[QA] Emergent Observations surfaces stop-word noise ("about", "talked") #166

New issue

Closed

opened 2026-09-10 00:32:09 +00:00 by Mark · 0 comments

Mark commented

2026-09-10 00:32:09 +00:00

Owner

Finding

The predictive engine at src/lib/utils.ts (the computeDashboardObservations / generateObservationSuggestions pair) extracts keywords from interaction notes/topics and surfaces them as Emergent Observations when the same keyword appears across multiple entries.

The stop-word filter at utils.ts:541-578 is missing common English verbs and prepositions, so they pass through as if they were meaningful topics.

Evidence (real Kindle run, 2026-09-09)

Seeded 32 interactions for a single person over 6 months on a Kindle Fire HD with the current main APK. The home page surfaced 6 Emergent Observations. Two were noise:

You and Jordan Reyes both seem into about — 13 interactions logged
You and Jordan Reyes both seem into talked — 8 interactions logged

Both ranked above a real signal:

You and Jordan Reyes both seem into travel — 1 interaction, 10 conversations logged

So a verb the user used 8 times ("talked about X") outranks a topic they explicitly discussed 10 times. That breaks the "what's actually important here" framing the surface is supposed to provide.

Reproduction

adb shell pm clear com.kin.prm
# Seed 32 interactions with notes containing the word "about" 8+ times
# Open home → Emergent Observations card → see "about" and "talked" ranked

Suggested fix

Expand the stop-word list at utils.ts:541-578 to include at least:

Verbs: talked, said, told, mentioned, asked, thought, felt, knew, wanted, needed, liked, loved, hoped
Prepositions/articles: about, around, through, after, before, during, between
Generic nouns: thing, things, stuff, time, times, way, ways, day, days

Severity

Low — does not break functionality, but degrades the predictive signal-to-noise ratio. The Emergent Observations surface is supposed to surface themes (hiking, photography, cooking), not narration verbs.

## Finding The predictive engine at `src/lib/utils.ts` (the `computeDashboardObservations` / `generateObservationSuggestions` pair) extracts keywords from interaction notes/topics and surfaces them as Emergent Observations when the same keyword appears across multiple entries. The stop-word filter at `utils.ts:541-578` is missing common English verbs and prepositions, so they pass through as if they were meaningful topics. ## Evidence (real Kindle run, 2026-09-09) Seeded 32 interactions for a single person over 6 months on a Kindle Fire HD with the current `main` APK. The home page surfaced 6 Emergent Observations. Two were noise: ``` You and Jordan Reyes both seem into about — 13 interactions logged You and Jordan Reyes both seem into talked — 8 interactions logged ``` Both ranked **above** a real signal: ``` You and Jordan Reyes both seem into travel — 1 interaction, 10 conversations logged ``` So a verb the user used 8 times ("talked about X") outranks a topic they explicitly discussed 10 times. That breaks the "what's actually important here" framing the surface is supposed to provide. ## Reproduction ```bash adb shell pm clear com.kin.prm # Seed 32 interactions with notes containing the word "about" 8+ times # Open home → Emergent Observations card → see "about" and "talked" ranked ``` ## Suggested fix Expand the stop-word list at `utils.ts:541-578` to include at least: - Verbs: talked, said, told, mentioned, asked, thought, felt, knew, wanted, needed, liked, loved, hoped - Prepositions/articles: about, around, through, after, before, during, between - Generic nouns: thing, things, stuff, time, times, way, ways, day, days ## Severity Low — does not break functionality, but degrades the predictive signal-to-noise ratio. The Emergent Observations surface is supposed to surface themes (hiking, photography, cooking), not narration verbs.