Checklist: Risk Classification

Running a scoring session that produces the same tier whoever runs it, and a record that survives a challenge.

Classification is where implementations most often go quietly wrong, and the cause is almost never bad faith. It is anchors applied loosely, in a document, by one person who also owns the delivery date.

This checklist assumes forty minutes and two people. The full reasoning is in the scoring guideline.

A completed checklist is not evidence. The evidence is the record it told you to complete — the templates hold those.

Before the session

The use-case canvas exists and has been read by both scorers
The second scorer does not carry the delivery date for this system
The relevant sector overlay section 3 has been read (twenty minutes)
The decision the output supports is written in one sentence
Someone can answer what the system does when it is wrong, not only when it is right

D1 — Decision consequence severity

Scored on the consequence of being wrong, not the value of being right
Scored assuming the output is acted on — the human catch belongs to D2
Justification names who is harmed and how
The 3-versus-4 test applied: serious harm, or an inconvenience they could get fixed?
If D1 = 4, the safety-override floor is noted now, not discovered later

D2 — Autonomy

Scored against the highest-autonomy path the system has, including the routine bulk path
Straight-through, auto-close and auto-approve thresholds identified and counted as autonomy
Where a human reviews: the override rate is measured, or the floor is agreed today
Friction symmetry considered — is departing harder than accepting?
Planned autonomy trajectory recorded, so a later increase is a planned change

D3 — Reversibility deficit

Scored on consequence reversibility, never on mechanism reversibility
Asked explicitly: who has already seen, received or relied on the output?
Money movement, disclosure, communication and physical effect treated as irreversible
Appeal or complaint routes not counted as reversibility
Elapsed harm counted: a decision reversed after four months did not undo the four months

D4 — Exposure and scale

Counted the reach of the records the system writes into, not the login list
Cascade considered: which downstream systems or decisions consume this output
Scored on the deployment path, not the pilot population
External exposure identified, including indirect exposure through a partner

D5 — Sensitivity and uncertainty

Both sub-scores recorded separately
Composite taken as the higher of the two, not the average
Sensitivity assessed on the joined record, not source by source
Uncertainty scored honestly for generative components, where omission is the failure mode

Computing the tier

Impact = average(D1, D4, D5), rounded up — arithmetic shown in the record
Control Deficit = average(D2, D3), rounded up — arithmetic shown
Matrix cell recorded, not only the resulting tier
Override floor checked deliberately: D1 = 4, or D2 = 4 with D3 = 4
Result sanity-checked against a comparable system already classified

Countersignature and after

Countersigner was present, or scored independently and compared Risk Lead
Adjustments made by the countersigner are recorded, including none Risk Lead
Disagreements recorded with both positions and who decided Risk Lead
Effective-autonomy threshold and measurement method recorded System Owner
Re-score triggers noted, with the next scheduled review date Risk Lead
Previous scores preserved, not overwritten Risk Lead
Classification published where other teams can see it and argue with it AI Governance Lead

Re-score triggers — check quarterly that these are wired

Any change classified Material or Major
Any autonomy increase, of any size
User population change
New or changed grounding source at or above current sensitivity
Control Plane risk-relevant drift alert
Scheduled review: annual minimum at Tier 3–4
Override-rate evidence contradicting the recorded D2