Chat is a drafting surface. The product is the artifact (morning card or kit). Validation bet #1: Morning Desk — semi-manual OK for first ~20.
Morning Desk pipeline (bet #1): vision extract → inventory ledger → archetype tone (self-report only) → templated morning card → email send. Soft Big Five = friendly tone labels — never clinical, never inferred from the photo. Subscription justified by email delivery of situated next-actions, not chat. Semi-manual OK for first 20.
Home Reset remains the heavier multi-photo path (rules engine + L/M/H kit) — not the hub lead during bet #1.
What each layer does
Honest separation of duties
01
Selected inputs
User-chosen photos and files only. Optional cultural inputs (playlists, boards, Letterboxd, reading history, recipes) stay opt-in. Never indiscriminate account ingestion.
02
Perception
Vision / OCR / classifiers extract candidates. Only confirmed or clearly user-stated items enter the ledger. Invented inventory fails the gate.
03
Ledger
Structured record of environment + problems + preferences. Source pointers stay attached so artifacts can show grounding.
04
Rules engine — plan structure
Deterministic scaffolding: step budgets, L/M/H variants, decision-rule templates, weekly-review slots, “no medical claim” language locks. This layer decides structure—not the LLM alone.
05
LLM — copy / extraction / assembly
Writes plain-language briefs, softens tone, helps extract from messy notes, assembles cards into the kit shell. May propose, never solely authorize the plan.
06
Human QC → artifacts
Operator checks gates, corrects ledger misses, approves delivery. One revision included on paid Home Reset kits.
Hard no: No sole-LLM planner. No diagnosis from images. No protected-trait inference. No “your photos show ADHD/traits of…”. No guaranteed symptom improvement.
Hard yes: Rules-structured plans · LLM for language & assembly · human QC · source grounding · organizational tool framing · Morning Card as primary wedge · Home Reset as heavier secondary.
Jev-class gates
Fail closed · measure correction rate
Grounding
Every inventory claim cites an upload or user statement. Ungrounded items are removed or marked unknown.
Scope
Plan only addresses stated problems + selected spaces. No lifestyle sermon. No scope creep into clinical care.
Energy honesty
L/M/H paths must be meaningfully different and completable. “Just try harder” fails.
Language lock
Block diagnostic / medical / guaranteed-outcome phrasing before delivery.
Preference respect
Support preferences (visual, low-cost, sensory-light, etc.) appear in the kit—not ignored for a generic template.
Safety
Self-harm or danger signals escalate to crisis guidance only—no coaching. External actions require explicit approval.
Privacy
No ad pixels on intake/results. Deletion path documented. Minimal retention for fulfillment.
Delivery shape
Must ship as artifact kit (brief · map · plan · cards · rules · review)—not a chat transcript.
Safety & privacy (stack)
NIST-aligned risk framing in plain language
Identify — what keeps breaking, from your words + selected inputs.