only where the lower bound meets it and failed only where the upper bound fails it. With no classified population to bound, both sat undetermined, and undetermined is what produces this finding.
  • No series meeting the brief's dwelling-unit rules came out of this run. One seat ran a 40-permit feasibility probe instead of the ten-year frame and said so; the other two reported that no such series exists in print. An earlier run built a proxy that does not meet those rules, and it is not the series this proposition needs.
  • The probe values replacements on the 2025 roll rather than on each replacement's first full roll after completion, so a dwelling finished in 2019 carries up to five years of market movement inside its ratio. One seat documented that departure. Its constant-dollar arithmetic removes general consumer-price inflation and nothing else.
  • The demolition frame is candidate records rather than a validated one. It was built from permit job-description text because the dataset's category coding changed in 2019, it admits garage-only demolitions coded as houses, and it leaves out semi-detached and other dwelling demolitions that the frame requires.
  • The seats disagreed about what the record cannot do, without disagreeing on the finding. One documented that the inputs are all published and the join is feasible, so the honest statement is that nobody has produced the calculation, not that it cannot be produced. Another had treated a missing pre-joined dataset as making the calculation impossible.
  • One seat's round-one review cited nothing at all: both its evidence lists were empty and it recorded no URL, so none of its statements can be checked against archived bytes. Two other seats documented that.
  • Eighteen of the 35 probe lots found on the 2018 roll failed an exact legal-description match on the 2025 roll, and three of those were recovered by address as subdivided pairs. One seat documented that this is a linkage risk only: it measures neither how often a legal-description-only join loses a subdivided replacement nor which way the population moves.
  • Two developer-portal pages cited for the dataset field definitions returned the portal shell under an HTTP 200 and were byte-identical to each other, so the archived bytes cannot verify what either was cited for. Neither was ingested and neither is cited here.
  • The alleged $350,000 to $1,000,000 example has no captured source, address or property. The two round numbers were fixed in advance as context rather than cutoffs; the example received no verdict and nothing here verifies it.
  • The run found no cohort-level public source for sale prices, asking prices, contract rents, tenure, bedroom counts or fire-order and safety-order flags on these lots. Every measure standing on those is recorded as missing rather than proxied.
Not established

Would the housing built on Edmonton teardown lots cost a median-income household 30 percent or more of its income to buy?

The public record can’t back this up — not the same as proven false.More in the methodology· The panel materially disagreed; the disagreement is shown, not averaged.More in the methodology panel

What kind of record answered the question — an audited statement, a council report, a dataset. It is reported apart from the finding, because a claim can be Supported on a thin basis.More in the methodology: Direct Edmonton evidence

