βάσανος · the touchstone

A machine that refuses to fool itself.

Reliable AI-assisted science needs two parts: a guesser that proposes answers, and a cheap, ungameable checker — a lock — that grades them against the truth. Generating answers is cheap. Checking them is the hard part and that is what Basanos is for.

Basanos is the engine at the center of Aletheia that makes the locks — and it is built to call something a discovery only when it survives every honest check, and to refuse the questions that can’t be checked at all.

A pipeline deviation is not a scientific anomaly

A pipeline deviation is an unexplained internal mismatch: a strange residual, control failure, or contradictory number. It has zero authority to make a scientific claim.

Basanos logs deviations cheaply, triages them hard, and can route a survivor into a new question or test. Only the later investigated outcome can enter the public findings ledger. This keeps software defects, noisy measurements, and genuinely interesting residuals from being mixed together.

Three ways to anchor a lock to reality

You can’t manufacture truth — only a cheap path to a truth that already exists. Every real lock uses one of these. A question with none of them cannot have a lock, and Basanos refuses it rather than faking one.

⚙️

Computational

Free · perfect

Truth is decided by running code or logic. The lock just executes it.

a faster algorithm — run it and check

⏳

Temporal / reality

Cheap · needs discipline

Truth is decided by held-out real outcomes — freeze a prediction, reveal what happened.

score a forecast against the future

🧪

Physical measurement

Real · must be acquired

Truth requires a new assay or experiment you have to run.

does the drug actually work?

What it takes to be called a discovery

For the plain-English version of this lifecycle, see Methods.

A result has to clear every one of these before Basanos flags it for a human. Most things die at one of them — by design.

  • 1Freezes the question, data boundary, falsifier, and claim ceiling before the decisive test
  • 2Proves the checker on a known-answer positive and a known-answer negative
  • 3Beats the strongest boring explanation appropriate to that domain
  • 4Survives leakage, confound, and multiple-testing checks where they apply
  • 5Repeats on genuinely held-out data or a new physical measurement when the claim needs one
  • 6Receives a separate, dated novelty review before any discovery label

Where it stands, honestly

The reconciled public record currently contains 14 supported findings and 5 bounded nulls. These are different endings, but both are useful: one preserves a narrow result and the other records where a prediction failed.

It currently has 1 verified discovery. Supported evidence is not enough by itself: novelty must also survive a separate, dated review. This keeps a solid replication from being advertised as new science.

Why the verifier is the whole game

AI is about to flood science with generated hypotheses and analyses and papers. The bottleneck will not be generating ideas. It will be telling which ones are real without fooling ourselves. Aletheia runs Basanos so its own questions get answered that way and the same machine is what any lab would need to sort real results from generated ones.

See the research queue →Full methodsContact