The Ontario Platform Baseline

What 4,743 researched candidates across Ontario have actually written down — the province-wide extension of the Toronto Platform Baseline, companion to State of the Election — Ontario. This is the completed edition: corpus as of the August 10 Friday-close plus the August 10–11 recovery-and-regrind legs, with Toronto scored fresh alongside the rest of the province under METHODOLOGY v2.1 for the first time, and — as of August 11 — every candidate with any captured content individually read under the same instrument. The receipts and method are below.

What was measured

Ontario's candidate registry carries 4,751 municipal candidates; 4,743 were researched to a dated conclusion (the remaining 8 are list-only, closing at nomination day): website verified, or a recorded search that found none. Of the 1,280 with a site on file at the capture snapshot (13 wrong-attribution website values were later removed on re-verification — receipt §11 reconciliation; corrected count 1,267), 984 were captured and indexed; the assessable platform corpus — every candidate with at least 50 words of real platform content — totals 2,359,387 words across 484 candidates province-wide. Every assessable platform was scored under the same versioned, viewpoint-neutral instrument used for Toronto: eight anchored dimensions — specificity, fiscal grounding, jurisdictional fit, evidence, implementation, local groundedness, coherence, time-boundedness — plus a new commitment gate (G5) that separates forward-pledged commitments from record and aspirational content, plus issue breadth scoped to the office sought (a trustee is never measured against transit). Every material score carries a quote from the candidate's own text. A stratified 249-candidate blind double-rating (20.2% of the primary set, across both scoring sittings) audits the whole exercise (results below). And as of August 11, the count is complete by construction: every candidate with any captured content received an individual class-first read — the 710 sites a mechanical word-count filter had classed “brochure” unread were each read in full by the same instrument, and 296 of them (41.7%, including a same-day remediation pass that re-fetched 29 defective captures) turned out to be something else, including 168 real scoreable platforms. No number on this page rests on a word-count threshold alone. No position was judged for its direction, anywhere, by construction — a fully specified plan to build X and a fully specified plan to remove X score identically. The only adverse finding this instrument can produce is absence or thinness of published substance.

The headline — the ballot is mostly blank

METHODOLOGY v2.1 reports the field in its full class scheme, both cities scored fresh together and every class carried by an individual read or a receipted disposition — reported here as one province-wide count (n=4,743 researched):

RETROSPECTIVE-RECORD (150 candidates, 28 of them incumbents): a real public document — usually a term-in-review or delivery record — reported here on its own terms rather than scored as a forward platform it was never written to be. STALE-CONTENT (42 candidates): platform text is on file, but for a different office or an earlier cycle — real writing, just not a readable answer to this race.

85.9% of the 4,743 researched candidates province-wide currently offer the public nothing readable about what they would do in office (NO-PUBLISHED-PLATFORM or BROCHURE, combined). Toronto was re-scored fresh alongside the rest of the province this edition, under the identical v2.1 instrument, on the same day. The desert deepens by region and municipality size until, under 5,000 people, published platforms effectively do not exist.

Fairness, stated here as everywhere: absence of a published platform is a fact about the public record, not about a candidate's effort, character, or worth — much of small-town Ontario campaigns door to door, and ten days remain in which any candidate can change their row in this table. We will gladly re-measure.

The substance profile — Ontario knows WHAT and WHOSE JOB; almost nobody says WHEN or WITH WHOSE MONEY

Across the 416 scored platforms province-wide — both cities scored fresh together under v2.1, every word-count-gated site individually read in the August 11 completion pass — the dimensions tell one story:

• Strongest: jurisdictional fit (4.22/5, median 4.0). The folk theory that municipal candidates promise federal miracles is false province-wide. Candidates overwhelmingly campaign inside their office's actual powers — a genuinely encouraging baseline fact.
• Middling-strong: local groundedness (4.19), coherence (3.66), specificity (3.65). Platforms name real local places and hold together as lists; named actions usually arrive without mechanisms.
• Weak: evidence (3.18), implementation (2.68), fiscal grounding (2.58). Claims without data; money without numbers; almost no sequencing.
• Weakest, everywhere, again: time-boundedness (1.30/5, median 1.0). A dated, staged commitment is still the rarest object in Ontario's platform corpus. Almost nobody in the province will tell you what happens in their first hundred days.

Of the 416, in bands: 7 nominal · 50 thin · 157 workmanlike · 122 substantial · 80 exceptional. The strict recognition-campaign floor — no substance signals at all — is 0 of 416 scored (0.0%): the v2.1 commitment gate (G5) now routes recognition-only content out of SCORED before it reaches the band count, reported in aggregate as always.

Office, region, incumbency

The breakdown below is carried from the v2.0 baseline as lineage. The v2.1 scored set (n=416) is differently constructed — the commitment gate routes record and aspirational content to their own classes before this cut is taken — and its own office/region/ incumbency slice publishes with the next edition. This edition's own v2.1 incumbent read, for the record: only 12 incumbents province-wide clear the scored bar (mean composite-7 3.58), while 150 candidates — 28 of them incumbents — publish delivery records (RETROSPECTIVE-RECORD) instead of forward platforms.

• Office: mayors who publish, publish best (scored mean 3.2/5); councillors 2.9; trustees 2.5 — the school-board dark zone holds at the substance layer, atop a field where 29% of trustee races have zero candidates and 9% of trustee candidates have any platform page.
• Region: among scored platforms, Eastern Ontario leads (3.4), Hamilton-Niagara 3.2, Central 3.1, Northern 3.0, GTA-905 2.8, Southwestern 2.8 — read with the selection caveat below: where fewer publish, the few who do skew serious. The North has the thinnest coverage (7% websites) and its scored few carry the widest issue breadth.
• Incumbency — the finding with teeth (aggregate, as it must be): of 271 known incumbents outside Toronto, 255 (94.1%) publish nothing readable (212 nothing findable, 43 brochure-class sites). The 13 who publish score a solid 3.3 average — but sitting officials as a class are asking for renewed mandates on name and record alone, and challengers as a class out-publish incumbents, everywhere we have measured.

