decision

ADR 0006: The static probe's STATIC verdict is advisory, not dispositive

Context#

ADR-0003 made the visual leg skippable and gave the decision a free mechanical signal: scene-detect plus pHash dedup already run as the first shell step, so a video yielding <= 3 distinct frames auto-degrades to transcript-only. ADR-0005 then froze that computation in tools/ingest.py probe, precisely so that "distinct" would not vary by agent.

Both were right. Neither anticipated the failure mode.

Scene detection measures whole-frame delta. A conference recording is usually a composited stream: a fixed background, a fixed logo, a fixed speaker-at-podium video inset, a fixed track footer, and a slide body occupying a minority of the pixels. When the slide changes, most of the frame does not. The delta never crosses SCENE_THRESHOLD, and the probe reports a continuously changing 20-slide deck as visually static.

Two occurrences, one root cause, at two different stages:

The second is the dangerous one, because ADR-0003 makes STATIC actionable: honouring it discards the entire visual leg and gates every node from the source single-leg by construction. On that source it would have destroyed three specific nodes whose numbers exist only on the slides - the skill-length lift curve (n11), "rewriting the description alone fixed 5 of 7 failures" (n13), and the 39.2% -> 91.6% case-study table (n28).

Why this is not just a bad constant. Lowering SCENE_THRESHOLD would trade one false verdict class for another (webcam noise and cross-fades would start reading as RICH), and worse, it would break the property ADR-0005 exists to protect: a verdict computed with different constants is not comparable to any previous source's. Five sources have already been measured with the current values. The metric is wrong for this input class, not mis-tuned.

The asymmetry that decides it. The two errors do not cost the same. A false RICH wastes one view call on a contact sheet. A false STATIC silently destroys a source's second leg, and the contract's own degrade rule says that once you skip the visual leg you must not retro-mark nodes corroborated - so the loss is not cheaply recoverable. An asymmetric error deserves an asymmetric check.

Decision#

A STATIC verdict is advisory. It must be confirmed by looking before it is honoured.

tools/ingest.py probe, on STATIC only, now samples 9 frames spread evenly across the runtime, tiles them into one chronological contact sheet, and prints the path with an instruction to view it before skipping the visual leg. Cost: one view call, on the branch that would otherwise throw away the entire visual leg. RICH is unchanged and costs nothing new.

The change is deliberately additive. The verdict, STATIC_FRAME_THRESHOLD, SCENE_THRESHOLD and PHASH_DISTANCE are untouched, so every past probe result stays comparable and no source needs re-measuring. Confirming a verdict is not the same as recomputing it - that distinction is what lets this coexist with ADR-0005 rather than eroding it.

Escape hatches, both explicit:

• --no-confirm skips the sheet and prints that you are honouring an advisory verdict unchecked, which must then be recorded in SOURCE.md.
• If the runtime cannot be read, the tool says the sheet could not be built and instructs the agent not to record a confirmed STATIC. Failure is never silent - the same rule that caught the swallowed ffmpeg exit code during ADR-0005's shakedown.

A user opt-out ("transcript only") still skips the probe entirely and therefore this too. Explicit instruction outranks the mechanism, as in ADR-0003.

Alternatives considered#

• Lower SCENE_THRESHOLD - rejected. Breaks cross-source comparability (ADR-0005's whole point), and trades a false-STATIC class for a false-RICH class without fixing the mismatch between a whole-frame metric and a partial-frame change.
• Add --crop to probe so it measures only the slide region - rejected as the default. You cannot know the crop region without looking first, which is circular, and a per-source crop reintroduces exactly the parameter variance ADR-0005 froze. frames --crop already exists for use after you have looked.
• Make the agent always eyeball a sheet before any probe - rejected. That is a token cost on every ingest to defend against one branch, and it makes the free signal not free.
• Leave it to the agent's judgement, documented in prose - rejected, and this is the ADR-0005 argument restated: the failure already happened twice with the rule in prose. The check belongs at the point of action, in the tool's own output, not in a paragraph the agent may or may not recall.
• Drop the probe entirely - rejected. It is right for the case it was built for (podcast, webcam interview) and costs nothing. The verdict is useful; its authority was too high.

Consequences#

Easier. A false STATIC now costs one view instead of a source's second leg. The 260726_dont-ship-skills-without-evals override becomes the documented default procedure rather than an ad-hoc save. Agents get the warning where it is actionable - in the tool's output - rather than in a contract paragraph.

Harder. A genuinely static video now costs one view call it did not before. That is the price of the asymmetry and it is small. probe also now depends on ffprobe for the runtime, which ships with ffmpeg and is already a hard dependency.

To revisit. If a third occurrence appears at a stage neither scene detection nor confirmation covers, the metric itself should be reconsidered - most likely by measuring delta over a centre-weighted region rather than the whole frame. That would be a constants change and would need its own ADR plus a re-measurement note.

Follow-ups completed in this pass: tools/ingest.py (implementation + a reproduction test on a synthetic templated deck: 9 distinct slides, candidates=0 -> STATIC, differences plainly visible on the confirmation sheet); AGENTS.md § "The visual leg"; prd.md §5.1 and change log; a dated line in ../log.md.