Avoid it. Shrink it. Put it back. Four rungs on the ladder and it only goes one way. And there’s no rung on the ladder for the thing you can’t repay.
The mitigation hierarchy begins before damage. Avoid the impact if possible. Where avoidance fails, minimize it. Repair or restore what can be restored. Compensate for what remains.
The sequence appears throughout environmental and social practice because its order matters. The International Finance Corporation’s Performance Standard 1 favors avoidance over minimization and leaves compensation or offset for residual impacts. In conservation practice, the sequence is often rendered as four steps: avoid, minimize, restore, offset.
Each descent accepts a smaller return. Avoidance preserves the prior condition. Minimization concedes some damage. Restoration attempts to recover what existed. Offset supplies value somewhere else because the original condition cannot be recovered in full.
The hierarchy is a reversibility ranking hiding in familiar language.
A determination as impact
Apply the same sequence to an automated determination.
Avoidance asks whether the system needed to make the determination at all. A workflow can keep certain questions outside automation, especially where an adverse result would change a person’s access to work or public service.
Minimization narrows the determination’s reach. The output might inform a human judgment rather than execute one. It might expire quickly. It might remain barred from reuse in unrelated decisions.
Restoration is where the vocabulary begins to fail. Correcting a database row can restore the visible record. It cannot restore a missed deadline or an opportunity given to someone else. A later approval does not create the conditions that existed when the first refusal arrived.
Offset is stranger still. Money can recognize a loss and may fund recovery. The sharper test is whether compensation has identified the whole residual loss and valued what can be valued. A wrongful determination may leave measurable costs alongside lost time, opportunity, and standing, for which no agreed unit exists.
The cheap direction
Forward action is usually designed as a single transaction. A score crosses a threshold. A status changes. A notice goes out. Other systems consume the result because integration was the point of building the stack.
Reverse action must discover every recipient, determine which consequence followed, and establish authority to amend it. The original transaction may take milliseconds. Its remedy becomes a small institutional investigation.
This asymmetry is easy to miss in a process map. The forward arrows are explicit. The return routes appear as policy words: appeal, correction, review. A word on a diagram says nothing about whether the technical path exists.
The field already knows that an offset occupies the weakest end of the hierarchy. IFC’s 2025 management-system guidance states that compensation or offset does not eliminate the adverse impact. It seeks a comparable positive result elsewhere. That limitation is clear when the impact is a damaged habitat. It should be equally clear when the impact is a determination.
Classification before action
An impact assessment that treats only the model’s error rate as risk misses the irreversible part. The relevant unit is the action the workflow permits: refusal, referral, suspension, escalation, or accusation. Each action needs a reversibility class before anyone automates it.
Can it be prevented? Can its reach be contained? Can the previous condition be restored? If restoration fails, what loss remains even after compensation?
Those questions belong near the beginning of design. Once a determination has moved through several systems, the hierarchy has already collapsed toward its weakest rung.
One calibration belongs here before the descent continues. Remedy does not require recreating an untouched past. It requires ending the original action’s authority, correcting what remains reachable, preserving the evidence, and stating the residual loss for what it is. The rest of this arc measures remedies against that standard, because a changed status is easy to mistake for a completed reversal, and a remaining loss is easy to leave out of the account.
The first reversibility test concerns the action. The next concerns the material expected to disappear when someone presses delete. A person asks for data to be removed. The live record disappears. The model may still carry what it learned.
Companions
- The offset limitation in the lender’s own toolkit: IFC, Environmental and Social Management System Toolkit, General, on compensation and offset measures that do not eliminate the identified adverse risks and impacts.
- The hierarchy at source: IFC, Performance Standard 1.
- The fiction where the scale first appeared: Kafka, The Trial, Titorelli’s three outcomes, in the Muir translation used in Definite Acquittal.
- The last time the scale passed through: Public Eligibility.
- The synthesis this extends: Attribution at the Moment of Action.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.