The schedule thatcrosses things outand then lights upwhat she liked.
A behavioral-activation planning agent for clinicians, built on Qloo.
Agent trace plan e0adf88a · re-plan
- 1start week 2 · 7 slots · LLM qwen:qwen3.8-flash → groq:openai/gpt-oss-120b
- 2findOutings day=thursday kept=true rated=8 → 1 returned → Dun-Well Doughnuts ✓
- 6findOutings day=saturday hour=11 → 25 returned · review 1 → Ceremonia Bakeshop ✓
- 8placeSocial day=sunday hour=14 → 11 returned · refused R-KIND 2 → Bushwick Pita Palace ✓
Week 2: 2 kept · 5 new · 2 rekindled by a rating (Qloo's attribution, not causation)
Maya is a fictional case · live re-plan of 8 Oct 2026 · rekindle.edycu.dev
Maya · a fictional case
She can't name one thing she'd enjoy this week.

Where scheduling breaks
Scheduling breaks in three places.
1
She can't name anything she would enjoy.
2
Generic suggestions don't stick: go for a walk.
3
Nobody checks the place: open at the hour, near, in budget, alcohol-free in recovery.
README · The Problem · evidence base for behavioral activation: Richards et al., Lancet 2016 (COBRA)
The truth boundary
The model picks where to look. Qloo picks the venue.
Clinician
What she loved
3–6 loves, confirmed against Qloo.
LLM tool loop
Where to look
Rung, domain, query; one coaching line.
qwen3.8-flashQloo
The venue
Real places, explained by her loves.
/v2/insightsVerifier
What is true
PASS, REFUSE, REVIEW.
R-ID … R-HISTClinician
Approve week
Nothing reaches her unapproved.
Her ratings
P and M
≥ 7 a signal, ≤ 3 excluded.
19 Qloo parameters to 4 endpoints, counted from the code · ARCHITECTURE.md
Thursday · one outing
Open at the hour, near her, explained by her loves.
SOLO · 15:00
Dun-Well Doughnuts
place
because Phoebe Bridgers 0.54 · Fleabag 0.46
1421 m · Thu 08:00–18:00 · slot 15:00 ✓moved from 18:00 (closed then)222 Montrose Ave Brooklyn, NY 11206
R-ALC · bar_service, wine +3 more · struck when a constraint changed
- 1feature.explainabilityQloo's weights for her loves.
- 2R-OPEN · R-NEAROpen at its slot, inside her radius.
- 3R-ALCWhat it replaced stays, struck, with its tag.
In recovery
Struck in place. Refilled live.
1 · one clickOnly the affected outings, re-checked live.
2 · the notice3 outings struck in place and 3 slots refilled live · 2.2 s
The clinician gate
Nothing reaches her until a clinician approves.
approved · Thu 8 Oct · 09:34 UTC
HOME · 10 MIN · 19:00 · SWAPPED IN
We Need New Stories
book
approved
CLINICIAN · swapped out by the clinician
SOLO · 15:00
Dun-Well Doughnuts
place
approved
R-ALC · bar_service, wine +3 more · struck when a constraint changed
SOCIAL · 14:00
Juice Generation
place
approved
R-ALC · full_bar, beer +3 more · struck when a constraint changed
Approve, swap or remove each card, then Approve week. A swap re-runs one slot live; the old card stays as a CLINICIAN strike.
SOLO · 15:00
Dun-Well Doughnuts
place
↻ kept from last week, rated 8 · re-checked live
because Phoebe Bridgers 0.539 · Fleabag 0.461
SOLO · 11:00
Ceremonia Bakeshop
place
← because of Dun-Well Doughnuts, rated 8 · 0.296, +0.044 over Phoebe Bridgers (original love)
SOCIAL · 14:00
Bushwick Pita Palace
place
← because of Dun-Well Doughnuts, rated 8 · 0.307, +0.073 over Phoebe Bridgers (original love) · with Lucy Dacus 0.239
Recreated from the live re-plan e0adf88a, 8 Oct 2026. Maya is a fictional case. A thread needs a lead of ≥ 0.02 over her best original love; it is Qloo's attribution, not causation.
The re-plan
What she enjoyed comes back. What she didn't stays out.
Maya rated Dun-Well Doughnuts 8: kept, re-checked live, and threaded to two new outings. Juice Generation, rated 2, stays out.
Receipts
Measured, with the caveats left on.
0of14
named venues wrong for Rekindle; the same LLM alone: 7 of 17.
LLM alone7 struck · 3 dashed
Rekindle0 struck · 6 dashed
0of30
in-recovery outings carry an alcohol tag or keyword, screen on.
4 bar tags21 of 28
14 tags3 of 28
+ screen0 of 30
Counts tags and keywords, not menus.
16/16
eligible re-plans put a ≥ 7 rating atop Qloo's attribution for a new card, at least 0.02 ahead of the profile's original loves (an ember thread's lead).
4 of 20 profiles had too few verified outings to rate. Rated ≤ 3 that came back: 0.
60,000
generated cases: in none did a passing outing break a hard constraint. 534 tests.
Manual audit: a random 20 per arm, checked one by one against public listings by our build agent (Claude, an AI; not blind); struck = wrong, dashed = inconclusive. The bench run of 7 Oct 2026 predates 31 later plan-path changes (listed in DEMO.md).
Honest limits
What it does not do.
“It is not a clinical tool.”
“A plan is slow: p50 is about 17 s from click.”
“The alcohol screen is a deny list.”
“The ember thread is Qloo's attribution, not causation.”
Her sheet
What a clinician gets.
Only approved activities, with empty P and M boxes. No weights, rules or strikes reach her.
Plan Maya's week, live.
no login · fictional cases · /judge · /integrations/qloo · /about
github.com/edycutjong/rekindle