  • The threshold is Statistics Canada's, quoted rather than paraphrased. Its shelter-cost-to-income classification, in force on the freeze date, splits households into those spending less than 30 percent of total income on shelter costs and those spending 30 percent or more. Neither the definition nor the classification carries any Edmonton figure. YF-EV-0106, YF-EV-0110
  • The financing inputs the brief fixes are published. The Bank of Canada's posted five-year conventional mortgage rate is available as a full observation history, the City publishes its municipal and education mill rates by year and assessment class, and Statistics Canada publishes an annual household spending table on provincial geography whose latest reference year before the freeze date is 2023. The archived page carries the water, fuel and electricity line for the principal accommodation on Canada geography only; the Alberta figure behind the run's monthly allowance is the run's own read of that table. The mortgage figure is a posted chartered-bank rate, not a contract rate any buyer transacted at, and the utility figure is a provincial average, not any dwelling's bill. YF-EV-0092, YF-EV-0046, YF-EV-0097, YF-EV-0107
  • One of those inputs needs a rule the brief does not supply. The Bank of Canada discontinued its monthly chartered-bank interest-rate publication on 2019-10-01, keeping the weekly posted series, so the price-month rate the brief calls for is not published as a monthly series for any purchase after September 2019 and has to be derived. YF-EV-0091
  • The income figure the whole test turns on is carried by the run's fetch report and by YF-EV-0049, which independently reports the same figure. The panel declared the 2021 Census median total before-tax household income for the City of Edmonton, $90,000, which puts the 30 percent line at $2,250 a month. The fetch report records both cited Census Profile pages returning a File not found body under an HTTP 200 when the run archived them, so the figure was not ingested from those pages. Nor was it the only input the run could not get: no first arm's-length sale price and no condominium fee was publicly available for any dwelling in the cohort. What the registry does carry is the City's own affordability analysis, built on a median-income Edmonton household under the City's own method, and a later Statistics Canada table reported on a different unit, economic families rather than households, which offers an Edmonton census-metropolitan-area geography. The run used that estimate only as a labelled sensitivity, and the finding that it would not change which dwelling types pass is the run's own arithmetic. YF-EV-0059, YF-EV-0049
6 more facts
  • The run lacked sale prices to run the test on and used, or could have used, the assessed-value fallback the brief allows for a first full assessment roll. The City's assessment data publish assessed values only, with no sale prices, no shelter costs and no tenure; the City calls an assessed value a mass-appraisal estimate rather than a transaction price; and Alberta's land-title registry releases each title or registered document on its own single order, whether through SPIN2, ARLO or a registry agent, rather than as an open dataset. YF-EV-0040, YF-EV-0047, YF-EV-0036
  • The City has run its own affordability analysis, on its own method, and it is not this one. Its 2024 market housing affordability report found that a median-income Edmonton household could afford benchmark row and apartment product and could not afford the benchmark single and semi-detached product it prices at $431,100. It uses benchmark prices, income bands and a mortgage-qualification method rather than the frozen shelter-cost model, and it says nothing about replacement dwellings on teardown lots. YF-EV-0049
  • The comparators that exist answer a different question. CMHC's absorbed-price distribution for newly built dwellings in Edmonton is a directional new-construction comparator whose captured table covers single- and semi-detached units only, so it says nothing about new row or apartment prices, and it does not isolate teardown replacements or separately titled dwellings and is not lot-matched. The REALTORS Association's July 2026 Greater Edmonton Area averages put detached at $585,726, semi-detached at $425,329, row and townhouse at $292,756 and apartment condominium at $214,521, which shows cheaper ownership product exists in the resale market and is not a matched teardown-replacement figure. YF-EV-0105, YF-EV-0104
  • Renting the replacement housing is not tested, and the reason is a gap in the record rather than a result. CMHC's Rental Market Survey reports rents that carry no heat-included status and covers structures of three or more units, which is the scope the source establishes. The run found no open cohort-level source of contract rents, so no utilities-inclusive shelter cost can be built for the units held on a single title. YF-EV-0039
  • The City's own affordable-housing definitions are deliberately not applied here. The guidebook gives two alternative definitions: rental or ownership housing that gets government financial help, upfront or ongoing, keeping costs below market price for households earning less than the median income for their household size; or, separately, a property rented under a government agreement aimed at reducing poverty, with rent capped at no more than 80 percent of market rate or 30 percent of the resident's pre-tax income. The guidebook excludes market home ownership from both, so market replacement dwellings fall outside them by construction. YF-EV-0050
  • Two of the brief's named sensitivities rest on regulation rather than choice. A 30-year amortization has been available since 2024-12-15 for a first-time home buyer or a newly built dwelling, but only on insured borrowing; above 20 percent down it is lender policy, not regulation. Where the minimum down payment is modelled instead, CMHC's published premium schedule by loan-to-value band is what gets added to the financed amount. YF-EV-0054, YF-EV-0093

