Disclosure: DROZlegal publishes this guide and builds the disclosure_review classification agent described below. This article is general information about Ontario criminal disclosure procedure, not legal advice, and does not create a solicitor-client relationship; it reflects Ontario and Canadian law, including a Supreme Court of Canada decision released this year, as of August 25, 2026.
What AI disclosure review software actually classifies
"AI disclosure review" gets used as a catch-all for anything that touches a Crown disclosure package — a search index, an OCR pipeline, a generic summarizer. DROZlegal's version is narrower and more specific: a single agent, disclosure_review, that looks at one document at a time and returns a classification along three axes, plus a short reasoning string a lawyer can read before accepting it.
Category, completeness, and a triage flag — nothing else. Category sorts the document as Stinchcombe first-party material, a McNeil police-misconduct record, or an O'Connor-eligible third-party record. Completeness flags whether what arrived reads as a full production or looks partial — a truncated transcript, a garbled OCR scan. A "hot" flag marks something for counsel to look at first. None of the three is a merit or guilt conclusion; the agent's own design brief is explicit that the hot flag means "look at this," never "this decides the case."
| Field | Values | What it means |
|---|---|---|
| Category | first_party / mcneil / oconnor | Stinchcombe first-party, McNeil police-misconduct, or O'Connor third-party disclosure |
| Completeness | received / partial | Whether the document reads as a complete production or looks incomplete |
| Hot flag | hot / not_hot | A triage signal for counsel to review first — never a merit or guilt conclusion |
| Case scope | matched, or refused as case_mismatch | A document from a different case is refused before any AI call, never reclassified into the wrong file |
Source: backend/app/services/agents/disclosure_review_agent.py (module docstring and classification logic), read directly for this article.
The Stinchcombe / McNeil / O'Connor split — and why one 2026 SCC case just moved the line
The three-way split isn't DROZlegal's invention — it's the actual shape of Canadian criminal disclosure law. R. v. Stinchcombe set the baseline: the Crown must disclose all relevant, non-privileged information in its possession. R. v. McNeil extended that first-party obligation to police misconduct records connected to the investigating officers. R. v. O'Connor covers material genuinely held by a third party, gated behind a higher two-stage test. A tool that can't tell these three apart isn't classifying disclosure — it's just labelling files.
The Supreme Court of Canada ruled unanimously that administratively expunged police misconduct records remain subject to the McNeil first-party disclosure regime, not the higher O'Connor third-party threshold — rejecting the argument that expungement removes the disclosure obligation. Source: Edmonton (Police Service) v. McKee, 2026 SCC 24 (decided June 26, 2026); coverage via Canadian Lawyer, June 2026.
That boundary moved this year, not in theory. For a classification agent, that's not academic. A record an older heuristic might have filed as O'Connor-tier — expunged, so "obviously" third-party-adjacent — is McNeil first-party material after McKee, with the lower disclosure threshold that implies. disclosure_review's system prompt encodes the McNeil/O'Connor split as currently understood; it's a live legal boundary, not a fixed lookup table, and it's exactly the kind of line a lawyer should independently confirm rather than take a tool's word for.
Wagg/Juman per-case isolation
Crown disclosure obtained in one matter isn't free to use in another. The restriction runs through R. v. Wagg and Juman v. Doucette, 2008 SCC 8's implied-undertaking rule against collateral use of material obtained through one proceeding. disclosure_review builds that restriction into the classification pipeline itself, not just into firm policy.
Every job is silo-scoped to exactly one case_id, enforced before any AI model sees a document. The agent resolves the case a job belongs to, then checks every document against it. A document that belongs to a different case is refused — flagged with an explicit case_mismatch error — never silently dropped, and never reclassified into the running case's batch. That refusal happens before the document reaches Claude at all, not as a check applied after the fact.
