Assess each system against the DADM, with evidence a reviewer can check
Tovris works out which requirements of the Directive on Automated Decision-Making apply to a system, checks each one against the evidence you linked, and returns one of five results with a confidence score and citations. A person reviews every finding before it counts.
- Who uses it
- Assessor · Reviewer
- Instruments
- DADM · AIA · PSD
- Starts when
- An assessor runs an assessment
- In the console
- Assessments · Review queue
- DADM 2025 requirements apply at level III
- 39 of 45
- Samples per requirement behind each confidence score
- 3
- Possible results, including “unable to determine”
- 5
From the build and the simulated Government of Canada test environment · synthetic data
Where the usual process breaks
Checking a system against the directive by hand means reading 45 requirements, working out which apply at its impact level, and finding the evidence for each. It is slow, and two people rarely reach the same answer.
- 01
Appendix C requirements change with the impact level, and it is easy to apply the wrong set.
- 02
Conclusions cite whole documents, not the passage that supports them.
- 03
Uncertain answers get recorded as yes or no.
- 04
Later, nobody can tell which edition of the directive a conclusion was based on.
Four steps, each one a screen and an API route
Each step is part of the product, and each is written to the audit log as it happens.
- 01
Decide what applies
The applicability rules select the requirements for the system's AIA level and attributes, and store the reason for each. Without a final AIA, level-specific requirements are kept and marked undetermined.
- 02
Gather evidence
Assessors upload documents or link existing evidence. Each file is hashed, attributed to the person who added it, and its text extracted for search.
- 03
Assess
The assessment agent checks each requirement against the retrieved passages. Every citation must exist in the evidence; an answer that fails the checks is never stored.
- 04
Review
Findings land in the review queue, lowest confidence first. Reviewers validate, reject, modify or annotate them, and the model's original answer is kept.
assessment run
Applicability
level III · 39 apply
Evidence
hashed · cited
Assessment agent
3 samples
Confidence
calibrated
Review queue
a person decides
What you can show
A finding per requirement that a reviewer can check quickly, because the evidence is one click away.
One of five results per requirement, with an explanation in English and French.
Citations that open the exact passage in the evidence.
A confidence score from the model, citation support and agreement across samples.
“Unable to determine”, with the reason, when evidence is missing.
A run record with the model, region, prompt version and policy version used.
In the console
- System detail: run an assessment
- Assessment results for each system
- Finding detail with citations
- Review queue
Through the API
- POST /api/v1/assessments
- GET /api/v1/findings/{id}/trace
- GET /api/v1/review-queue
Instruments in scope
Who does what
Assessor · Reviewer. Every action is checked against the person’s role on the server and written to the audit log.
The rest of the loop
Inventory
Know every AI system your department runs, in the register's own format
Start from the public Government of Canada AI Register, add the governance fields it lacks, and keep one record per system that owners maintain and auditors trust. Tovris imports the register, reports what is missing and exports it back in the published format.
Admin · Owner · GC AI Register · AIA
- Register entries imported, 0 failures
- 412
- Organizations in the published register
- 43
- Published fields, kept under their own names
- 24
Policy change
When a directive is amended, see every system it affects before you publish
Tovris compares the old and new editions requirement by requirement, shows which systems each change affects and why, and opens reassessment tasks for the systems already assessed. Only the changed requirements are assessed again.
Admin · Owner · DADM 2023 → 2025
- Requirements added by the June 2025 DADM amendments
- 11
- Requirements modified
- 34
- Requirements retired
- 6
Risk monitoring
Catch drift and unfair outcomes between assessments
Tovris watches data quality, input drift and output measures for each system. When a measure crosses its line, the result is filed as evidence, findings are raised on the linked requirements, a reassessment task opens, and your ticketing system hears about it.
Owner · Assessor · DADM 6.3
- Default drift alert line (Population Stability Index)
- 0.2
- Injected breaches caught in testing, no false alarms
- 2 of 2
- Ticket priorities on the signed webhook
- P1–P4
Walk through this on synthetic data
We run this use case end to end on the simulated Government of Canada environment, in English or French, and answer questions as we go.
Live walkthrough · English or French · Synthetic data only