Posts tagged “benchmarks”

5 posts on SDAD Writes.

A Subagent Stops Inheriting Effort. The Model Name Does Not Move.

A real behaviour change shipped on 3 November 2026 in a changelog line about effort inheritance, six weeks after Opus 5.5 quietly started ignoring a user-level effortLevel and defaulting to medium. The claim that followed was that the model got worse. The only frozen, pre-registered panel says nothing it can detect moved, and its own validation says it could not tell an Opus 5 for Opus 5.5 swap apart anyway.

ExploitGym's Flag Was Computable From Its Own Code Until August. The Leaderboard Still Cannot Say Which Scores Predate the Fix.

ExploitGym's first gate is a flag the agent must capture by exploiting a target. The flag is derived as HMAC-SHA256(seed, task_info), and until 5 August 2026 the repository shipped the seed. The changelog says so plainly: the constants let anything inside an agent container derive the expected flag without exploiting anything. The fix is in. The record of which scores predate it is not.