The body of work, in public.
Client work is confidential by default. So the work we show is the kind you can verify yourself: published standards, open source you can run, research you can check, and pilots you can play. Every artifact below is real, versioned, and testable before a single email is exchanged.
open source
Code you can run today
github.com/ravnlab · public · MIT
ravnlab-eval-harness
The runnable form of the Plausible-Wrong Benchmark: the public 20-case set, the grading harness with named failure codes, and baseline regression diffing. No dependencies beyond Python 3, with an Inspect AI adapter included. Point it at your own answers file to score any system - ours, yours, or one you're deciding whether to buy.
$ git clone https://github.com/ravnlab/ravnlab-eval-harness $ python3 harness.py --cases cases.example.json --answers answers.example.json
pwmetrics
Reference metrics for confident-wrong failure: grounding, calibration (ECE), the overconfidence gap, and refusal correctness. Zero dependencies; the test suite ships with it.
on GitHub → MCP SERVER · MITreview-gate-mcp
A working MCP server that asks a human when it isn't sure - document extraction where uncertain fields go to a review queue instead of being guessed into your systems. Plugs into any MCP client.
on GitHub → RESEARCH · CC BY 4.0confident-wrong-field-guide
The methodology note behind the benchmark, with the six-mode taxonomy in machine-readable form - fork it, extend it, hold us to it.
on GitHub →research & standards
Published, versioned, citable
The standard we hold AI systems to, and the research behind it - published so clients, and anyone else, can hold us to it.
The Confident-Wrong Field Guide
How production AI ships answers that are fluent, coherent, and wrong - a cross-domain taxonomy built from documented incidents in law, medicine, finance, and construction, with every figure source-labeled.
read it → RL-PWB-1 · V1.0The Plausible-Wrong Benchmark
Twenty expert-anchored trap cases across four regulated domains, each pairing a fluent-wrong answer against a correct one - with the full grading protocol and results policy.
read it → RL-EV-1 · V1.0The RavnLab Evaluation Rubric
Five graded dimensions, twelve named failure modes, golden-set protocol, pass bars, and regression policy. Also available as a citable PDF.
read it → FREE TOOL · V1.0The Confident-Wrong Checklist
Six tests any AI team can run against their own product in an afternoon, with a self-scoring rubric. Genuinely free; run it before you ever talk to us.
run it → NOTES ON METHODField notes
Shorter, sourced writing on evaluation practice - including a reference build graded under our own rubric.
read them →interactive
Work you can play
The pilots are not descriptions of the work - they are the work, in miniature. The studio page shows how the same interactive muscle presents itself as client-facing craft.
The evaluation desk
Pick a domain - legal, clinical, finance, engineering - and grade real trap cases yourself. The fluent answer and the right answer diverge; see which one you'd have shipped.
play it → PILOT · LIVEThe voice director & the extractor desk
Direct one piece of raw material into three platform-native cuts; clear a pile of messy documents and watch the system ask a human at the right moment.
play them → STUDIOThe creative studio
Interactive explainers, launch and explainer films, and brand identity for AI companies - presented as worked, honestly-labeled case studies.
see it →