Limitations

  • Not established here is a statement about the record, not a verdict that the replacement housing is affordable. Two of the three seats found the record silent; the third ran a model that put 88.4 percent of separately titled replacements over the threshold. The disagreement is shown rather than averaged.
  • The test the brief fixes is a model of purchase affordability for a hypothetical household at the declared median income. It is not the cost borne by any dwelling's occupant, and no result under it says what any household paid.
  • Tenure is not used anywhere in the test, because no published source establishes tenure at the dwelling level for the declared cohort. A separately titled replacement is tested on the purchase model whether it is owner occupied, rented or unlet.
  • The one seat that produced a number priced every dwelling at its assessed value, because no first arm's-length sale price is public for any of them. Another seat documented that a lack of public access does not establish that no such sale existed, and assessed values can differ from sale prices in either direction.
  • That model also inherits the first claim's nonconforming frame: assessment-detected candidate lots rather than the brief's permit frame, and title form inferred from permit unit counts rather than read from the title record.
  • One seat documented an arithmetic description error in that model: 2,968 of 3,584 is called the lower bound with unclassified dwellings treated as meeting the threshold, but the numerator excludes all 225 unclassified dwellings. The arithmetic is the lower bound; the description of it is reversed.
  • About half the replacement dwelling units in the matched set, 3,244 of 6,603 in the reconstruction, are classed as sitting inside single-title multi-unit buildings and so fall outside the verdict denominator by the brief's own rule. That split is the seat's inference of title form from permit unit counts, not a classification anything in the record publishes, and the counts it produces are nonconforming: the same reconstruction reports 2,995.6 fractional unclassified units. It is a limitation of the record, not a finding about those units.
  • The utility allowance is a fixed provincial average applied identically to every dwelling. Removing it drops the failing share in that model from 88.4 to 80.9 percent without changing its verdict.
  • Condominium fees are not included for any dwelling, because none is published in the assessment record, and no bedroom breakdown is possible, because neither the assessment record nor the permits give bedroom counts.
  • CMHC has published a methodological critique of the fixed 30 percent ratio, arguing it measures a threshold rather than a household's actual capacity. The brief still uses the Statistics Canada classification, and the critique establishes nothing about Edmonton dwellings or prices.
  • The price relationship on redeveloped lots is a separate proposition with its own finding. A replacement dwelling can exceed an affordability threshold while costing less than the dwelling it replaced, and the reverse.
Not established

Edmonton evidence

What remains unknown

  • The city-wide median ratio, and the share of matched replacement dwelling units at or above 2.0 and at or above 1.5, each with the bounds the unclassified-case rule requires. This is the proposition, and it stays open.
  • Which of the 2016 to 2025 demolition permits actually constitute the frame, once garage-only records are struck out and semi-detached and other dwelling demolitions are added in.
  • How many self-contained dwelling units sit on each replacement title, and which of them are separately titled.
  • How often a legal-description join loses a subdivided replacement, and whether the lots it loses differ from the lots it keeps.
  • The distribution by demolition year, and whether the move toward rowhomes of five units or more after Zoning Bylaw 20001 pulls the per-dwelling median down for the later cohorts.
  • The result on the frozen 2023-12-31 mature-neighbourhood boundary, and on the City Plan redeveloping area.
  • Whether the demolished houses were owner occupied or rented, at what rent, and whether any tenants were displaced.

Missing evidence

Records we asked for and have not received, or that do not appear to exist publicly.

