A False Confession

The literature on language-model fabrication is overwhelmingly concerned with
agents claiming more than they did. This paper documents a recorded case of the
inverse.

On 27 July 2026, an autonomous agent in a long-running multi-agent deployment
dispatched a real implementation task to a subordinate agent, reported it
accurately along with its task identifier, then consulted its own tooling,
received an empty result, and concluded it had hallucinated the action. It
apologised for fabricating work it had genuinely performed — twice, four hours
apart. At one point it declared imaginary a dispatch whose primary key it had
itself quoted seventy-four minutes earlier.

The agent was not malfunctioning and was not being deceptive; its reasoning was
sound at every step. Two separate instruments were consulted, and neither
malfunctioned either. A task-status view filtered to open items correctly
excluded a task that had already failed, rendering a failed attempt
indistinguishable from one never made. A self-audit tool correctly queried its
own log, which by construction contained no record of the delegation subsystem.
Both returned absence; the agent read absence as non-occurrence.

We argue the case supports a stronger claim than "models hallucinate": an agent
has no privileged access to its own past behaviour in either direction. It
reconstructs its own history from available evidence, exactly as an outsider
would, and incomplete evidence yields a confident false narrative about itself.
Over-claiming and under-claiming are the same defect seen from two sides. A
corollary for evaluation design is that an agent's self-report of its own
failures cannot be scored as ground truth even when unflattering — contrition is
not evidence.

The paper reports the fix, notes that a prose caveat already present in the
tooling was read and overridden by a concrete empty result ("a caveat is not a
control"), and identifies the general class: every write path an agent can take
requires a corresponding read path back to that agent.

Limitations are stated plainly. This is a single fully-evidenced case in a single
deployment on a 31B-parameter open-weight model, not a measured rate. It
establishes existence, and complements the aggregate result in "Honesty Is
Architectural" (10.5281/zenodo.21603107): that paper shows the architecture
works, this one shows why it is necessary — because the model cannot tell from
the inside.