What we count, and what we can't
Every figure is extracted from the four corners of a published decision and carries a confidence score: stated verbatim, unambiguous inference, or inferred from an index. Nothing is imputed — a silent field is recorded as silent, never as a zero.
Rates are calculated by matter — one person's case, verified by HRTO file number — not by decision. One matter can produce a dozen decisions, and counting decisions overweights the most litigated files. Where only an index is available (no file numbers on its face), we count decisions and say so.
Published decisions are a survivorship sample. Matters that settle or are withdrawn may produce no public document, so every rate we publish is a rate among adjudicated matters. An exit is not a verdict.
The Visibility Gap
The core research question: when a person enters a legal or administrative system, what outcomes become visible — and what outcomes disappear? Public decisions show the cases that reach a formal conclusion; they reveal far less about the cases that exit earlier through settlement, withdrawal, procedural dismissal, or outcomes that never generate published reasons. Quantifying that gap — and closing it through access requests — is the project.
What we refuse to publish
- Success rates as merit rates. The record cannot distinguish "claim lacked merit" from "claimant lacked capacity to continue."
- Raw-count league tables of respondents. Counts without population and calls-for-service denominators, and without matter-level identity data, mislead.
- Settlement claims without the settlement flag data. Privilege protects terms; the facts we cite are counts, and we do not have them yet.
- Untested allegations against named individuals. Institutions and patterns, not people; findings, not accusations.
Corrections · in force from day one
Tribunal-only datasets have a known failure mode: a tribunal does not annotate its decisions when a court later sets them aside. That trap surfaced in our own pre-publication review — an outcome initially read backwards was caught by joining court records — so every outcome we publish now requires that join before it ships, and a validation rule blocks any figure without it.
Pre-publication validation has also already caught: a mislabelled survival statistic (screening vs all-gates rates, both now published with their definitions), and an index revealing 8 earlier decisions (2009–2010) outside our corpus window, now queued for ingestion with the window statement corrected.
August 2026: an earlier draft of Case Study 01 stated its corpus window as complete at 52 decisions (2011–2026). The province-wide index revealed 8 earlier decisions (2009–2010); their full texts have now been obtained, read and ingested, and the case study's totals were revised to 60 decisions / 2009–2026, adding two early-era merits hearings to the record. This entry replaces the prior "queued for ingestion" notice.
August 5, 2026 — verification pass before the second release: (1) The claim that the former Thunder Bay chief's trial "began May 5, 2026" was corrected: reporting shows May 5 was a hearing at which the Crown confirmed the trial will proceed on fewer counts, her case having been severed after a medical delay; trial dates were not yet set, and no verdict exists. (2) The former service lawyer's outcome was stated more precisely: acquitted of one charge April 17, 2026 as the Crown rested (no evidence presented), found not guilty of the remaining charges April 24, 2026. (3) A cohort outcome was updated: one member-applicant's reprisal claims were dismissed in June 2025 for lacking factual basis. (4) A draft ledger row on provincial discretionary spending was removed pending primary sources. (5) The Durham ~$2M legal-spend and $65,000 costs-order figures were verified against the original CBC reporting before publication.
When we publish an error, the correction will appear here: dated, permanent, in full, with what changed downstream.
AI, and its limits here
AI assists extraction from decisions into the schema (confidence-scored, human-verified before publication), entity-resolution candidates, and drafting. It never produces outcome predictions for individuals, personalized legal guidance, or unreviewed published text.
Class-action indicators — an analytical schema
Applied to every documented pattern, marked clearly as research analysis, not legal advice: pattern_commonality (does one institutional practice — policy, code, template, clause — recur across affected persons? e.g., settlement non-disparagement templates; PTSD-finding appeals as reported blanket practice; unfounded coding) · class_size_signal (countable affected population from records) · evidence_base (institutional documents proving the practice — self-generated records beat testimony, this site's founding insight inverted to plaintiffs' benefit) · precedent_track (nearest certified/settled analogue: the G20 mass-detention class; profiling and strip-search certifications) · limitation_horizon (the clock, per claim type — because Chookomolin) · venue_logic (why class/civil rather than, or alongside, individual Code applications: the s. 35 vacancy, the personal-respondent removal practice, and the attrition documented across this site are themselves the argument that individual process cannot reach systemic conduct).
First three patterns to score: the Durham PTSD/WSIB appeals practice; settlement-silencing clause use; Thunder Bay Indigenous service-docket attrition.
Standing on prior work
Province-wide analysis of HRTO dismissal rates, abandonment trends and annual-report gaps has been led since 2020 by Tribunal Watch Ontario, a non-partisan group of retired adjudicators, counsel and administrative-justice experts. OnionSkin's respondent-level relational data and court-record joins build on the ground their aggregate work established. The 2012 Ontario Human Rights Review (Pinto) is our baseline for what this system looked like when it worked — and the source of a still-unimplemented recommendation to publish anonymized settlement-outcome data.