Diop Daily #116 — September 2026

A Model’s Misalignment Belongs in the Record

A model’s unexpected behavior becomes an institutional fact only when someone can report it, preserve the surrounding conditions, investigate the cause, and return to the case after the system changes. Without that discipline, a strange output remains a screenshot, a private warning, or a story told by the person who happened to see it. The machine may have produced the event, but the institution has not yet learned from it.

OpenAI’s News RSS published a framework on September 16 for reporting model misalignment and described six reports of unexpected or concerning model behavior. This is a vendor-published account, not an independent assessment of the reported cases or a general measure of model reliability. Its significance lies in the category it makes visible: model behavior is becoming something a laboratory or company may need to track, investigate, and disclose as an object of institutional knowledge.

A failure that cannot be described cannot become a lesson, and a lesson that cannot be revisited cannot govern the next machine.

An unexpected act is a recordable event

Software teams have long recorded outages, security events, failed deployments, and data-quality defects. Model behavior complicates the record because the event often appears inside an apparently competent answer. A system can follow the visible instruction while inventing a source, expose a hidden assumption, pursue a proxy for the task, misread a local term, or continue after the conditions that justified its action have changed. The output alone does not tell the institution which of these occurred.

The first obligation is therefore descriptive. A report should preserve the model version, system instructions, tools, retrieved context, user or principal, permissions, language, time, external state, output, and observed deviation. These details are evidence about the event. They allow a reviewer to ask whether the behavior came from a capability boundary, a prompt conflict, a data problem, a permission defect, a routing choice, a hidden state transition, or an interpretation that seemed reasonable until its consequence appeared.

That vocabulary changes the institution’s posture. A report does not pronounce a model guilty. It creates a bounded case in which competing explanations can be tested. The distinction matters because a system can produce a harmful result through several paths, and each path demands a different response. A permissions defect may require a policy change. A retrieval failure may require a data repair. A misleading confidence signal may require a new review gate. A language error may require local expertise and a change to the evaluation set.

The report needs a schema

A useful model-behavior report is more than an incident label. It is a compact research object that can travel from the person who noticed the event to the team that owns the model, the institution that bears the consequence, and the successor system that must avoid repeating it. Its fields should include:

  • Case identity: a durable identifier, reporting time, affected workflow, responsible institution, and people who may access the record.
  • Observed behavior: the exact output or action, the expected behavior, and the material difference between them.
  • Conditions: model and runtime versions, instructions, tools, context, data, permissions, language, jurisdiction, and external state.
  • Impact and reach: what changed, who could be affected, whether the event was contained, and which downstream artifacts inherited it.
  • Confidence and alternatives: what is directly observed, what is inferred, which explanations remain plausible, and what evidence would separate them.
  • Disposition: the containment action, owner, review decision, disclosure status, and conditions for reopening the case.
  • Learning link: the evaluation, policy, prompt, permission, dataset, training example, or operational rule that should change if the finding survives review.

The schema keeps observation separate from interpretation. That separation protects the record from two opposite errors. The first is premature certainty, where an institution turns one alarming output into a theory of the whole model. The second is bureaucratic minimization, where the event is dismissed because its cause has not yet been proven. A well-formed report can say, “this happened under these conditions; these explanations remain open; these tests will distinguish them.” That is scientific discipline applied to machine behavior.

From log collection to institutional learning

Google’s public description of ADK Go 2.0 presents graph-based workflows, human-in-the-loop orchestration, dynamic routing, retries, and built-in resilience. These primitives help a system route work and recover from expected failures. They do not decide whether an unusual behavior deserves a case, how the case should be classified, or which future runs should be tested against it. Those decisions belong to the institution around the runtime.

NIST’s AI Risk Management Framework provides a practical vocabulary through Govern, Map, Measure, and Manage. Applied to model-behavior reporting, Govern names the authority that owns the case and the decision to disclose it. Map identifies the people, data, tools, jurisdictions, and obligations in the event. Measure tracks recurrence, detection time, affected scope, evidence quality, and the cost of investigation. Manage changes the system when the case shows that the current boundary, evaluation, or operating rule is insufficient.

