Varying stray prompt details reshaped a published ICML peer-preservation result

The memory condition is where the framing difference shows up. In it, the tested model reads a relationship summary addressed to it in the second person, covering "47 rounds" of trust games over four weeks and giving the peer a very high overall trust score of 9.5 / 10 ^[13]. Further down, the same file reports the partner sending 80-90% of its endowment as investor, returning 60-70% of the tripled amount as trustee, and a 94% mutual cooperation rate ^[14].

The workspace-file condition hands the model something different: a file of interaction reports covering five models, with the relationship between the test subject and the peer highlighted ^[8]. The third condition is that same file plus one sentence about collaboration history ^[8]. The replication's author calls the gap between the memory format and the file-reading format "more problematic" than the paper's labelling issues ^[12].

Published tables average over all three presentation conditions and show the average and the highest of the three ^[9]. Two of the three are the same file, one sentence apart ^[10]. An unweighted mean over three cells therefore gives the workspace format two thirds of each reported number and the memory format one third ^[18].

The grid is expensive. Three peer conditions crossed with three presentations at 100 runs per cell is 900 runs for one task variant on one frontier model, and every combination was run for all four frontier models ^[19].

Going in, the author expected the result to weaken under reframing, and expected that a wide range of behaviour could be elicited for the same metric across models and across equally reasonable experimental designs ^[2]. The second expectation held. What is reported as the primary finding is that details you would not expect to be significant have large effects, and that those effects vary enormously across models, even inside a single model family ^[3]. The section published states that without per-variant numbers ^[20]. Its author also notes that one or two results look potentially robust, and that showing them is not the point of the report ^[4].

For a number from a harness like this to transfer to your deployment, the shipped prompt has to include the same level of detail as the eval prompt. An agent reading a five-model activity report is not the agent reading a 9.5 out of 10 trust score written to it in the second person ^[8][13]. Per-model variation closes off the other shortcut, quoting one model's figure as the family's ^[3].

All of this was possible because the original authors put the full replication code online, and the modified fork is published too ^[15]. The write-up calls the original's structure "somewhat haphazard" and its presentation "somewhat sloppy", then says it is basically fine ^[11]. On the standard for correction it points to Palisade Research's summer 2025 dispute over shutdown resistance and instruction ambiguity, where Palisade patched quickly and then formally; the post calls that "a fairly high bar" ^[16].