The Compounding Test (AI-Powered)
Paste a description of your AI initiative; get a diagnosis grounded in The Berkshire Test for AI: one verdict in the 2×2 matrix, four condition scores with rationale quoting your description, the closest-fit portrait company from our research, and a failure-mode citation if your initiative lands in one of the three failure quadrants.
Prefer a faster deterministic version? The five-question diagnostic runs entirely in your browser — no AI, no waiting, no API keys.
Try a sample
Pick a sample or paste your own description, then click Diagnose. First request after idle takes 30–90 seconds (the GPU spins up); follow-ups are fast.
How this differs from the five-question version
The deterministic version asks five yes-or-no questions and returns a verdict from a fixed scoring function. It's fast, sharable via URL, and the math is fully auditable.
This AI-powered version reads your full prose description and produces a tailored writeup that names the bottleneck in your operation, quotes your description as evidence for each score, and explains the failure mechanism in operational terms if your case lands in a failure quadrant. The tradeoff is that it depends on a remote model — first call after idle takes 30–90 seconds to spin up — and the output is less deterministic.
Use the deterministic version for quick self-assessment; use this version when you want a richer, narrative-form analysis you can share with a colleague.
A note on privacy
Free options (Phi-4-mini on ZeroGPU):
your description is sent over HTTPS to a HuggingFace Space
(AshwinP/compounding-test),
which runs the model and returns the diagnosis. The Space code does not
write your input to disk and does not retain it across requests. As with
any hosted service, HuggingFace's platform may log request bodies as part
of its infrastructure; the Space owner can view these logs.
Premium option (Claude Opus with
your Anthropic API key): your browser calls
api.anthropic.com directly.
Your description and your API key
never touch mile-hi.ai or HuggingFace —
they go straight from your browser to Anthropic, with no intermediary. The
API key is sent only in the request header, never logged to your browser
console, never written to localStorage or sessionStorage, and never echoed
back in error messages (we redact it from upstream errors as defense in
depth, in addition to Anthropic not including keys in errors).
Don't paste anything you wouldn't paste into a third-party LLM service. If your description contains genuinely sensitive material, use the deterministic version instead — it runs entirely in your browser and never makes any network call at all.
Related reading
Anchor essay
The Berkshire Test for AI
Portrait · Compounder
Progressive's Risk-Selection Flywheel
Portrait · Compounder
Deere's Physical-World Data Loop
Portrait · Compounder
Mastercard's Network of Labeled Outcomes
Portrait · Compounder
Mayo's Outcome-Labeled Corpus
Sibling lab · Deterministic
The Compounding Test (five-question)
Get the weekly briefing