Receipts

Research & Receipts

Every claim made anywhere on this site, with a link to the primary evidence: published papers, merged pull requests, and case studies. If something can't be linked, it isn't listed here.

Published benchmark work

CL-bench: a context-learning benchmark arXiv:2602.03587

1,899 expert-authored tasks with 31,607 verification rubrics. Contributed evaluation tasks across Gaming, Coding, and Mathematics; served as strategic project lead on the delivery engagement.

CL-bench Life arXiv:2604.27043

405 human-curated context-task pairs and 5,348 rubrics grounded in messy real-life context. The strongest of ten frontier models solves only 19.3% of its tasks.

Open-source contributions to evaluation frameworks

inspect_ai (UK AI Safety Institute, 2k★)

5 merged PRs, hand-merged by JJ Allaire

  • #3690 — Judge-prompt injection hardening (22 tests)
  • #3794 — Pluggable URI resolver registry for custom schemes
  • #3836 — HuggingFace dataset retry on transient errors
  • #3722 — Anthropic Opus 4.7 sampling-parameter guard
  • #3668 — vLLM provider restart-after-close via epoch invalidation

lm-evaluation-harness (EleutherAI, 12.5k★)

Fail-fast guard preventing a silent ~17× accuracy regression on MoE models

  • #3734 — data_parallel × expert_parallel incompatibility guard

Python scientific stack

8 merged PRs

Open-source tools

judge-calibration

Zero-dependency Python toolkit that measures whether an LLM-as-a-judge agrees with human labels, and catches the four biases a good correlation score hides: position, verbosity, self-preference, and sycophancy. Exits non-zero so it can gate CI directly.

Writing

Talks & podcasts

None yet. If you run a podcast or meetup covering AI evaluation, agent benchmarking, or what actually breaks frontier models, get in touch — I'll bring real numbers rather than opinions.

Want this rigor applied to your product?

The same discipline that gets tasks into a published benchmark is what I build into client evaluation systems.