Manual dispatch, propose-only — and where the hard ceilings actually sit
A lawyer starts every classification run. On the case file's disclosure workspace, dispatching disclosure_review is an explicit action — an audited API call carrying its own idempotency key — not something that fires automatically when a new document lands.
17 of DROZlegal's 22 registered agent classes are auto_dispatchable=False, and disclosure_review is one of them. The distinction matters: DROZlegal's six permanent hard ceilings — trust money movement, court filing, settlement, commencing litigation, engagement approval, and agent-initiated external communication — never applied to disclosure classification specifically, because classification was never designed to touch any of the six in the first place. What does apply, the same way it applies to every agent on the platform, is propose-only output: disclosure_review returns a category, a completeness flag, and a reasoning string for a lawyer to read. It doesn't file anything with a court, send anything to the Crown or opposing counsel, or reach any conclusion about the case.
Maturity is worth stating plainly rather than glossing over. Per DROZlegal's internal capability inventory, the criminal module — disclosure_review included — is complete but unused: the standby criminal design partner hasn't been activated yet, and no firm has run a matter through it. Everything above describes what the agent does on a test run, not a track record.
The disclosure workspace — override and completeness tracking
Classification output lands on the case file's disclosure workspace, where a lawyer reviews it item by item — and where the tool's read is never the last word.
Two actions, both audited server-side. A lawyer can override the category the agent proposed — the workspace records the change with classification_source set to "manual," distinguishing a lawyer's call from the agent's — and can update the completeness status directly as new material arrives or a gap gets filled. Both actions go through the same PATCH endpoint and the same audit trail as the rest of the case file.
"Every case is now much more complicated to prosecute and involves so much more data than it used to." — Donna Kellway, president of the Ontario Crown Attorneys' Association, on Crown staff being swamped with body-camera footage to review for court-ready evidence packages. Source: CBC News, coverage of the London Police Service body-worn-camera rollout, January 2026.
That volume problem sits on both sides of a criminal file — Crown staff assembling the disclosure package, and defence counsel sorting what lands. A classification pass that gets most of a batch right on the first read, with a workspace built to correct the rest in place, is a narrower and more honest claim than "AI reviews your disclosure" — and it's the one disclosure_review is actually built to make good on, once a firm puts it in front of a real file.
What to check before you adopt AI disclosure review software
Whatever tool your firm evaluates — DROZlegal or otherwise — these questions separate real disclosure classification from a labelled folder:
- Does it classify into the actual legal taxonomy, or just "important/unimportant"? Stinchcombe, McNeil, and O'Connor carry different disclosure thresholds; a tool that collapses them into one bucket isn't doing the legal work the category split exists for.
- Does it isolate disclosure by case, or risk collateral use across matters? Wagg/Juman-style per-case refusal — rejecting a mismatched case_id before any AI model sees the document — should be a design choice you can point to, not a promise.
- Does anything file, send, or decide without a lawyer? It shouldn't. Classification is a proposal; court filing, sending anything to the Crown, and deciding what a document means for the case all stay with counsel.
- Has a real firm run a live matter through it? A vendor should be direct about whether a claim describes a design or a track record — DROZlegal's own answer for this module, honestly, is design so far, not yet a track record.
- Where does the data live, and under what retention terms? Stored client data staying in Canada (DROZlegal: AWS ca-central-1) and AI processing that doesn't train on client data and auto-deletes within roughly 30 days is a specific, checkable claim — see DROZlegal's security architecture. "Zero retention" isn't a real category; press any vendor claiming it for the actual terms.
Disclosure classification is one piece of a criminal file. For the Jordan s.11(b) ceiling clock, charges and court-appearance tracking, and the rest of what a criminal-defence cockpit needs to do, see our broader guide to AI criminal defence software in Ontario. For how DROZlegal's review agents work outside criminal law, see AI legal document review software, or the product overview for the platform as a whole. New to how AI fits into a law practice generally? Start with the Lawyer AI Academy guide.
Not ready to subscribe? Join the DROZlegal waitlist instead.