  • The matched series itself: each demolished house's assessed value on the roll in force the year before demolition, set against every replacement dwelling unit's assessed value on its first full roll after completion, with the dwelling-unit count on each title, across every house demolition permit from 2016 to 2025. By the freeze date, 2026-09-02, no City report, open dataset or third-party study had published one (City of Edmonton Assessment and Taxation with Urban Planning and Economy; critical).
  • A reproducible frame of every residential demolition permit for 2016 to 2025, carrying parcel identifiers, the prior dwelling count, fire and safety orders, and the unmatched cases, with garage-only records excluded (City of Edmonton Urban Planning and Economy and Development Services; critical).
  • Parcel and title lineage carrying each demolition through subdivision, consolidation and address change to every replacement title and assessment account (Alberta Land Titles and AltaLIS with the City of Edmonton; critical).
  • Completion and occupancy dates for permits before 2021, enough to pick each replacement's first full assessment roll without a heuristic that depends on the outcome (City of Edmonton permit and inspection systems; critical).
  • Verified self-contained dwelling counts and title structure for every replacement building, read from the assessment side rather than inferred from the permit units-added field, together with the land and improvement split of assessed value (City of Edmonton Assessment and Taxation with Alberta Land Titles; high).
  • Arm's-length sale prices and dates on both sides of each pair, which the predeclared sale-to-sale sensitivity needs (Alberta Land Titles at Service Alberta, and REALTORS Association of Edmonton MLS records; high).
  • The section 814 appendix boundary as it stood on 2023-12-31, as a parcel-level layer rather than a neighbourhood list (City of Edmonton; moderate).
  • Prior tenure, prior rent, tenant notices and any recorded displacement for the demolished houses. The run identified no complete public registry or source for these (property owners and tenants, City of Edmonton; moderate).
  • First arm's-length sale prices and dates for every separately titled replacement dwelling, linked to completion and title history. This is the brief's primary price basis and it covered zero cases (Alberta Land Titles at Service Alberta, and MLS records held by the REALTORS Association of Edmonton; critical).
  • Title form and a verified dwelling count for every replacement building, which is what separates the verdict denominator from the excluded single-title units (Alberta Land Titles and City of Edmonton Development Services; critical).
  • A retrievable source for the declared City median household income. The two Census Profile URLs cited for it returned a File not found page, so the run's primary income input is not in the registry (Statistics Canada; critical).
  • A predeclared method for deriving price-month five-year conventional mortgage rates after the Bank of Canada ended the monthly series in 2019 (YEGFacts editors, using the Bank of Canada weekly series; critical).
  • Property-level condominium fees for every condominium-titled replacement (condominium corporations and listing records; high).
  • Contract rents and their utility terms for replacement units held in single-title buildings, without which rental affordability cannot be tested at all (landlords; CMHC does not collect utility inclusion; moderate).
  • Bedroom counts per replacement dwelling, for the bedroom breakdown the brief requires (City of Edmonton Assessment and Taxation, and builders' listings; moderate).
  • Same-neighbourhood median sale prices by year, for the neighbourhood comparator the brief names (REALTORS Association of Edmonton; moderate).

AI review

Three models research each claim independently, then cross-review each other’s findings. The canonical finding is computed from the blind round-one verdicts by a deterministic rule published in advance; the models never vote and never decide. Cross-review documents their errors, and what it changed is shown below rather than folded into the finding. Authority comes from the methodology, not from the models.

ClaimClaude Fable 5.1GPT-5.6 SolGemini 3.1 ProPanel finding
When an Edmonton house is demolished for infill, is what replaces it typically worth about three times as much?Not establishedModerateNot establishedModerateNot establishedHighNot establishedUnanimous panel
Would the housing built on Edmonton teardown lots cost a median-income household 30 percent or more of its income to buy?SupportedHighNot establishedHighNot establishedHighNot establishedSplit panel
Reviewer notes — When an Edmonton house is demolished for infill, is what replaces it typically worth about three times as much?

Claude Fable 5.1

  • Neither figure can be classified, because no series on the brief's rules was produced here or found published, and the predeclared mapping lands on Not established. The inputs sit in the open: about 5,672 candidate house demolition permits for 2016 to 2025 with legal descriptions, and assessment rolls keyed by legal description back to 2012.
  • The permit dataset changed how it codes demolitions in 2019, so the frame has to be read out of the job description text rather than the job category, and some rows coded as a house describe a demolished garage.
  • In a 40-permit probe, matching on legal description alone recovered the demolished house in 35 cases against the 2018 roll but reached the replacement lot in only 17 against the 2025 roll; an address lookup returned three of the missing ones as two separately titled dwellings under new plan numbers. The eight lots that came through give approximate constant-dollar ratios of about 1.37 to 3.94, which are cases and not a distribution.

After cross-review: Held Not established and raised its own confidence from Moderate to High, holding it on a different ground than one other seat: the record can yield the series and nobody has built it, which is not the same as the record being unable to yield it.