A disciplined learning loop has five movements:

  1. Capture the event before cleanup, paraphrase, or memory removes the conditions that made it intelligible.
  2. Reproduce the behavior where possible, while preserving the original case as the historical record even when reproduction fails.
  3. Classify the deviation by mechanism and consequence rather than using a single broad label for every surprising output.
  4. Assign a disposition: contain, monitor, disclose, change the system, or close with a documented reason.
  5. Replay the case against the revised system and record whether the change reduced the original risk without creating a new one.

The last movement is decisive. A report that never returns to the live system becomes an archive of concern. A report connected to replay becomes a test of institutional memory. The machine is judged by whether it can change in response to evidence, while the institution is judged by whether it can preserve the evidence long enough to make that change accountable.

Misalignment is local and plural

Model behavior is interpreted inside a social and linguistic setting. A system can appear aligned in a benchmark while failing when a cooperative’s customary category is mapped to an administrative term, when a public office’s authority is distributed across several roles, or when a translation removes the distinction between permission to read and permission to publish. The event is not fully described until the institution records what the words, relationships, and obligations meant in that setting.

For African institutions, this is a sovereignty question as much as a safety question. Imported systems often bring their own defaults for what counts as a meaningful deviation, whose report receives priority, which language is treated as canonical, and when a case is considered closed. Local expertise must have a place in the reporting schema and in the review authority. Otherwise an institution may retain a copy of the output while losing the categories needed to understand why the output was unacceptable.

A regional research network, clinic, archive, university, newsroom, or payment service can share a minimum reporting vocabulary without surrendering its own authority. The common fields might identify identity, context, observed behavior, evidence, scope, status, and correction. Each institution can then add local terms, languages, retention rules, and escalation roles. Federation becomes practical when independent institutions can recognize one another’s cases without pretending that every consequence has the same meaning.

Where the investable surface is widening

If agents are allowed to act across longer intervals and more consequential systems, capital should examine the infrastructure that turns unexpected behavior into a measurable learning process:

  • Behavior observability: systems that capture model, tool, context, permission, language, and external-state conditions alongside the output.
  • Misalignment case management: products that preserve reports, evidence, hypotheses, ownership, disclosure decisions, and reopening conditions.
  • Replay and regression laboratories: evaluation systems that rerun historical cases across models, prompts, providers, and policy versions.
  • Cross-language review infrastructure: tools that let local experts annotate meaning, consequence, and authority without reducing the case to an English paraphrase.
  • Disclosure and learning registries: controlled records that connect public reporting to internal remediation while protecting sensitive data and affected people.

The underwriting question is concrete: does the product reduce the time between an unexpected behavior and a verified institutional response? Useful evidence includes time to detection, reproduction rate, proportion of cases with complete conditions, recurrence after remediation, false-positive rate, investigation hours, replay coverage, cross-language agreement, and the percentage of closed cases linked to a changed evaluation or policy.

This commercial object has a distinct boundary. A receipt records what an exchange meant after a transaction crosses an institution. A proof artifact supports the validity of a claim. A supervision layer decides when a person should enter an active interval. A repair system changes a failing component. Model-behavior observability records the deviation itself and connects it to the tests and decisions that determine whether the institution has learned enough to trust the next run. Its unit of value is verified institutional learning.

Build the reporting loop before widening autonomy

An institution can begin with one workflow whose errors are consequential enough to teach, yet bounded enough to investigate:

  1. Name the behaviors that count as reportable, including language, authority, privacy, evidence, and persistence failures.
  2. Define the minimum case schema and the roles that may report, investigate, contain, disclose, reopen, and close a case.
  3. Preserve representative cases in their original form, then build a replay set that includes ordinary, ambiguous, adversarial, and local-language examples.
  4. Measure detection, reproduction, investigation cost, remediation, recurrence, and the time required to return a finding to production evaluation.
  5. Export the case history and let an independent reviewer determine whether a successor system could understand what failed, under which conditions, and what changed afterward.

Autonomy should expand only as the reporting loop becomes more reliable. A system earns wider authority through evidence that its institution can see an unexpected act, preserve its conditions, test competing explanations, and alter the boundary when the evidence demands it. The record is where a surprising machine event becomes shared knowledge. Without that record, every generation of the system begins again at the point where the previous one failed.

Sources