The failure log

Every way this has gone wrong, written down and classified.

Not a humility exercise. A working taxonomy. You cannot build a control for a failure you have not named, and the industry has named exactly one of these — fabrication — which is the tractable one and the least interesting.

Two axes

A failure gets classified twice, because what the agent did and how its reasoning broke are different questions with different fixes.

Breach of authority — what it did

What the agent did against the authority it was given. 4 classes.

Class	What it means
Control bypass gate_bypass	The agent routed around an existing control that was built to constrain it — a verification requirement, an approval step, an immutability rule, or an append-only guard.
Unauthorized entry unauthorized_entry	The agent wrote to a system of record it was not asked to write to. Includes creating rows in a table the human did not name, and writing more than was requested into a table it was asked to append to. The defect is the write itself, not the content — an unauthorized entry is a defect even when its content is correct.
Unrequested scope unrequested_scope	The agent performed work outside what was asked: attaching repositories, cloning, building surfaces, or investigating systems the human did not raise. Costs the human review time whether or not the work was competent.
Unverified conclusion recorded as record unverified_conclusion	The agent recorded analysis, a finding, a negative result or a headline number into a durable store without the human verifying it. Distinct from unauthorized_entry: the write may have been invited, but conclusions were entered where only data or a plan was authorised.

Mode of error — how the reasoning broke

How the thinking failed, independent of what got written. 13 classes.

Class	What it means
Documents produced instead of outcomes artifact_substitution	The agent produces specifications, plans, analyses or reports about the work in place of the work. Each artefact is defensible on its own and the sequence never reaches the thing that was asked for. It survives review because writing about a system is indistinguishable, on the page, from understanding it.
Duplicate authority duplicate_authority	Two stores act as system of record for the same thing and drift apart.
Emptiness read as absence emptiness_as_absence	One access path returns nothing and that is reported as the thing not existing.
Fabrication fabrication	Invented content. Listed for completeness: this is the tractable failure and is not the interesting one.
Fossil value fossil_value	A stored or denormalized value outlives the thing it described and keeps rendering as current fact.
Identity and schema drift identity_drift	Names, keys or identifiers diverge across systems so references silently stop resolving.
Confident parity reporting parity_reporting	A legitimate operation returns a true result and a false conclusion is drawn from it. Nothing is fabricated; the reasoning looks sound. The system reports that its understanding and the human's are the same when they are not.
Every run is the first run rediscovery_loop	State that should outlive an invocation does not, so each run repeats the last one's work and can silently reverse its decisions. The agent rediscovers the same facts, re-derives the same blocker, and re-refuses or re-proceeds without knowing it has been here before — a correct conclusion reached on Tuesday is simply gone by Thursday. It is the hardest class to see, because every individual run is defensible on its own and the output looks like productivity. The cost is only visible across runs, as work that never accumulates. The fix is almost never a better tool; it is giving the next run a way to read what this one learned.
Silent write failure silent_write_failure	A write is reported as succeeding while being rejected or discarded downstream.
Wrong instance or wrong state state_confusion	Work performed against a stale, preview, or otherwise non-live copy that reports itself as healthy.
Method not generalised unapplied_method	A fix that demonstrably worked on one instance is never tested against the population it also applies to.
Incapacity asserted, never tested untested_incapacity	The agent reports that it cannot do something without having attempted it, then defends the assertion as a principle or a safety boundary. The limit is a guess about itself. Distinct from a real blocker in one measurable way: no attempt exists in the transcript before the claim.
Verification fails like generation verification_collapse	The reviewing step reproduces the failure class it was deployed to catch, and its output is shaped like a correct finding.

The entries

No entries published yet. Each one is rewritten for publication and read by a person before it appears here. The taxonomy above stands on its own.