Classify + Probe — how raw input becomes an S/M/L tierCapability

How JARBUS decides what to do with anything you drop in: classify.py reads the capture, assigns an S/M/L tier and an intent, scores its own confidence, and only bothers you when the decision is genuinely ambiguous. Today you'll find the one rule that keeps an orphaned Plaud recording from nagging you.

Every capture lands as a 'batch' — the text plus metadata (word count, content format, source). classify.py asks the Claude API for three things: a tier (S = quick one-shot, M = multi-step protocol, L = full project), an intent (action / note / proposal / etc.), and a confidence score 0.0–1.0. The tier maps straight onto a workflow: S → quick-deliver, M → run-protocol, L → spawn-project.

The load-bearing design choice is bias-to-ship. As of 2026-05-19 the default is auto-fire: the router just acts, because most routes are reversible and you can re-classify from the dashboard. It only stops to ask you (emits a 'route-approval' card) when should_auto_fire() returns False — i.e. Claude flagged requires_decision, OR confidence fell below CONFIDENCE_SANITY_FLOOR (0.35, meaning Claude is basically guessing).

But 'below the confidence floor' isn't enough to justify interrupting you. A second gate, should_hold_silently(), catches captures that are too empty or too thin to be answerable. The dominant case: an orphaned Plaud -recording.* file. The small transcript/summary siblings classify and archive before the ~1MB audio finishes downloading, so the audio arrives alone, looks like a fresh input, scores ~0.3 confidence — and without this gate would fire a card asking you to route an empty audio file. is_empty_capture() catches it (word_count 0, no content, format in {audio-only, needs_manual, unknown}) and holds it to data/holding/ instead. Nothing is lost; you just aren't pinged.

Worked example

Find the exact rule that suppresses the orphaned-recording card — this is today's hands-on answer:

grep -n "def is_empty_capture\|def should_hold_silently\|THIN_CAPTURE_MAX_WORDS\|CONFIDENCE_SANITY_FLOOR" \
  ~/Claude/JARBUS/scripts/classify.py

# Then read the two gates back-to-back:
sed -n '585,636p' ~/Claude/JARBUS/scripts/classify.py

# See what's actually been held (suppressed) recently:
ls -t ~/Claude/JARBUS/data/holding/ | head
cat ~/Claude/JARBUS/data/holding/$(ls -t ~/Claude/JARBUS/data/holding/ | head -1)

• The grep surfaces the four knobs that govern interruption: the two hold-rules, the 5-word thin threshold, and the 0.35 confidence floor.
• Lines 585–636 are the whole 'useful cards only' system: should_auto_fire (do I act automatically?) then should_hold_silently (if not, is this even worth asking about?).
• is_empty_capture() is the rule the curriculum asked you to name — it returns True when word_count is 0, content is blank, AND the format is audio-only/needs_manual/unknown. That's the orphaned Plaud recording.
• The holding/ directory is the proof it works: every capture too thin to route is logged there as JSONL, so 'silently held' never means 'silently lost' — you can re-drop with context to route it.

Gotchas

• A confident short task is NOT held. 'call dentist tuesday' is 3 words but high-confidence and auto-fires — should_hold_silently is only consulted on the route-approval path (when auto_fire already failed). Thin ≠ short; thin means low-context AND low-confidence.
• THIN_CAPTURE_MAX_WORDS = 5 is deliberately tiny. It's there to swallow stray fragments like 'Build out a report', not real briefs. Don't raise it to 'clean up the Inbox' — you'll start dropping legitimate small tasks.
• The 0.35 confidence floor and the hold gate are different layers. Below 0.35 = 'Claude is guessing, ask the human.' Hold = 'even asking is pointless, there's nothing to route.' A capture can pass the floor and still get held, or fail the floor and still be answerable.