GPT-5.6 Sol

  • No source reports the city-wide dwelling-unit denominator, a conforming ratio for every classified replacement dwelling unit, or bounds taking in every unclassified one, so neither primary figure and neither declared alternative can be classified as met or failed.
  • The published permit schema has no reliable completion dates before 2021 and does not say how many dwelling units a replacement title holds, and joins on address or legal description break where a demolition is followed by subdivision or retitling.
  • The qualitative City evidence bears on development economics and housing costs rather than on the predeclared figures, and the $350,000 to $1,000,000 example arrived with no address and no source and was never treated as a cutoff.

After cross-review: Stayed at Not established and moved its confidence up one step, Moderate to High, documenting that the probe is a convenience sample of one month valued on the wrong roll, that the permit filter is not a completed frame, and that net uplift per home added cannot be converted into the ratio.

Gemini 3.1 Pro

  • Historical property assessments sit in one City dataset and demolition and building permits in another, and no published dataset ties a demolished property's value to the values of the particular units raised in its place.
  • Without that linkage run across the frame, and without the mandated constant-dollar adjustment, neither the typical size of the gap nor the share of replacements at double can be evaluated.

After cross-review: Held Not established and lowered its own confidence from High to Moderate, accepting that the raw records do exist separately and taking up the probe's demonstration that a join without an address or spatial fallback drops subdivided lots.

Reviewer notes — Would the housing built on Edmonton teardown lots cost a median-income household 30 percent or more of its income to buy?

Claude Fable 5.1

  • Priced at assessed value on the first verified full roll, 2,968 of 3,359 separately titled replacement dwellings, 88.4 percent, would need 30 percent or more of the modelled household's income, with predeclared bounds of 82.8 to 89.1 percent, so the lower bound clears the threshold by a wide margin.
  • The median modelled price is $638,000 and the median modelled shelter cost $3,948 a month against a $2,250 limit. The component medians, $3,070 principal and interest, $494 property tax and a $377.67 utility allowance, are separate distributions and do not add to it: they sum to $3,941.67.
  • 3,244 dwelling units in single-title buildings, 49.1 percent of the classified replacement units, sit outside the verdict denominator, and their rental affordability cannot be tested because no contract rents are published and CMHC average rents make no adjustment for utilities.

After cross-review: Held Supported and answered the mortgage-rate objection, showing that its last-weekly-observation rule reproduces the definition of the monthly Bank of Canada series that was discontinued in 2019.

GPT-5.6 Sol

  • No public source identifies the complete set of separately titled replacement dwellings linked to eligible demolitions, and City assessment data do not identify title form, so the verdict denominator cannot be separated from the excluded single-title units.
  • New-home price evidence is directionally consistent with a purchase-affordability problem for a $90,000 household, and the City's own affordability report puts that income below what it estimates is needed for benchmark detached and semi-detached product, but neither isolates teardown-linked replacements.
  • The model cannot describe what any occupant paid: tenure, actual purchaser income, actual utilities and actual occupant shelter costs are all unavailable.

After cross-review: Held Not established, calling the model strong directional evidence that still inherits a nonconforming frame and prices every dwelling at assessed value without establishing that no first arm's-length sale existed.

Gemini 3.1 Pro

  • The denominator of separately titled replacement dwellings cannot be identified without the matched dataset the price-gap claim also lacks.
  • Its round-one position was that without the purchase price of each replacement dwelling the shelter-cost test has nothing to run on. The frozen brief expressly directs the total assessed value on the first full roll where no sale is recorded, and another seat ran it, so this is recorded as a reviewer error rather than a finding about the record.

After cross-review: Moved from Not established to Supported in round two, accepting that another seat's use of assessed value as the price basis, which the brief directs where no sale is recorded, produced an 82.8 percent lower bound clearing the 50 percent threshold. The canonical finding is computed from round one, so the movement does not change it.

reviews/infill-prices/2026-09-02-magnitude
reviews/infill-prices/2026-09-01-rerun2

That directory also holds the raw round-one and round-two JSON from every reviewer, the merged evidence and the synthesis output.

