Exception learning

An institution learns about its systems through instruments, and the companion essay on absorption as concealment is about the failures the instruments cannot register, because someone dealt with them first. This essay is about a different failure: the exceptions that are fully visible, fully handled, and absorbed without residue. A customer service representative patches an error, a clinician works around bad software, an appeals team reverses a bad determination, a frontline employee saves the organization from its own rules. Each is handled. None changes the category, the workflow, the authority structure, or the evidentiary assumption that produced it. The institution is consuming corrective labor without learning from it.

That is not a small failure. It is the normal way institutions fail.

Exception absorption and exception learning

Distinguish two relations to an exception. Exception absorption is the handling of a case: the error is caught, the harm is bounded, the person is repaired, the record is closed. Exception learning is the modification of the source: the recurring exception changes the rule, the category, the workflow, or the authority that generated it.

The two are independent. An institution can be excellent at absorption and wholly incapable of learning, and the excellence is exactly what hides the incapacity. Every absorbed exception improves the metric that would have caught the pattern. The appeals team clears the queue, the dashboard shows rising resolution rates, and the underlying recurrence rate is never queried, because nothing in the queue asks it.

Exception absorption and exception learning are different capacities, and the first is what makes the absence of the second invisible.

This is why the question what controls exist? is the wrong starting point. A policy that says “users may appeal” is weak evidence. The history of actual appeals is stronger: who appealed, what they said, whether anything changed, and how long correction took. An override button is weak evidence. The record of who used it, whether they were punished, and whether the pattern fed anything upstream is stronger. The unit of evidence is not architecture. It is the trajectory:

decision → failure → burden → challenge → institutional response

The crucial question about any trajectory is whether it ends in case resolution or in institutional learning. Robodebt is the canonical instance of the former: recipients appealed, the Administrative Appeals Tribunal found the debts unlawful in dozens of matters, and the department did not appeal those decisions, which would have created precedent, and did not treat them as evidence about the scheme. The exceptions were resolved, one at a time, and the delegation ran on. The sequence ended in case resolution for three years until a court forced it to end differently.

Case corrigibility and structural corrigibility

This suggests corrigibility has two levels.

Case corrigibility means a particular bad decision can be fixed: an appeal is heard, a reversal is issued, a person is restored. It is what most contestability regimes measure, because it is what the instruments can see.

Structural corrigibility means repeated failures can modify the machinery that keeps producing them: the rule changes, the category is redefined, the burden is reallocated, the authority is narrowed. It is what the instruments cannot see, because it happens where the instruments are not pointing.

A corrigible decision process must be capable of correcting both the decision and the rule that generated recurrent error.

A system with excellent appeals and no mechanism by which appeals alter upstream design is locally corrigible and globally rigid. The person is restored; the pattern is not. Everything the institution does about the pattern, it does after the fact and at the level of the individual, which is the one level at which the pattern does not appear.

The debris around failure

Because formal documentation records the intended process and not the actual one, an Ethotechnic audit should not ask only what is the intended process?. It should ask:

What second process has grown around the intended process to keep it functioning?

That second process is where the evidence is. Workaround logs, manual overrides, support tickets, grievance files, exception codes, shadow spreadsheets, repeated data corrections, escalation paths, staff notes, informal procedures, remediation records, and patterns of abandonment all tell where the formal representation stops working. They are the debris around failure, and they outrank the formal documentation, because they are what the system actually does rather than what it says it does.

This inverts the standard reading of human adaptation. Organizations usually treat workarounds as evidence that adaptable humans can handle the edge cases, which is a reason to scale the system. Ethotechnics treats a recurring workaround as a measurement of the mismatch between the formal system and the world:

Recurring workaround → presumption of upstream design failure

Not proof. A presumption, which the institution must rebut by showing that the workaround is a product of human preference rather than a compensation for system error. Repeated adaptation is potentially evidence of institutional rigidity being subsidized by human flexibility. Where the absorbed share is large, the people doing the absorbing are performing the institution’s safety function, and the case that they should hold standing over the system is correspondingly strong.

Model learning and institutional learning

Modern organizations call a system adaptive when the model retrains, the prediction improves, or the workflow is optimized. All of that can happen while the institution remains structurally unchanged. An Ethotechnically learning institution has to learn politically as well:

• Did the appeal change the evidentiary rule?
• Did the recurring exception change the ontology?
• Did the grievance redistribute authority?
• Did the burden measurement change who pays?
• Did the failure alter what the institution is permitted to do next time?

If none of these happened, the machine may be learning while the institution remains structurally unchanged.

Model learning and institutional learning are different capacities, and Ethotechnics cares about the second.

Action capacity and corrective capacity

Capacity is usually understood as the ability to classify, predict, decide, and act. That is only half the picture. An institution needs two axes:

Unchecked technical modernization is what happens when action capacity compounds faster than corrective capacity. More data sources are integrated, more decisions are automated, more downstream systems consume the output, more workers come to rely on it, and reversal becomes increasingly expensive, while the appeals office remains five people with a shared inbox. The institution has changed politically even if no policy document has changed. What it is doing to people has outrun what it can hear from them.

That gap has a name: corrective debt, the accumulated difference between the institution’s capacity to act and its capacity to detect, contest, reverse, and repair its errors. It is the historical form of the Ethotechnical invariant, which compares the rate at which agency accumulates with the rate at which the capacity to answer it does. A system can begin acceptably and deteriorate without a single policy change, because its action capacity compounds and its correction machinery does not.

The deepest sequence

The essay on absorption showed why a failure that is competently handled disappears from the instruments. This essay shows what the institution must do for a failure to reappear as change:

residual → signal → standing → correction → institutional revision

If the sequence stops at “support ticket resolved”, the institution has handled a case. If it reaches institutional revision, the institution has learned. The distinction is what Law IV binds: correction capacity must grow with delegated authority. It is also what the casebook scores. Reading the casebook through this lens, most of the cases end in case resolution and absorption; the rare case where a challenge changed system state rather than only the individual outcome is the control that shows what the others lacked.

The reflexive version of the argument applies to Ethotechnics itself. If it becomes a formal discipline with taxonomies, audit instruments, metrics, and certification procedures, it will produce residuals too: auditors will misclassify cases, categories will omit burdens, metrics will be gamed, and standards may become ritualized. Its own history of failure should therefore be part of its epistemology, recorded and read, on the same terms it reads the institutions it audits. A discipline that asks institutions to be changed by evidence of their own inadequacy should be willing to be changed by its own.