Post Oak Labs

Research

The public record of our measurement work on AI models and engineering tooling: every benchmark page keeps the prompt as pasted and the reply as saved, with the grading method beside the result, so a reader can check any claim against the primary files.

This section is written for engineers and risk practitioners who want the primary record behind a published number, from the prompt that produced it to the pass that graded it.

Benchmark

Free Web AI Red-Team Benchmark

Sixteen free web chat models on ASIC physical-design and regulatory-compliance tasks, every prompt pasted by hand and every reply saved verbatim. Most of it was graded by mechanical recomputation against primary sources, across eight collection days and 663 responses, at zero spend. The evidence files live in the public data repository.

Read the eight days →
Repository

engineering-decision-schema

Shared JSON Schemas and a small validator for engineering-decision artifacts, under JSON Schema draft 2020-12. It is the one artifact several downstream repositories depend on, kept deliberately small and free of data.

GitHub →
Repository

openeda-decision-dataset

Flow harness, metric extractors and episode tooling for reproducible physical-design decisions built on OpenEDA, from environment setup through sweep tooling to the episode records the benchmark methodology draws on.

GitHub →
Repository

zkprof-harness

A browser-native profiling harness for zero-knowledge proving workloads, measuring prover behaviour where the workload actually runs for most users.

GitHub →

How to check a number here

Goal of this animation

Show how any number in this section leads back to its primary record: the grading method beside it, the prompt and reply in a public repository, and a check a reader can rerun.

Day 2 pageJCS vectors passedkimi12/12a published numbergraderrecompute in scratchPython, independentlyRFC 8785 canonical formfree-web-ai-redteam-benchmarkpublic data repositoryprompt as pastedreply as savedreview notesa reader reruns the check and gets the same 12/12engineering-decision-schemaJSON Schemas + validatoropeneda-decision-datasetflow harness + episodeszkprof-harnessbrowser prover profilingPublic repositories, each with its own README and licence.
Step 5 of 5

The tooling is public too

Three more repositories carry the schemas, the physical-design flow harness and a browser prover profiler that the wider research draws on.

Left and right arrow keys step back and forward once the player has focus. Link to a step with #xp-hub-step=N.
🔒 All inputs are processed locally in your browser. No data is transmitted. Do not enter real personal data — use synthetic or anonymised inputs only.