Monterey Bay AI Lab
Can this answer be defended?
Why even “what is the ocean temperature right now?” is genuinely hard — and a working prototype that answers it honestly instead of confidently. LIVE PROTOTYPE
Try it yourself — it’s live.The real working prototype, running on public data. No login, no mockups.
The problem
AI is being pushed into science, public agencies, and data portals with thousands of layers. But there’s a wall:
domain experts won’t stand behind an AI answer they can’t check. A confident chatbot number is worse than useless when
someone has to defend a decision to their director, a regulator, or a funder. Trust — not cleverness — is what actually blocks adoption.
Why even “what’s the temperature?” is hard
It sounds like a lookup. It isn’t. “Sea-surface temperature” is not one number. A buoy measures the water ~1.5 m down at a
single point. A satellite measures the skin of the surface over a ~1 km square. A third “blended analysis” smooths many sources into
a coarse 25 km grid. Here is what three trusted public sources actually reported for the same day and place:
15.4°CIn-situ buoy — a thermometer ~1.5 m deep, one point
15.3°CSatellite analysis — ~1 km surface cell
14.4°CCoarser analysis — 25 km cell, ~10 km away
A full degree apart — at different depths, different-sized areas, and a day apart in time. And the two satellite products
quietly ingest the buoy’s data, so they are not independent confirmations. There is no single “true” value.
What we built — and it’s running now
An agent that refuses to hide any of that. Ask it the temperature and, instead of guessing one number, it shows its work:
it pulls the real public sources, traces every value back to the raw data with a checksum, reports where they disagree and where they’re
secretly dependent, and returns one of SUPPORTED / QUALIFIED / ABSTAINED — abstaining when the evidence
can’t support a claim. A human expert can inspect the reasoning and correct it.
The live agent: one bounded question → a fully-sourced, qualified answer, with every limit surfaced before it answers.
What makes it different
Defensible, not confidentIt says supported, qualified, or “I can’t answer that” — and shows exactly why.
Nothing inventedEvery value traces to a raw public source, a timestamp, and a checksum.
Honest about dependenceThree feeds are not three independent votes — and it says so.
Expert-in-the-loopA domain reviewer can correct an assumption; the record is versioned and immutable.
Why it matters
It turns AI from a confident-answer machine into a defensible-evidence machine — the missing piece that lets
scientists and agencies actually trust and adopt AI for decisions where being wrong is expensive, because the answer arrives with its
receipts and its limits, and it knows when to stay quiet. Temperature is just the proving ground; the same discipline applies to
any question where someone has to defend the answer.