Prove it works

Think of it when anyone, you or an agent, says “done”.

In everyday life

Like a pilot who checks the fuel gauge, not the fuel receipt. The receipt says fuel was bought; only the gauge says it is in the tank.

One case

An AI agent changes a shop’s discount codes and reports: “Done, all 212 tests pass.” Every test checks that a code is accepted; none checks the price the customer pays. On the live site, a 20% code takes off only 2%, and the first 6 customers pay almost full price.

212 tests passed, and every customer paid too much
The case, cell by cell1 cell = 1 order
Price after 20% off
402464208048
Price charged
4929.478.424.59858.8
Orders charged too much
6
Tests that passed
212
The tests checked that the code was accepted. Nobody looked at the price.

Every case

Every check is either the real thing or a stand-in for it. A passing build, tests with fake services and a report from the agent are all stand-ins: each one is a proxyA check that stands in for the real thing, such as “the build passed” in place of “the feature works”. It is cheaper, and sometimes wrong.. They are cheaper, and they usually agree with the real thing. The bug lives in the gap where they do not. An agent reports success in good faith, from what it meant to do, so its report is a proxy too.

The word

Engineers call this

verificationEvidence you saw yourself that the real thing works: the output, the page, the saved record. A claim, a passing build or “the agent said so” is not evidence.

Before you say done, check the real result directly: run it, open it, read the actual value. Report what you ran and what it showed, and report a skipped check as skipped.

Where this idea leads

Skills that encode it

  • principle-prove-it-works pstackVerify against the real artifact: run the feature, read the actual value. Not a proxy, a self-report or “it compiles”.
  • full-send tstackProof is evidence you observed yourself in this session. Green mocked tests do not prove that the user experience works.