Independent agent-tooling builder. I ship deterministic, no-LLM software that agents and teams can trust: 4 projects live in the Official MCP Registry (skill security audit, evidence-gated claim reporting, benchmark-grading hygiene, secret scrubbing), all MIT + fully audited. I also documented and fixed a real benchmark-harness bug (VulcanBench #79) where a leaked pytest --cov option silently scored passing tasks as failures. I build custom MCP connectors, agent-verification layers, and eval/benchmark hygiene for teams that need auditable AI tooling — reproducible outputs, no LLM in the hot path.
by Robert Garrison
Detect pytest config-leakage that silently corrupts agent-eval grading: repo-root --cov/--cov-fail-under/--maxfail addopts leaking into workspace runs score functionally-passing tasks as 0.0. Deterministic, no-LLM.
by Robert Garrison
Force evidence-gated, honestly-tiered claim reporting. Classifies any agent claim into FACT / INFERENCE / SPECULATION / UNVERIFIED based on attached evidence, blocks fabricated 'it's done' completions. Deterministic, no-LLM, no-network.
by Robert Garrison
Deterministic pre-install audit of any agent skill/prompt against the 8 malicious-skill supply-chain patterns (R1 injection, R2 exfil, R3 secrets, R4 dangerous commands, R5 obfuscation, R6 untrusted fetch, R7 credential access, R8 privilege escalation). No LLM, no network — PASS/FLAG verdict with line-level evidence per finding.