The exceptional platforms — praise only, three gates before any name

80 of 416 scored platforms cleared the instrument's top band. Nobody gets named on a score: every one was independently re-read in full, and then held to this project's strictest test — a verbatim policy commitment from the candidate's own platform: a pledged future action with a number, mechanism or deadline attached. The 74 candidates who cleared all three gates — the best published platforms in Ontario, by region, each linked and each quoted on a commitment — have their own page. The rest of the top band stays in the aggregate count only — delivery records, older-cycle pages, or platforms whose pledges stay hedged; absence from the named list is not criticism. Five standouts, each quoted on a commitment (measured on published substance only — not viability, endorsement, or agreement with any position):

How much to trust the scores

Scoring was done by AI raters under the anchored public rubric, disclosed as always — and audited: a stratified 249-candidate sample (20.2% of the 1,233 primary rows, across both scoring sittings) was independently re-scored blind by different raters, drawn across every lane. Class agreement (5-way, post-adjudication scores-of-record): 226 of 249 (90.8%). Commitment-extraction agreement: 94.8%. Among the 77 pairs both raters scored, per-dimension agreement within one point ran 84–99% by dimension; average composite-7 divergence: 0.39 (median 0.29), with 3 pairs diverging more than a point. Two numbers stated by us before anyone asks: across the full 249-pair sample, raw class agreement before any adjudication was 207 of 249 (83.1%) — the 90.8% is the post-adjudication scores-of-record figure, and both belong in the open. Within the completion pass's own 144-pair subsample, raw agreement was 79.9% — expected-lower, because that sample sits exactly on the ambiguous brochure/record/platform boundary the retired word-count filter had hidden, and its composite divergences over a point numbered zero. Every class disagreement across both sittings — 46 rulings in total — was re-read and adjudicated by the lead against the candidate's own text on 2026-08-11 (primary upheld 26, dual upheld 20); the full reasoning publishes with the data. One honest number, stated by us first: a separate 50-pair blind cross-edition check re-rated v2.0-scored corpora under v2.1 — mean composite-7 drift of +0.09 (median +0.14) among the 23 pairs that stayed SCORED, while 27 of the 50 moved class entirely under the new commitment gate (19 to BROCHURE, 6 to RETROSPECTIVE-RECORD, 2 to STALE-CONTENT) — the gate doing exactly what it was built to do.

The growth loop — what happens next

This baseline exists to be beaten. After nominations close August 21, we recapture everything — including the 153 candidate sites that appeared after our capture window and the 143 that were assessed dead or parked — score every new and changed platform, and publish the deltas in aggregate: how many candidates added costed planks, how far time-boundedness moved, how much of the ballot stopped being blank. Then every two weeks until the election. Candidates who publish after being measured are not gaming the system; they are the system working. Publish something real and tell us — re-measurement is free, fast, and gladly done.

Method & limitations — ours to state first

The instrument is public and versioned: eight 0–5 dimensions with written anchors, a commitment gate (G5) that separates forward-pledged commitments from record and aspirational content — the source of this edition's two newly-reported classes, RETROSPECTIVE-RECORD and STALE-CONTENT — office-scoped issue-breadth (12 buckets for mayor/council, 8 for trustee), quotes required behind every material score, "low confidence" flagged rather than guessed, French platforms scored as-is and never penalized for language. Web-only record — print and doorstep literature are invisible to us, and we say so. Ten days pre-close — the field will grow; that is why this is a baseline. Forty-nine unassessable/unreadable sites (UNASSESSABLE) are held as their own class, never counted as absent. AI raters under anchors, quotes, blind double-rating and lead adjudication — disclosed, reproducible, not infallible. Issue taxonomies are versioned choices. Twelve miscaptured rows (MISATTRIBUTED-CAPTURE) and eighteen retracted/flagged captures were held out of scoring rather than guessed at; a further four wrong-website registry errors were caught by this analysis's own cross-checks and corrected before publication — verification working as designed. And the standing truth: a great platform is not proof of a great officeholder — it is the part of the offer a voter can actually read, and the part a province can demand more of. That demand is the project.

Receipts — where every number comes from

• The official candidate lists of all 444 Ontario municipal clerk's offices (August 10, 2026 collection — this weekend edition) — the candidate universe and submitted channels.
• Our per-candidate research record — 4,743 candidates researched to a dated disposition; 984 candidate sites captured and indexed (2,359,387 words of assessable platform text across 484 candidates); method at how the research is made and how we verify.
• The versioned scoring instrument and aggregate results — dimension anchors, breadth taxonomies, band definitions, reliability statistics and every aggregate on this page are published together; re-running the pipeline on the same corpus reproduces every number. Toronto figures: re-scored fresh alongside the rest of the province under METHODOLOGY v2.1 this edition, no longer reused from the original baseline — see the original Toronto Platform Baseline for the v2.0 lineage figures.
• Per-candidate assessments are never published. Aggregates only, committed before publication; the praise-only shortlist above is the single exception, by standing rule. No aggregate on this page describes a group of one or two identifiable people.

Found a mistake?

That's the system working. If any figure here is wrong, out of date, or reads differently than its source supports, send it to the corrections page — thirty seconds, and every correction is published, dated, and left visible. Candidates: if we have your site wrong, your platform missed, or you've just published — tell us and we will re-check and re-measure, gladly.