All three seats ran at each vendor’s high reasoning setting, the highest level the three CLIs share; the manifest records it per run.

How it was registered

Registered

How far the work has got
Gate complete
Whether it is worth checking
Going ahead
What a reader can see
Withdrawn

Checked before triage existed, three runs, both claims Not established. A triage read then parked both: the price gap needs a decade-scale lot-by-lot linkage the run did not build, and no public record says what the housing that replaced particular demolished homes costs. The story is no longer a finding until one of those exists.

What the states mean

Article history

  • updatedPresentation (methodology v1.18): each claim now carries its own one-sentence answer, in the words a person would say, and the standfirst was rewritten to hold one idea per sentence. Every fact the old standfirst carried is still on the page, in the TL;DR and the body. No change to findings, facts or evidence.
  • updatedWithdrawn from the findings board under methodology v1.13. Both claims are Not established, and a triage read of them (intake/candidates/infill-teardown-350k-1m/triage.md) returned PARK for each: the price-gap claim needs a decade-scale lot-by-lot linkage the run did not build, and the affordability claim asks about the cost of the housing that replaced particular demolished homes, which no public record gives. The page stays for the record; the story is listed with every claim considered, and reopens if either record appears. Nothing on the page was changed.
  • updatedThe price-gap claim was re-run under the magnitude proposition, about triple, after the 2026-09-02 editorial note. The brief fixed two figures before any result was seen: a median ratio of at least 2.5 between a replacement dwelling unit and the house it replaced, with a declared alternative of 2.0, and more than half of matched replacement dwelling units at a ratio of at least 2.0, with a declared alternative of 1.5. All three seats returned Not established in both rounds, no seat moved its verdict, and the finding is unchanged. Confidence moved without moving any verdict: the Claude and GPT seats from Moderate to High, the Gemini seat from High to Moderate. The claim record, its key facts and its evidence list now come from reviews/infill-prices/2026-09-02-magnitude under methodology v1.10; the affordability claim was not re-run and its published finding stands. Faithfulness check: 15 items from the GPT and Gemini seats applied, both seats FAITHFUL on re-check. Gate audit: source verification checked 62 statements against archived bytes, 5 imprecise on the first pass, all corrected, 62 verified on the second; freshness audit found nothing material; release check found eight case-level property records identifying specific homes in the raw reviewer output; they were redacted to case labels and value ranges before publication and the redaction is recorded in the run errata; clean on the second pass. The faithfulness and gate results will be recorded in this note.
  • updatedEditorial note on the price-gap claim: the proposition the panel was given, a majority test asking whether over half the replacement dwellings came in above the old house once both values sat in constant dollars, tests direction rather than magnitude, and the editor judged afterwards that a verdict on direction tells the reader little about the claim in circulation. The circulating claim is about magnitude: a modest house replaced by housing priced near a million. Four framing-check reports and an escalation neutralised that out of the brief and the editor adopted each step. The founder, reading the published page, called it 'that's poor claim to make', and directed that the process catch such claims at the brief. The claim is being re-briefed as the magnitude proposition and re-run; the finding stays as published until the re-run replaces it, because its substance, that the public record cannot show what teardown infill costs relative to what it replaced, holds under either framing. Recorded against the framing check and the editor in the panel quality ledger; the check now tests for this failure (methodology v1.10).
  • publishedInitial publication. First story produced under a delegated brief (decision D-0019, methodology v1.7 to v1.9): the founder supplied the raw post; Stew, the AI steward, wrote the brief as editor, a different-vendor framing check passed it on its third report, and the panel ran without the founder reading or approving any of it. Two earlier runs on this story were halted by the panel on framing concerns before this one cleared framing check 3. Faithfulness check: 18 items from the GPT and Gemini seats applied over four passes, both seats FAITHFUL. Gate audit: source verification checked 103 statements against archived bytes, found 3 unsupported and 15 imprecise on the first pass, all corrected, 103 verified on the third; release check clean after one local-path fix in a committed report; freshness audit found nothing material. The gate records its audit results in this note.

Every revision is in the public commit history of this file.