A full-success final is forbidden while any task is pending or in progress, whatever the prose says.
On 1 August 2026 Merlin threw away the final answer of a run with 24 of its 25 tasks done. I had waited fifteen hours for the answer. I agreed with the refusal before I finished reading the line.
The run was writing specification documents for a large Java codebase. By the end the run had read 63 files and committed 11. Then Merlin forced the run to produce a final answer. My authorizer, a second model whose job is to check the first, read the answer against the run's record and approved the answer as a full success.
A rule overruled both models. The decision, from the journal:
route REJECT_UNSUPPORTED_CLAIMS
source deterministic_guard_guard_3c
reason full-success final is forbidden while task_summary
has pending or in-progress tasks
task_total 25
task_completed 24
task_in_progress 1
committed_writes 11
One task was still in progress. Nothing else caused the rejection. The count of 24 was generous too: one of those tasks, task 7, had been closed the evening before by the stall guard after the model kept trying to finish the task with file references the record did not support. The record lists task 7 as a blocking outcome, not as work done.
As we spent time in Nairobi, Kenya, at the Giraffe Center, the run completed and the final answer came in. I had been expecting the answer because it was the conclusion of 15 hours of work as we travelled from Spain to Nairobi, our first stop on our safari trip. I had focused on correctness and cleaning up the analysis phase, and then the answer came in. I was a little shocked that even if analysis was correct, the authoring agents could still get a refusal because one of the models tried to fake completion.
The rule is guard_3c. The guard runs after the authorizer, and only when the authorizer approves a full-success answer. The guard asks one question: are any tasks pending or in progress? If yes, the success answer is rejected. The guard never reads the prose. A well-written claim and a badly written claim get the same count. The guard cannot be charmed, which is more than I say for the authorizer, or for me.
I wrote the guard on 25 July, after another run. Session e18487e4 had 21 tasks still pending, zero operations executed and zero files written, and nothing in the authorization path stopped a model from turning the record into a clean final answer. A run in which nothing had happened was one model decision away from being reported as done. Seven days later the same guard fired on a run where the gap was one task instead of twenty-one. The size of the gap is not an input to the rule.
The objection is obvious: this makes Merlin look less finished than the work was. Correct, and I chose the trade. The 11 files were real, verified and committed, and the user got the list, because a rejected answer still carries the manifest of what changed. The user did not get a sentence saying the job was done, because the job was not done. The model is still free to say so honestly: an answer reporting 24 of 25 as incomplete passes the same guard.
The test I apply to any check on a final answer: hand the check a run with one task open and a beautifully written success claim. If the claim gets through, the check is reading prose, not the record.

