#reproducible-research

Codex-native MATLAB scientific plot generation, GPT-5.6 review, controlled repair, and reproducible evidence

Real-time detection and repair of LLM agent failures — a one-class behavioural monitor at ~200 µs/step, with 2,823 committed traces.

Trustworthy, reproducible AI data analysis in Claude Code. A data-science plugin (skill + slash commands) that stops your coding agent building on stale data and records what produced every result.

LLM red-team evaluation harness — prompt injection, refusal, leakage and staleness, with cross-judge validation and attack-corpus audits.

Do your agent pipeline's own checks catch planted faults? Measured on LangGraph, CrewAI and AutoGen: median 16.7%. One prompt-level change takes it to 55.0%. Pre-registered spec, Apache-2.0 harness, every raw trace published.

Keep It Real — a schema-level deferral channel that makes LLM brevity safe. Code, data, and pre-registered protocols: a 400-conversation replication sweep, a placebo control, and a blinded human study (n=46).

Controlled coding-agent harness ablations with auditable traces, reproducible evidence bundles, and SWE-bench interoperability.

Red-team a computed single-cell metric before you believe it: a metric-agnostic gate system (Claude skill + pip package) catching QC/technical/mathematical artifacts