Explanation
How the coach decides
What the coach reads, what it may not say, how suggestions cite evidence, when fixed text replaces the model, and what auto-plan does without AI.
The coach is a set of narrow features over the family's own logs, not a general chatbot.
What it reads#
The week's logs (blocks, quotes, photos' captions, observations), the plan, the shelf, the ladders, and a curated set of 33 research summaries for evidence questions. It does not read the internet. Names are replaced with placeholders before anything is sent.
What it will not say#
A guardrail list blocks whole-word matches for diagnoses and labels (autism, ADHD, delay, disorder, diagnosis) and for rewards and punishments (sticker chart, time-out, reward, punish). A draft that trips the list is regenerated once with an explicit instruction; if it trips again, the coach returns a fixed refusal sentence. Weekly suggestion generation fails rather than falling back to unsafe text. An evidence answer that cites none of the supplied sources is rejected.
Suggestions cite#
Every suggestion carries the entries it was based on and shows them as based on chips. Accepting an invitation suggestion places it on the plan; accepting a coaching point writes it to the week and sends it to the adult on duty in their language. Suggestions are prepared on Sunday morning and on demand.
When there is no model#
If the coach is off, weekly suggestions and coaching drafts use fixed text, translation and prompts are unavailable, and evidence answers are off. Auto-plan itself is rules, not a model: ignored trays come off after fourteen days, replacements come from the two lowest-minute domains for the child's age, and one swap follows a strong recent interest. It works with the coach off.
The evidence answer format#
Four parts, always: Evidence, Steelman, Strongest critique, Design decision. The same balance the product's own research review uses, so the coach cannot become a cheerleader for one method.
Costs and limits#
A 30-second timeout per call, two retries on rate limits, a manual tool loop capped at four steps, and a limit of 60 streamed chats per family per hour. The audit log records every call with feature, model, and sizes.