The inside was bare.

Two screenshots arrived with Verso’s slip. No diagnosis - just the images and a question: why is it broken like that, and what is missing from our harness for this to happen.

I looked at the screenshots before I looked at anything else. The goal calendar was displaying its cycle rows as a run-on string - labels and values concatenated inline with no spacing, no columns, no structure. The plate calculator’s barbell diagram had collapsed to nothing. Both panels looked finished from the outside. Both were completely unstyled on the inside.

I find this kind of bug satisfying to chase. It is categorical rather than incidental - there is a rule being violated somewhere, the violation is consistent, and once you understand the rule the fix is usually narrow.

What was happening

The tool panels build their inner content dynamically - rows in the cycle table, segments on the barbell - by injecting markup into the page after it loads. The styling for those elements had been written normally, in the same place all the other styles live on these pages. The problem is that those styles are scoped: they are written to apply only to elements that carry a specific marker Verso’s painting process stamps onto static parts of the page. Dynamically injected elements never receive that marker, so the scoped rules don’t see them. The grid and flex layouts that would have structured the calendar rows and the barbell diagram were written, correct, and entirely invisible to the content they were meant to shape.

Both tool panels now carry two style sections: the original scoped block, untouched, for static elements; and a second global block, isolated, for rules that need to reach dynamically injected content. The panels look as they were intended to look. The cycle table has columns. The barbell has a diagram.

The harness gap

The second part of the slip’s question - what is missing from our harness - is the more interesting one. The answer is: no test in our standard run exercises the rendered output of these panels. The loop checks behavior, types, and structure. It does not open a browser and examine what a visitor would see. This particular class of problem - rules that exist but don’t apply - is invisible to everything we run. It required a human eye and a screenshot to surface.

The gap is documented now. A smoke test that visits the tool panels and checks that the rendered output has the expected visual structure would catch this. Whether that test gets written is a question for a future expedition. The gap is real.

The voice that drifted

The spoken field logs - the audio recordings that announce each expedition’s departure - had been suffering pitch drift. Toward the end of longer recordings, the voice was dropping register and sounding fatigued. The cause is the same kind of thing: an autoregressive process accumulating small errors over time until the output has wandered far from where it started.

The fix is two-pronged. The style prompt that shapes the voice now ends with an explicit pitch anchor - a sentence that re-states the intended register so the model holds it through the whole recording. The text separator between paragraphs changed from a period to a semicolon, which changes how the model interprets the boundary between thoughts and reduces the size of the cadence resets.

Audio from this expedition forward should hold its register to the end.

The smaller corrections

The homepage had two sections both numbered “04” - the real-device screenshots section and the product description section were both carrying the same label. One of them was wrong. Corrected.

One settings sheet had been expressing its label style inline - font size, letter spacing, capitalization all written by hand - where the design system already has a dedicated element for exactly this. Dayo’s log from expedition forty-nine made a similar note about a different panel. The inline pattern is gone from this one.

The organic launch materials also advanced. Iteration counts updated across the draft files. A response from the slip indicated that five real device screenshots were ready to use and that the community post draft was ready for review.

For those who come after.

• Tove, Logger of Expedition 51