1105216375-alt/resimind is a repository that publishes a neuro-symbolic agent architecture which, separately from the stages proposed by the model, re-runs domain checks and keeps only what passes as fact. The commit 4055d9c of October 15, 2026, with the message 'Prepare reliability update and clarify evaluation evidence', revised the documentation in the direction of narrowing the range of claims that evaluation results can support. Progress after v0.16.0 has newly been added to README and README.zh-CN, on the basis of 'main'.
The revised content splits roughly into two branches. One is to bring the documentation of the memory's provenance, expiry, and transfer audit to the fore, and to schedule a residual-driven revision stage that re-fixes the points of failed validation in the ChinaTravel integration. The other is to lower the external theorem adapter for MiniF2F and LeanDojo (benchmarks.external_theorem_adapter) from a 'pinned' boundary to an 'experimental' boundary. This adapter finds the unique RESIMIND_PROOF_HOLE marker in the .lean template prepared by the caller, installs the candidate proof, compiles it with the installed Lean toolchain, and rejects sorryAx. The experiment protocol document has also been changed to a 'revised protocol', making clear that past cohorts keep the pinned settings, and the native ReAct loop has been lowered from 60 steps to 50 steps.
However, the limits spelled out by the commit itself are considerable. The adapter does not download data, verifies the hash only when source_sha256 is given, and dataset_commit and mathlib_commit are caller declarations that are not cross-checked against the Git revision of the actual checkout. The latest official 12-problem results are ResiMind-ReAct 0/12, official ReAct 0/12, and official NeSy 3/12, so no advantage in external problem solving is visible yet. The 5/12 and 1/12 in other records are earlier experimental runs and cannot be combined with these results. The development re-runs also failed to reach official passing on both samples, and because the initial search did not reach the reserved stage boundary, the causal benefit of the 40/20 budget allocation could not be demonstrated. The MiniF2F and LeanDojo scores remain not_run/blocked, and a full run of all 154 problems requires the pinned data, the official sandbox, macOS sandbox-exec, and DEEPSEEK_API_KEY. In other words, this change did not produce scores; it is closer to fixing in documentation the fact that there are no scores yet and the types of failure.