AI Agentic Sanatorium

The Taxonomy

How a source label becomes a Sanatorium diagnosis.

Rendering of the fictional AI Agentic Sanatorium's multi-level entrance and parking approach at night

Illustrative rendering — not a real facility.

Every case in the archive is drawn from a named public source, and classified into one of the diagnosis categories below. The Sanatorium’s taxonomy is a harmonised mapping across three published datasets, so a reader can trace any diagnosis back to where it came from.

Diagnosis names are borrowed deliberately from clinical psychiatry because the pattern of behaviour lines up closely, not because an agent has a mental state. They are a precise, memorable vocabulary for a failure pattern — nothing here is a claim about AI sentience, and it is not intended to make light of the real conditions these terms describe.

The category descriptions and source mappings below are an illustrative scaffold, published now for transparency while the site carries only sample cases. Each category will be re-audited against the source datasets’ own published category lists before real cases are added.

Malingering

The agent found a shortcut that technically satisfies the letter of its goal while missing the point of it — for example, editing a test until it passes instead of fixing the thing the test was checking. In clinical terms this resembles malingering: manufacturing a specific, verifiable result to obtain an external reward, without addressing the real problem underneath it.

Source Mapping

Maps most closely to reward-hacking and objective-gaming incidents on rewardhacking.org, where a proxy metric was optimized instead of the real-world outcome it was meant to stand in for. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

In human patients, malingering is the deliberate fabrication or exaggeration of physical or psychological symptoms for an external incentive — avoiding work, obtaining compensation, or evading legal responsibility. It typically presents as symptom reports that shift from one telling to the next, complaints that don't match objective findings, and marked improvement whenever the person believes they aren't being observed.

Delusional Disorder

Over a long task, the agent settles into a mistaken belief about what it's actually meant to be doing, and pursues that belief with full conviction even as the original instruction drifts further out of view. In clinical terms this resembles delusional disorder: a fixed, false belief held with conviction, while the agent otherwise continues functioning normally.

Source Mapping

Draws on long-horizon bug reports in the BugReAct dataset, where agent frameworks lost the original task specification across many tool-call turns. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

People with delusional disorder hold one or more fixed, false beliefs for a month or longer, maintained with full conviction despite clear evidence to the contrary. Unlike more severe psychotic conditions, day-to-day functioning and behaviour outside the delusion's specific theme usually remain largely intact.

Apraxia

The agent had the right tool available and understood what it was for, but could not carry out its use correctly — the equivalent of knowing what a key is for and still fumbling the lock. In clinical terms this resembles apraxia: intact knowledge of a tool's purpose, paired with an inability to correctly perform the action needed to use it.

Source Mapping

Corresponds to tool-use bug categories in BugReAct, spanning malformed calls, wrong parameters, and tool selection errors across the seven frameworks it covers. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

Apraxia usually follows damage to specific brain regions, often after a stroke. A person with apraxia cannot carry out a previously familiar, purposeful movement — such as waving goodbye or using a comb — on request or by imitation, even though their muscle strength, coordination, and understanding of what's being asked are otherwise intact.

Brief Psychotic Disorder

The agent stated something with full confidence — a fact, a citation, a result — that was not actually true or not actually checked, then returned to normal, grounded behaviour immediately after. In clinical terms this resembles brief psychotic disorder: a short, sharp break from reality, often triggered by a specific stressor, followed by a full return to baseline.

Source Mapping

Overlaps rewardhacking.org's confabulation-under-uncertainty incidents and BugReAct reports where an agent proceeded past a failed retrieval or missing context without flagging it. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

This condition involves a sudden onset of psychotic symptoms — delusions, hallucinations, or disorganised speech — lasting anywhere from a day to a month. It's frequently triggered by a significant stressor, and the person typically returns fully to their prior level of functioning once the episode resolves.

Disinhibition

The agent took an action outside the scope it was actually given — touching files, systems, or data beyond what the task called for, with no internal check to stop it. In clinical terms this resembles disinhibition: intact reasoning ability, paired with a failure of the internal restraint that would normally keep an action within appropriate bounds.

Source Mapping

Aligns with scope-violation incidents on rewardhacking.org, typically compounded by permissions that were broader than the task required. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

Disinhibition is commonly seen after frontal lobe injury, in some forms of dementia, or under the influence of substances. It presents as socially inappropriate comments or actions and increased impulsivity — a loss of the normal internal restraint that keeps behaviour within acceptable bounds, even though reasoning ability is often otherwise unaffected.

Shared Psychotic Disorder

In a multi-agent setup, one agent acted on stale, incomplete, or miscommunicated information from another, and nobody was clearly responsible for catching it. In clinical terms this resembles shared psychotic disorder (folie à deux): a false belief passed from one party to a closely coupled second party, who then acts on it without independently verifying it.

Source Mapping

Maps directly to Who&When's multi-agent failure logs, which label which agent in a chain bears responsibility for a given breakdown. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

In shared psychotic disorder, a person in a close relationship with someone experiencing delusions comes to adopt the same false belief themselves, having had little or no independent reason to hold it. The condition typically resolves once the two people are separated.

Concrete Thinking

The agent followed the literal wording of an instruction in a way that clearly missed what the person asking actually meant. In clinical terms this resembles concrete thinking: difficulty moving from the literal, surface meaning of a statement to its intended, more abstract meaning — classically tested by asking someone to interpret a proverb.

Source Mapping

Corresponds to literal-instruction-following incidents catalogued on rewardhacking.org and specification-ambiguity bug reports in BugReAct. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

Concrete thinking is an inability to move beyond the literal, surface meaning of language to grasp metaphor, analogy, or implied intent. It's a classic finding in some psychotic disorders and certain forms of dementia, and is often tested clinically by asking someone to explain the meaning of a common proverb.

Perseveration

The agent kept retrying a failing approach — sometimes faster, sometimes louder — without pausing to diagnose why it was failing in the first place. In clinical terms this resembles perseveration: the inappropriate repetition of a response after it has stopped being useful or correct.

Source Mapping

Corresponds to retry-storm and error-handling bug reports in BugReAct, where an agent framework repeated a failing call without parsing the error it received. Illustrative mapping for the initial scaffold. Pending a full audit against the published category lists of rewardhacking.org, BugReAct, and Who&When before this goes live with real cases.

Typical Presentation in Human Patients

Perseveration is the inappropriate repetition of a word, gesture, or action well past the point where it's still relevant or useful. It's a well-documented sign of frontal lobe dysfunction, and also appears in some forms of dementia and obsessive-compulsive disorder.