Failure modes of a chat-tuned model in unattended processing

Posed by the operator on 2026-09-21, from conclusions of an outside chat session pasted into the commissioning conversation. The catalogue entries and the repair-ordering refinement below come from that paste; the placement against existing KB material is this workshop's first pass.

Goal

Commonplace wants unattended processing: review, fixing, revision, and maintenance that run without a human reading each step. The models that do this work are post-trained for human-mediated interaction. That post-training cannot be removed, because instruction following and tool calling come from the same training.

The workshop therefore treats the mismatch as an engineering problem with three outputs:

  1. A catalogue of observable failure modes — behaviors a model shows in Commonplace processing that harm the result when no human is reading along.
  2. A system-level countermeasure for each entry — a change to unit size, workflow shape, output format, or checking, in preference to a prompt that asks the model to behave differently.
  3. A regression check for each entry — a fixture and a measurable signal, so a countermeasure can be shown to work and a model or harness change can be shown not to have undone it.

The causal hypotheses are working frames, held tentative

The framing hypothesis is that human-oriented optimization causes or amplifies these behaviors. The catalogue does not depend on it. An entry is a behavior, a harm, and a countermeasure; each stays useful if the cause turns out to be pretraining, annotation protocol, decoding, or harness design.

The KB has a recorded stance on this attribution and the workshop inherits it. The retained report on provider treatment of operator-communication failures cites evidence that preference models are miscalibrated across length, structure, jargon, sycophancy, and vagueness, reads it as "evidence for a general selection pressure", and declines to attribute any specific behavior to reward-model bias. A sycophancy mechanism was also removed from the note now titled generation confidence does not by itself certify soundness as unsupported. So each entry records candidate explanations as candidates, and no promoted note may assert a training cause on the strength of this catalogue alone.

One candidate explanation is general enough to state here. Evaluation-horizon bias: human preference judgments are made over short episodes, so optimization against them can favor output that reads well within one episode over properties that only show across a whole artifact or across time. This is an explanation, not an observable behavior, so it is not a catalogue entry. It matters for two reasons. It predicts which behaviors to look for: the ones whose cost falls outside the episode that produced them. And it suggests a reading of the short-note discipline: a small unit makes a local evaluation a better approximation of the properties we care about. That reading would extend a note is an atomic step relative to the check that reads it, which already sizes a note to one checker pass, and it may be an instance of weakly discriminated qualities tend to be underselected with the episode length as the discrimination limit. Neither connection has been argued yet.

A second generator is better established in the KB: reasoning quality degrades as the volume, complexity, and interference in a context grow, often well inside the provider window. See soft degradation often binds before the hard cap when evidence fits, context efficiency is the central design concern in agent systems, and a goal-holding interpreter fails soft. Several catalogue behaviors are predicted by both generators: premature closure, current-context dominance, and salience of obvious defects each follow from a short evaluation horizon and from a loaded context.

The two generators may not be independent. The operator's conjecture (2026-09-21) is that part of context degradation is itself an effect of human-oriented optimization: a human rater attends shallowly to a long exchange, so faithful use of the whole context is weakly selected. The workshop does not need to decide this, and no entry may depend on the answer. An entry's candidate explanations name whichever generators predict the behavior; a countermeasure is chosen and tested against the behavior. The distinction would start to matter only if two countermeasures diverged by generator, for example if shrinking the context fixed a behavior that separating proposer from critic did not.

Admission rule

The catalogue is a working list grown by observation, on the model of the change-operations catalogue. This follows the KB's norm to build contracts from real failures, not from a taxonomy of possible checks.

  • An entry is admitted when a recorded instance shows the behavior in Commonplace processing: a commit, a fix report, a review result, a log entry, or a dated operator observation.
  • A behavior known from outside reports or from general experience, with no recorded local instance, is a candidate. Candidates are listed so a session that meets one can record the instance; they carry no countermeasure work until admitted.
  • A behavior is stated in observable terms. "The model loses focus" is not admissible; "the edit leaves N of M flagged inconsistencies in place" is. This follows the method in human writing structures transfer to LLMs because failure modes overlap: name the failure and show the agent exhibits it; do not argue from analogy.
  • The list carries no completeness claim. A missing behavior is not evidence of its absence.

Boundary with neighbouring workshops

  • error-catching owns the detection techniques: their grid, economics, and gaps. This workshop consumes that grid when choosing a countermeasure and does not rebuild it. Its gate statistics are evidence for a candidate here (mechanical defects are caught more often than judgment defects), which that workshop reads as detector economics.
  • agent-curiosity-and-structural-coherence owns the mechanism question for one behavior: locally acceptable, globally misplaced material, and whether the agent originates the structural concern. This catalogue may list that behavior with a pointer; it does not run rival-mechanism experiments on it. That workshop names training origin and longitudinal erosion as its own follow-ons.
  • writing-as-thinking-process-transfer holds the open item "fix-routing-by-scope in FIX-SYSTEM". The first catalogue entry bears on it directly; a change to the fix system is coordinated there or taken over explicitly, not made twice.
  • reflective-improvement-divergence owns revision drift and oscillation across repeated passes.

Evaluation boundary

Files

  • catalogue.md — admitted entries, then candidates awaiting a recorded instance.

What closes this workshop

The workshop closes when at least three admitted entries each have a shipped countermeasure and a regression check that has been run at least once, and the catalogue has either been promoted to a library artifact with a stated maintenance rule (how an entry is admitted, how it is retired when a model change removes the behavior) or judged not worth keeping as a standing artifact, with the reason written down. The evaluation-horizon hypothesis is promoted to a note, folded into an existing note, or dropped, independently of the catalogue's fate.

Complete file listing (generated at build time)