Arc Consolidation | Episodes 256–262
Attribution at the Moment of Action closed with three conditions for accountable action: a named principal, engineered reversibility, and evidence held beyond the actor's exclusive control. The first tells us who must answer. The third gives everyone else a record from which to ask. Reversibility carries the consequence. It decides whether the person acted upon can recover anything after the answer and the evidence establish that the action was wrong.
That condition sounds almost administrative. Systems have pause buttons, rollback procedures, correction forms, deletion requests, and appeal routes. Each offers a verb that points backward. The difficulty appears when the verb meets the object. A paused system leaves its completed determinations untouched. A corrected row cannot recall copies already sent elsewhere. A successful appeal may change the official status after the missed opportunity has gone to somebody else. The interface moves backward while the consequence continues forward.
Definite Acquittal found the right taxonomy in Kafka's court. Titorelli offers Josef K. three outcomes. Definite acquittal removes the case so completely that no file remains to be reopened. Everyone has heard of it and nobody can produce an example. Ostensible acquittal frees the accused while the file stays ready for another judge. Indefinite postponement keeps the case near its beginning through continuous procedural motion. The first is perfect. The others are available.
Automated systems often produce ostensible acquittal with excellent paperwork. The adverse status changes. The record that supplied it survives in another database, a vendor log, or a model's training history. Indefinite postponement appears when the affected person must keep submitting evidence to prevent the same wrong result from settling again. Activity replaces closure. The burden of remembering the correction falls on the person whom the system forgot correctly.
The scale beneath the promise
Environmental and social practice already has a disciplined account of return. The mitigation hierarchy begins with avoidance, moves through minimization and restoration, and leaves compensation or offset for residual harm. The Scale That Only Runs One Way read that sequence as a reversibility ranking. Each descent accepts that less of the prior condition can be recovered.
An automated determination belongs on the same scale. Avoidance asks whether a system should make a consequential determination at all. Minimization restricts its reach or lifespan. Restoration tries to recover the previous condition. Compensation acknowledges the loss that remains. The action matters more than the abstract accuracy of the model. A recommendation that expires inside one session and a refusal distributed to several institutions may share an error rate while carrying radically different prospects of return.
The forward route is usually explicit because integration depends on it. A score crosses a threshold, a status changes, and another system consumes the result. The reverse route appears on the diagram as a policy word such as review or appeal. Executing it requires discovery. Someone must locate every recipient, determine what each recipient did, secure authority to amend the resulting records, and preserve proof that the correction arrived. A transaction completed in milliseconds becomes an institutional investigation on the way back.
This is why compensation cannot silently stand in for restoration. Payment may recognize part of a loss and support recovery. It cannot recreate a planting season, reopen a deadline, or return standing consumed by an accusation. A remedy has to name the residual loss rather than letting the existence of a payment close the account.
The disappearing object
Deletion Is Not Erasure followed a deletion request through the stack. The visible record is the easiest object to remove. Backups follow retention cycles. Vendors hold copies under other arrangements. Derived data may no longer resemble its source. Model weights do not contain a labeled drawer from which one person's contribution can be lifted.
Machine unlearning research tries to recover the state that would have existed without selected material, or a sufficiently close approximation under a defined test. The test carries the claim. Silence at the interface shows what a model currently outputs. It cannot establish by itself that the targeted influence has disappeared, especially where later fine-tuning can revive a concept that seemed to be gone.
Erasure therefore needs a map. The map begins with the source record and follows known copies and derived artifacts. It assigns an evidentiary standard to each layer. Where a layer cannot be reversed, the person should learn that before the data enters it. A satisfying confirmation screen is a poor substitute for an account of what was removed, what was placed beyond use, and what remains.
The life of a corrected record
The same propagation problem appears after an appeal. The Record That Reopens examined determinations that retain influence after their formal reversal. The official record may be corrected while derived scores preserve the old value. A downstream recipient may have acted before the correction arrived. Training data can carry the disputed finding into later systems.
The Dutch tax administration's FSV blacklist gave the problem institutional scale. Inaccurate entries could obstruct debt restructuring, and the state was still tracing which bodies had received the list. A correction inside the originating administration could not complete the remedy without discovering where the finding traveled. The appeal process sees the determination presented to it. A full reversal must also see the route the determination took.
Correction propagation deserves the same engineering attention as data distribution. Every consequential result needs a stable identifier. Recipients need a way to accept a withdrawal or amendment tied to that identifier. Operational systems must stop treating the overturned result as current truth while an independent historical record preserves the fact that the original action occurred. Destruction of the evidence would make the remedy harder to prove. Continued operational authority would make it incomplete.
Authority across the boundary
Who Is Allowed to Undo located the social architecture behind the technical gap. An operator may roll back a transaction. A vendor may patch the application. A model provider may retrain. The affected person submits a request. Those powers reach different objects, and none necessarily reaches the consequence already carried by the person.
Rights create a claim on an institution. Capability determines whether the institution can honor it. The distance between them is often a contract boundary. A frontline organization may own the decision while depending on software it cannot modify. Its vendor may control the workflow while depending on a model it cannot inspect. The provider may replace the model without accepting responsibility for a particular deployment. Each party can describe its own compliance while the return route remains owned by nobody.
Reversal authority has to be assigned before deployment. For every consequential action, the contract and system design should name who may reverse it, which systems must accept the correction, how quickly the operation must occur, and what evidence completion returns. An escalation address promises communication. A reversal mechanism promises a changed state and supplies proof.
The strongest test is a drill. Choose a completed determination, reverse it, and trace the correction through every recipient. Preserve the historical record without allowing it to drive current action. The drill exposes missing permissions and silent dependencies while they can still be repaired.
A retraction with machinery
The Retraction That Worked supplied the constructive case. South Africa published its draft National AI Policy as a numbered government notice. After fictitious references were found, the minister announced its withdrawal. A later gazette withdrew the original notice in its entirety through the same public system that had given the draft its status.
The retraction succeeded because the original act was legible and an authorized person could countermand it. The correction traveled through a public channel strong enough to alter the artifact's official life. The historical record survived, allowing anyone to see both acts.
The case is narrow. A withdrawn draft had no need to recover a missed livelihood or unwind a model's learned influence. Its value lies in the visible machinery. The artifact had an identifier. Authority had a name. The correction reached the system that conferred standing, and independent evidence remained available afterward.
That sequence separates an apology from an executed remedy. Announcement can stop further reliance and explain what happened. Formal withdrawal changes the operative record. Consequence tracing follows what escaped before either action completed. Collapsing those stages into one word makes the unfinished work disappear.
Before commitment
Before the Commit turned the findings into a procurement standard. Every proposed automated action needs a reversibility class before the workflow is allowed to execute it. Designers should identify the action, the party authorized to reverse it, the systems required to accept the correction, and the residual loss that no return route can reach. Affected people and operators need clear warning where the action crosses into practical permanence.
An executed remedy ends the original action's authority and propagates the correction through what remains reachable. It preserves evidence of the action and its reversal. It accounts for the loss left behind. Changed status alone cannot carry that claim.
The design order matters. An appeal added after deployment wraps a social promise around whatever technical residue the system happens to support. A return route designed first limits which actions may commit. Some systems will need better lineage or stronger contractual duties. Some actions should remain outside automation because no acceptable reversal exists.
The next system inherits records produced by this one. It may also inherit promises about correction that its predecessor could never honor. Succession begins with that residue: the files on the shelf, the influence in the model, and the consequences still moving after the screen says cleared.
Nothing Comes Back ran from 13 to 19 September 2026. The next arc turns to succession: what happens when the system entrusted with institutional memory is itself replaced.