Three surfaces, three different calls

A visitor lands on sipnab.com, watches the terminal animation on the homepage, clicks through to /analyze, and presses “Load a sample call”.

Until 2026-09-01 those were three different captures:

The still and the animation come from two recordings of two different calls, so the page appears to cut between them as the animation loads. Then the tool the page is demonstrating hands the visitor a third call.

Everything was valid

This is the part worth being precise about, because it decides what kind of fix is possible.

Every asset existed. Every .tape rendered. Every capture parsed. Every link resolved. Every gate in the repository was green, and correctly so — nothing violated a rule.

What was wrong was an agreement between three files, and nothing expressed it. The three facts live in a .tape, an .html template and a .js fetch. No diff puts them side by side, and no reviewer reading any one of the three sees anything out of place.

A half-fix had already shipped for exactly that reason. On 2026-08-31 somebody pointed the hero still at a new capture and left the animation pointing at the old one, because the two values live in different files and only one of them was on screen.

Expressing the agreement

tests/demo_capture_agreement_test.rs is seven tests and reads three files. The two that matter state the agreement directly:

assert!(
    shown.ends_with(&offered) || offered.ends_with(&shown),
    "the homepage animation shows {shown} ({tape_name}) and /analyze offers \
     {offered}. A visitor who watches the video and clicks through is handed \
     a different call than the one they just saw."
);

and

assert_eq!(
    still, anim,
    "the hero still comes from {still} and the animation it swaps to comes \
     from {anim} ({anim_tape}). The page would cut between two different \
     calls as the animation loads."
);

Both messages describe what a visitor experiences rather than which constant disagrees with which. That matters when somebody hits this in six months and has to decide whether it is a real problem.

Reading a tape honestly

Getting the capture out of a .tape needs care, and the helper says why:

The -I argument and nothing else: a tape may mention other paths in its comments, and a scan that matched those would report a capture the recording never opened.

A scanner that reports the wrong file confidently is worse than no scanner. It would have made the three surfaces look like they agreed.

Two more of the seven cover the failure that reports nothing at all. A tape naming a moved or deleted capture still renders: sipnab prints an error, and VHS faithfully records the error. The result is a plausible-looking demo of nothing, and only a person looking at the output would notice.

The part that is about people

That gap is also why somebody reported the video as updated while the shipped asset was still the old one.

There was nothing to check against except memory. The assets were all valid, the tapes had all rendered, and the only thing that could have said “these three do not agree” was somebody remembering to compare three files by hand after the fact.

A claim with nothing to verify it against is a claim about what somebody intended, and it reaches the person listening as a claim about what shipped. The tests are what makes that difference visible before the sentence gets written.

Worth stealing

Look for facts that appear in more than one file and describe one thing. They do not show up in review, they do not show up in a diff, and they do not break anything until a user notices the mismatch.

The gate for that class is cheap — three file reads and an equality — and its value is entirely in existing, because the alternative is a human comparing three files every time any one of them changes.