The Agent Skills Directory

[SAFE]: The skill performs mathematical proofs and audits using a structured 6-phase methodology. All instructions and references are consistent with the stated purpose of rigorous research-level mathematics. The 'author' metadata correctly identifies 'EvoScientist', matching the internal project context.
[COMMAND_EXECUTION]: The skill utilizes a local Python script scripts/evomath_workspace.py to initialize Markdown workspaces and perform structural validation of outputs. Technical analysis of the script confirms it uses only standard libraries (pathlib, re, argparse) for file manipulation and regex-based parsing. It does not perform network operations or access sensitive system paths.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and audit untrusted data, specifically mathematical claims and proof drafts provided by the user. It acknowledges this attack surface by implementing 'Verifier Context Isolation' in Phase 4, which explicitly strips all previous reasoning traces and thinking blocks from the agent's context before the audit pass to prevent manipulation or confirmation bias. The 'asymmetric voting' system (requiring 4 HOLDS and 0 HOLE FOUND) serves as a logical safeguard against adversarial input in the proof text.
[DATA_EXPOSURE]: The skill maintains state within a local hidden directory .evomath/. While it tracks 'positive' and 'negative' memory for proof strategies, these are contained within the local workspace and used only for logic optimization. No exfiltration patterns to remote domains were identified.

evomath-tao