I build evidence-first systems for tool-using agents: capture what happened, preserve provenance, test adversarial failure modes, and keep operator authority explicit.

Current focus: agent security, context injection, evaluation infrastructure, reproducible execution evidence, and local-first tooling.

GTDataworks is the main line of work: an evidence pipeline for consequential agent runs, from captured execution to sealed evidence, evaluation receipts, and derived products.
gtdataworks.com

• Agent Flight Recorder — local-first recording for observable tool-using agent runs: model calls, tool calls/results, state snapshots, checkpoints, exports, regression cases, and eval seeds.
• Crumple Zone — Firecracker-based bounded agent exercises with host-owned mediation, quarantined raw output, trusted event streams, replay, and evidence verification.
• GTDataworks Portlock — shipped Linux USB mass-storage policy tool with soft/hard locking, session-lock behavior, attempt logging, .deb packaging, and a public apt repository.
• DIF Defense — experimental behavioral prompt-injection detection that evaluates recorded agent behavior against a frozen set of forbidden actions.
• GTDataworks-Labyrinth — public runtime architecture for captured and sealed agent execution, including a separately constrained range model for cyber environments.

• Kimi K3 in C — PR #6 — merged checkpoint-integrity work introducing immutable Hugging Face revision pinning and checksum verification so same-size corruption cannot pass size-only validation. The upstream first-run consolidation later retained and further hardened this path.
• Kimi K3 in C — PR #13 — merged regression coverage for a silent trunk-streaming corruption failure where one-slot async prefetch can overwrite the layer still being computed. Adds a synthetic checkpoint-free fixture covering the one-slot guard, two-slot prefetch isolation, ring wraparound, and failed-read publication safety; the mutation check fails when the guard is deliberately defeated.

• Omega Harness — receipt-bearing, operator-authorized runtime sessions with tamper-evident lineage.
• OpenAI Safety Hunt — scoped canary-only research harness for indirect prompt injection, MCP/tool-output poisoning, and reproducibility receipts.
• Card Forge — bounded card/deck workflow tooling with explicit operator approval and provenance rules.
• Constellation — local evidence mapping and planned-vs-actual drift receipts for agent-assisted work.

• Evidence before claims.
• Preserve primary artifacts; derive downstream products with lineage.
• Treat untrusted context as data, not authority.
• Make failure states explicit instead of silently widening scope.
• Prefer reproducible receipts over plausible completion narratives.