I break systems I actually use, trace failures to their root cause, and turn the evidence into fixes, tools, or upstream reports.

My current focus is offensive security around AI and developer tooling: real systems under real load, reproducible failures, preserved artifacts, and attribution that survives a second look.

Public lab: AI metrology · digital forensics · agent memory · security research · FS25 engine archaeology

• BRONCO — FC-001 · policy-driving evaluation provenance & reproducibility.
• schroedinger-sync — local-first AI data sovereignty · hardening under real workloads.
• MemPalace — upstream fixes and persistent-memory work from production use.

• 19 Aug — commented on issue EleutherAI/lm-evaluation-harness#3749
• 19 Aug — commented on issue SWE-bench/SWE-bench#621
• 19 Aug — commented on issue KeilerHirsch-Labs/BRONCO-AI-Metrology-Benchmarks-DIN-ISO-IEC#1

_{Auto-refreshed from public GitHub activity every 6 hours. Profile-maintenance commits are excluded.}

• Courseplay/Courseplay_FS25#1298 — fix: don't build the vehicle HUD on a dedicated server
• modelscope/evalscope#1587 — feat(metrics): Wilson CI + paired McNemar helpers
• tobi/qmd#901 — fix(embed): report actual completion and session expiry (additive)
• MemPalace/mempalace#1989 — fix(embedding): preload CUDA/cuDNN DLLs before the first GPU session

_{Selected automatically from my currently open public PRs, preferring one active contribution per upstream repository.}

• KeilerHirsch-Labs/schroedinger-sync v2.2.0 — published 19 Jul 2026 · SHA256SUMS

• Threat model / security policy — constraints, caveats, reporting path, and what the tests actually enforce.
• BRONCO Field Cases — evidence-linked measurement cases with falsifiers and explicit open questions.
• Release checksums are linked beside shipped releases when available.

• schroedinger-sync — local-first export and sync for claude.ai conversations, project docs, and memory. Windows, DPAPI + CDP, no telemetry, no cloud.
• BRONCO — AI Metrology, Benchmarks, DIN & ISO/IEC — research-first infrastructure for treating AI benchmarks as versioned measurement instruments, models/agents/harnesses as precisely defined systems under test, and results as measurements with provenance, applicability, and uncertainty.
• AFFLR_Anthropic-Failure-Forensics-Live-Radar — watches the public anthropics/claude-code issue space and prioritizes security/trust-boundary, evidence/provenance/integrity, and fresh critical signals for human review.
• MemPalace/mempalace — my daily persistent-memory layer for agent workflows. I contribute fixes found by running it under real-world load.

• Authorized targets. Offensive mindset.
• Hypothesis → reproduction → evidence → root cause → fix.
• Measure, don't guess.
• Evidence before attribution.