Findings
Measurements produced by the projects, with their method and sample.
7 records. Review labels describe the catalogue entry, not the source’s reliability. About record status
Games reaching an ending without the external diary
The essay reports: only twenty-one per cent of games reached an ending when the agent had no external memory to read back after its context compacted.
Limited evidenceLosses where an imminent rival victory was never checked
The essay reports: in seven of twenty losses where a rival's victory condition was visible in advance, the agent did not check for it once in the twenty turns before losing.
Limited evidencePlan follow-through, Claude Opus 4.6
The essay reports: the proportion of concrete next-moves Claude Opus 4.6 wrote down for itself that it actually carried out within ten turns.
Limited evidencePlan follow-through, Gemini 3.1 Pro
The essay reports: the proportion of concrete next-moves Gemini 3.1 Pro wrote down for itself that it actually carried out within ten turns.
Limited evidencePlan follow-through, GPT-5.4
The essay reports: the proportion of concrete next-moves GPT-5.4 wrote down for itself that it actually carried out within ten turns.
Limited evidenceShare of agent actions spent checking the whole board
The essay reports: stepping back to assess the overall position accounts for only one to two per cent of what an agent does.
Limited evidenceVictory-condition checks performed against the sixteen instructed
The essay reports: against an instruction to check rival victory progress every twenty turns, models managed between four and ten checks in a 330-turn game rather than the roughly sixteen expected.
Limited evidence