tapir/scripts/spike-media

spike-media — the /bygge toolchain, recovered

Throwaway spike tooling for #28 (upload → transcript → takeaways). Committed because it was found in tmpfs one reboot from gone, not because it is finished.

Nothing here is wired into tapir. No database, no HTTP handler, no ADR.

What produced /bygge

The prototype at tapir.d-ma.be/bygge was not hand-built, which the epic and the first round of spike issues both got wrong. It came from these three scripts:

IMG_1233.mov --transcode.yaml--> IMG_1233.web.mp4   (k3s Job, h264, +faststart)
IMG_1233.wav --whisper---------> IMG_1233.sv.srt    (interactive — the one manual step)
IMG_1233.sv.srt --analyze_srt.py--> analys.json     (berget/mistral-medium, temperature=0)
analys.json + video --build_page.py--> page HTML

The reference analysis was produced by berget/mistral-medium — a mid-tier cloud model, not a frontier one. That matters when judging what a local model has to clear: the bar is mistral-medium, and it should be re-run as the control arm rather than excluded.

The validator is the valuable part

analyze_srt.py grades its own output on four checks. Two of them catch classes of error that timestamp- and citation-checking cannot:

Check	Catches
citation exact / partial / absent	Paraphrase presented as a quote. Three tiers, because alternativ citations legitimately stitch non-contiguous passages — a strict substring test fails the reference output.
TALET … FINNS INTE I CITATET	A wrong number carried by a verbatim citation and a real timestamp. Converts Swedish number-words to digits first. It caught 45×230 where the transcript said 220.
TACKNINGSLUCKA	Spans of the clip no section covers. The recall term — without it an empty analysis scores perfectly on every other check.
timestamp exists in the SRT	Fabricated seek targets. Near-solved: 74/75 on the reference, because the model copies the cue starts it was handed.

Known defects — read before running

• max_tokens: 6000 is too low. The 4-minute reference analysis is ~4,700 output tokens, i.e. ~78% of budget. A 10-minute clip truncates for every model, and truncation reads as a JSON-discipline failure if you are not counting tokens. Record finish_reason and separate truncated from malformed.
• build_page.py hardcodes the title ("Jonas – genomgång av tillbyggnad").
• The schema carries fields nothing renders — projekt.sammanfattning, projekt.skede, avsnitt.bildtid, avsnitt.bildmotiv, material[], oklart[] — left over from an earlier still-frame/markdown output that build_page.py replaced. They are paid for in output tokens on every run.
• Speaker attribution is inferred, not diarized. The output attributes statements by name from an SRT with no speaker labels. All four checks above are blind to a misattribution.
• The web encode does not downscale. The source is 720×1280; -preset veryfast -crf 24 on handheld motion is what makes it ~5 Mbps, not the resolution. The lever is the preset.

Fixtures

fixtures/fake.sv.srt + fixtures/fake.analys.json — a synthetic 7-cue Swedish construction walkthrough with planted numbers. No real people, no real project. It exercises the parser and all four validator checks without a model call, which is what makes it usable in a test.

Why the real artifacts are not here

This repo is public. The prototype's transcript and analysis are a verbatim recording of a named person discussing a client's building project — private third-party content, and exactly the class of data #28 calls out as the reason uploaded transcripts cannot live in the shared transcripts table.

So IMG_1233.sv.srt, payload.json and the model-comparison runs stay out of git here. That leaves /bygge without a versioned source of truth, which is a real gap and a decision for #28: a private repo, an encrypted blob, or accepting that the reference artifacts live outside version control.