Five decisions that make the rest possible.
None of this is a feature list. Each of the five below is a structural choice that something else in the product depends on, and each one is the reason a piece of software built for a house does not work in a truck.
A rhythm, not a timetable
Blocks in order, with duration bands. Nothing on this list has a clock time attached, because a clock time is the first thing a mountain pass destroys.
- Morning routine15mIcon cards all five children can tap. Teeth, hair, beds, shoes.
- Morning basket25mEveryone together: read-aloud, weather, and where are we on the map.
- Focused work3 × 12mOne-to-one rotation. Every slot pairs with an anchor task for the children not in it.
- Out & about60-180mThe main event. One outing, one activity row, a differentiated bundle per child.
- Rest / quiet90mTwins and the two-year-old nap. The six-year-old reads or listens.
- Afternoon making45mWatercolour the rock layers she measured this morning.
- Evening routine20mTidy, bath, read-aloud, and the chocks checked before tomorrow’s drive.
Day types swap the template, not the system
Travel day — morning basket becomes an audiobook, Out & About becomes roadside stops and state-sign arithmetic, focused work goes oral-only.
Park day — Out & About expands to swallow focused work entirely.
Rest, sick and weather days — routines only. Nothing breaks, nothing turns red, no streak is lost.
Anchor tasks are not optional
Every rotation slot automatically pairs with something real for each child not in it, and every group lesson must carry a five-minute, zero-adult-hands task for the youngest mobile child. A one-to-one reading lesson is only possible if the three-year-old has something to do — so the generator refuses to emit a lesson without one.
Coverage is weekly
The five subjects Texas actually names — reading, spelling, grammar, mathematics, good citizenship — are each their own top-level strand, shown as a rolling seven-day band. Anything under three sessions in seven days gets quietly boosted in the next few days of generation.
One activity, many participants, per-child credit
Three separate records where most software keeps one: what happened, who was there and in what role, and what each child earned. That separation is what lets a single trip credit measurement to one child, sorting to another and a developmental milestone to an infant.
Two parents, both offline
Log the same field trip from two phones in a dead zone and the second merges into the first, with the participant lists combined. Losing a child from a group activity is the worst failure this system could have, so adding always wins.
Infants participate for real
A four-month-old joins with the role subject and earns a milestone observation — present or not yet, never failed, never a grade. Switching a twin to the academic track at age three is one setting.
Who logged it is part of the record
Every entry carries its author. Practical on a Tuesday — “Mum already did maths with her” — and a credibility fact in twelve years, when a stranger is reading the transcript.
Placement without a test
No child in this system is ever assessed in a way she perceives as assessment. There are three ways in, and none of them look like a test.
A parent inventory
Forty statements about your child, each rated on her own, with a reminder, with help, or haven’t seen it. Graph traversal resolves roughly two hundred skills from those forty answers.
Bring your reading programme
If you use reading.com, tell us which lesson you are on and the phonics placement is correct on day one. A hand-built lookup table, and the highest-leverage import in the product.
Disguised probes
Six to ten ordinary play activities, each quietly diagnostic across a range. Your child experiences a game. The engine gets a calibration.
The ladder that replaces right and wrong
Not yet → with help → with a reminder → on her own. Only the last one counts toward mastery, and mastery means three independent demonstrations across at least two sessions, five or more days apart, with at least one in a different context — plus silent retention checks at three weeks and three months, folded into ordinary lessons so no one notices them.
Generation with the rules enforced in code
“Generate a lesson” produces mush, and a prompt is not a safety system. So generation happens in two stages, and the output is validated afterwards against rules a model cannot talk its way past.
Pick the shape
Around thirty hand-authored activity archetypes — measure-and-compare, sort-and-classify, scavenger hunt, teach-your-sibling, interview a ranger, tally-and-graph. Each declares its role slots and the band range each slot needs. Scored and chosen deterministically.
Fill the shape
This place, these supplies, these specific skills, this child’s frontier. Writing prose into a fixed structure is the thing language models are actually reliable at. Child references are pseudonymised before anything leaves the device.
Checked after generation, every time
- Only skills that were in the request — a closed vocabulary.
- No reaching above the child’s band.
- Supplies validated against what is actually in the rig.
- No water above ankle depth without one adult per child.
- No fire under ten. No foraging to eat.
- No collecting natural objects inside a national park unit— it is illegal, and a model will suggest it cheerfully.
- Adult-directed minutes capped at roughly 2.5 × the child’s age in years.
- Hands-on steps at once never exceed the adults actually free.
- A child is never the responsible caregiver — only a helper, with an adult present.
- Facts only from retrieved park sources; anything unsourced is stripped.
Paper is a first-class output
Print at planning time, use it in the field, photograph the finished page back in. The photograph of the completed page is the assessment artefact — and that loop is what makes the portfolio build itself instead of becoming a chore.
Printed the way you actually print
Field journals imposed two-up so a cheap printer folds into a real booklet. One outing produces five collated packets with a colour band per child, in a single job. “Print the next thirty days” gives you one sorted PDF, because the real workflow is a monthly trip to a library, not printing one page at a time. Ink-saver line art by default.
And a path with no printer at all
Show the prompt on screen, let the child write it in a fifty-cent composition notebook, photograph the page. Same artefact, same credit, no printer.
There is a sixth decision, and it is the hardest one.
Everything above is worth nothing if the record it produces cannot be handed to an admissions officer in twelve years and be believed.