Archive

What we measured, dated, and left as it was written.

Finished soak runs and investigations on Rytmi, Karhu TV's software playout engine, and on the hardware it drives. Each entry is written once, for an engineer who arrived with a specific problem and wants to know what was actually measured, on what, and when.

  • Every figure is dated. Each number carries the date it was measured, or is marked as a floor (≥) because it can only grow. Nothing here is a live reading: the home page carries current figures whenever a soak is running, and this archive carries what was true on the day.
  • Nothing is changed silently. An entry is not rewritten after it is published. When something in one turns out to be wrong, a dated addendum goes at the end saying what was wrong, how we found out and what replaces it. The original stays readable, so you can see what we believed and when.
  • The bad days are in it. Runs that failed their own criteria, figures we retracted and results that came out inconclusive are published the same way as the ones that passed. A passing number means little from someone who only shows passing numbers.
  • Each entry ends with what it did not settle. What remains open is part of the record, stated with the evidence that would settle it. We do not give dates for work that has not been done.

Entries

Newest first, by the date each was written.

  1. The card refused our audio for thirteen minutes in every sixty-eight: 16 September 2026

    New counters on the second unit lit up within three minutes: after certain clip changes the engine offered the DeckLink card more audio than a frame holds, and the card refused the surplus. A prediction committed an hour ahead pinned the cycle to the schedule, one clip giving two different shortfalls pinned the cause, a capture on the wire measured the gaps it left, and a change to when the engine rebuilds its audio sizing cleared it on both units.

  2. Late wakes near 400 microseconds, August 2026: we found where Linux had put the timer, not what made them late

    Two late wakes failed our fourteen-day soak in August 2026. They were the top of 109 excursions near 413 µs that recurred with the content and fell only in items handled one way on their way to the output. Kernel traces found where Linux puts the wake timer of a thread on an isolated CPU, but not what made the wake late.

  3. Genlocking a DeckLink Duo 2 at 1080p50, and pulling the reference: 27 September 2026

    We froze this test before the sync generator arrived. A 1080i50 reference was seen but not locked, 720p50 locked on both units, and the generator was measured against TAI. While locked, the engine discarded surplus frames nearly four times as often as when free-running. When we pulled the reference, its loss alarm reached the operator's phone, and nothing was watching the SDI output itself.

  4. Fourteen days of SDI playout, 61 million frames produced: the August 2026 soak, including the part that failed

    Between 6 and 20 August 2026 the first unit ran one unbroken 339-hour playout with no late frames on the SDI output, and still failed the criterion we had frozen before it started, because twice the engine's clock thread woke more than half a millisecond late.

    Addendum 30 Sep 2026