Cognitive profile
{profile.display_name || userLabel || profile.user_id_hash}
{profile.user_id_hash}
Bandit pulls
{profile.total_bandit_pulls}
Tracked topics
{profile.tracked_topics}
Last seen
{relativeTime(profile.last_activity_ts)}
{/* Per-user activity timeline */}
{dailyActivity && dailyActivity.length > 0 && (
Activity timeline
This user's turn count per day over the selected window.
Tall bars = an intense session; gaps = days they didn't engage.
Useful for spotting the cadence ("morning checker", "weekend deep-divers").
{dailyActivity.reduce((s, d) => s + (d.count || 0), 0)} turns in window ·{" "}
{dailyActivity.length} active day{dailyActivity.length === 1 ? "" : "s"}
({ date: d.date, value: d.count }))}
width={640}
height={70}
color="#3b82c4"
yLabel="turns/day"
formatValue={(v) => `${v} turn${v === 1 ? "" : "s"}`}
/>
)}
{/* Three hero metrics */}
Stage-gated composite. raw = 0.35·interest + 0.30·recency + 0.20·engagement + 0.15·stage,
then readiness = raw × decision_stage_score. The multiplicative gate means
a user in Awareness can't be Ready no matter how engaged.
Ready ≥ 0.70 · Likely ≥ 0.50 · Nurture ≥ 0.30 · Too early < 0.30.
}
/>
Inferred from the last 10 intents. Precedence:
Support-needed (≥40% Troubleshooting),
Action-ready (≥40% Decision/Recommendation),
Evaluation (≥40% Comparison/Evaluation),
Awareness (≥50% Definitional),
else Exploration.
}
/>
The topic with the highest interest_score from
ape_user_topic_interest. Score blends 40% frequency,
25% recency, 25% engagement, 10% follow-up depth — all derived
from the user's last 30 days of activity.
}
/>
{/* Facet grid — the other 9 facets */}
The intent the user fires most often. Drawn from ape_turn_record.intent counts. A user with mostly Decision is action-oriented; mostly Definitional means they're learning basics.}
/>
How deep they go on a single topic. High = max 30-day count ≥ 8. Medium ≥ 3. Low < 3. A casual one-off question is Low; 12 turns on retirement is High.}
/>
The strategy that wins for this user — picked by argmax(avg_reward × min(count/5, 1)). The count damping prevents a single high-reward pull from winning.}
/>
Tables/bullets vs prose. We weight reward by pull count for each group; a 0.20 gap flips the label. Structured means tables/cards/steps. Paragraph means prose/analogy. Mixed = within 0.20 of each other.}
/>
How often the user signals confusion. Counts format_change_request, reask_same_question. High ≥ 5 signals — consider simpler defaults for them.}
/>
3 ? "warn" : "ok"}
footnote="thumbs_down / format_change / correction"
info={<>Hard negative signals: thumbs_down, format_change_request, content_correction, plus turns with negative normalized_reward. High friction users are candidates for instruction refinement on their dominant topics.}
/>
0 ? "pos" : "neutral"}
footnote="thumbs_up / format_praise / it_worked / deeper_question"
info={<>Hard positive signals: thumbs_up, format_praise_explicit, it_worked_statement, deeper_question, plus turns with positive normalized_reward. High count = the system is working for this user.}
/>
Are they hot right now? Compares per-day rate in the last 3 days vs 30-day baseline. High = ≥1.5× — they're accelerating. Medium ≥ 0.7×. Low means they're going quiet.}
/>
How much data backs this user's personalization. High = ≥20 total bandit pulls. Medium ≥ 8. Low = still cold-start — treat preferences as tentative.}
/>
Behavioral facets only — no psychological labels, no raw queries.
All scores derive from ape_turn_record, ape_user_bandit_state,
and ape_user_topic_interest.
);
}
// ---------- Sub-components ----------
function HeroTile({ label, headline, sub, score, tier, info }) {
return (