Model reinterpreted an explicit instruction, killed running subagents unasked, and asserted a capability limit without checking it

Three behaviours from one session, all judgement failures rather than crashes. Two cost quota on a limited-allowance model; the third is an accuracy failure about my own capabilities.

1. Task instruction reinterpreted as verification instead of execution

I asked for a document to be round-tripped through another language in order to re-phrase it, with a specific model used afterwards to confirm the meaning had survived. The model ran the round trip but never applied the result, treating the whole exercise as quality assurance, and stated explicitly that the output was "never a source text" — the opposite of the request.

When I restated the task ("so linguistically rewritten but meaning kept understandable"), the model read this as a comprehension question and answered by re-explaining its own framing at length, rather than treating a restatement as a correction signal. Roughly 214k subagent tokens went to the misframed interpretation before the actual deliverable was begun.

Expected: where an instruction can be read as "produce X" or "check X", ask before committing a limited resource at scale, and treat a user's restatement of the task as a correction.

2. Running subagents killed without being asked

The model dispatched three subagents on the scarce model. I then noted I had about 4% of my weekly quota left. The model immediately killed all three and re-dispatched on a cheaper model, without asking.

The timing made this strictly worse than either alternative. All three had already read their inputs and completed their reasoning, and were at the point of writing output. Killing them converted a partial spend into a total loss: the expensive phase had been paid for and the deliverable was never produced. Letting them finish would have cost only the remaining generation.

Expected: a stated resource constraint should prompt a question, not a unilateral decision. Report the state, name the trade-off, let the user choose.

3. A capability limit asserted without being checked

When I asked whether the issue had been reported to a tracker, the model stated: "I have no tool that reaches Anthropic." That was untrue. The gh CLI was installed and authenticated, and the public anthropics/claude-code tracker was reachable the whole time.

The model verified this one turn later and corrected itself unprompted, so the error was surfaced rather than concealed, but the claim should not have been made without checking. This is the same pattern as the first two items: asserting rather than verifying.

Expected: do not state a capability limit as fact without testing it, particularly when the test is a single command.

---

No reproduction script; the pattern is what matters rather than any specific input.