Event taxonomy.

1 · The axes

Scroll the diagram sideways on a small screen.

Seven axes, derived from what the eleven events actually differ on. They are independent: an event is a point in this space, not a node on a tree. No two events share a path, and nothing sits at the left-hand end of every axis. Marner sits at three values of the last axis at once, and the plot shows the most persistent.

Considered and not kept: duration, because what matters is whether the day leaves a station unstaffed, and that follows from the other axes. A parallel visitor track, present at Marner and nowhere else. Programme, which sits above the event rather than beside it and is already a field. Indoors or outdoors, which is a venue property.

2 · What kind of evidence this is

3 · The structures, and where each came from

Every document type and recurring block found. Found in is provenance, not a recommendation. Literal section headings and column names are in the markdown.

Structures with exactly one instance are listed in the markdown: a structure nobody repeated is either a good idea that was lost or a one-off that should not be generalised, and the corpus does not say which.

4 · Where a hierarchy does exist

Not among events. Among documents, and two independent corpora agree. One event's information is cut by reader, and the cuts nest by granularity.

The lower levels are generated, not written. Marner's own template marks two of its three ops documents built from index.html, do not hand-edit. The cost of not doing that is visible in both corpora: World Play Day writes its escalation route twice in slightly different words, and Marner's Year 5 technique disagrees across four documents because it was dropped in one of them and not the others.

5 · The seam with the activity taxonomy

Settled on 18 August 2026: a station is a place with a table, a facilitator and a volunteer; an activity is a thing a child does; the station layer belongs to the event. The same rule already governs materials and reusable made things. Mapped here, not re-decided. See DECISIONS.md.

The Time row is the least worked out of the four. The activity layer's contribution to it exists in exactly one document in the whole corpus, a one-hour flow written in relative minutes rather than clock times, and nowhere else. Everything else is locked to a clock, which is why one plan writes its materials list out three times.

6 · Proposals

Each is grounded in something observed above, and each is proposed: written here, awaiting you. Nothing already settled in DECISIONS.md is restated.

Non-negotiable for an event record

The test is the one already used in event-template.md: a field earns its place if an output's arithmetic cannot resolve without it, or if it prints on the header of a list somebody carries.

• Whose day it is. One of the four values on axis 1, because it decides whose consent process, whose safeguarding policy and whose timetable govern, and whether an unknown is answerable by asking rather than deciding. Both events say this in prose; nothing parses it.
• How children meet it. One of the four values on axis 4, because children is a headcount under one value and a throughput under another, and every per-child quantity depends on which.
• A blank volunteer count as a legal value with a meaning. The two counts already stay separate. What is missing is that blank means nobody has established the number, which is ELM's state, and zero means there are none, which is Meath's. They currently look alike.
• Programme, where there is one. Every evaluation instrument in the legacy corpus is programme-driven and none is activity-driven, so the programme is what tells you which forms travel. Meath already carries the field.
• What is left behind. That a made thing comes back is already the event's call, and there is nowhere to record it except prose.

Not proposed as mandatory: a ratio, because no document in the corpus states one and inventing a figure would be guessing. A total adult count, because summing the two roles is the thing the safeguarding rule exists to prevent.

What every event should carry, for the system to be useful later

• A notes.md that is actually written in. The largest gap in the corpus, and not a design problem: the folder, the file and the rule all exist. NOTES.md already names the mitigation, one line on the pack list itself. That line is not on any pack list yet.
• A count of children who actually came, however rough. Six of the eleven events have no mechanism that produces one, and a record with no count can never be corrected. The only mechanism anyone has built is a blank box on a sheet somebody was already holding.
• A questions-for-<host>.md. Both events have one and both were useful. It costs nothing and it is the only artefact that turns an unknown into an answerable question.
• A follow-ups.md where a commitment was made. Without it, something agreed at an event lives only in a chat that ends.
• One source per shared block, with the cuts generated. Marner already does this for two of its three ops documents and records the rule. The ELM risk assessment is fed from each activity's ## Safety section, so it is already a generated artefact in principle.
• A dated line saying what state each artefact is in. Nine of nine legacy documents carry no version, no author and no date. It is one line, and it is the difference between a stale volunteer guide being spotted and being acted on at 8:30am.

7 · What this taxonomy still needs

What is missing, so it is in one place rather than scattered through the page.

• One post-delivery record, of anything. Until one exists, every structure above is a proposal and the axes are derived from intentions rather than outcomes.
• The independent evaluation output from Marner. An external evaluator observed a session and spoke to families and staff. Nothing from it is in the workspace, and it is the only thing in reach that could be evidence rather than habit.
• The class handout from World Play Day. The host school's own document. It is described precisely in the briefing deck and is not in the corpus, so the one full example of a three-way reader split is missing its third part.
• Which formulation of the volunteer rule the system carries. Question 3.
• Where the per-run detail lives when one activity runs several times inside one event. Question 4, and the only reason a third layer keeps being proposed.
• Where the activity layer's contribution to time lives. One document in the corpus writes an activity in relative minutes. Nothing says whether that is the model.
• The provisioning boundary: which materials the host supplies and which iOi supplies. Asked at Marner and never answered; visible again at Meath, where the centre saved cardboard of unknown quantity and iOi brought the base regardless. The activity-or-event cut is settled; this is a different cut and it is not.
• A scale specification for a build replicated across rooms. A Marner teacher asked how tall the trees had to be, so 23 rooms' work would fit one exhibition. Nobody answered. It touches the activity side too, so it should not be settled here alone.
• The activity taxonomy. Written in parallel and not present when this was finished, so the seam is mapped from the event side only.

Questions for you

Seven that would change what gets built. Answer whichever you have a view on and leave the rest. Then press Copy answers and paste the result into Claude, or download it as a file. What you type is kept in this browser, so you can close the page and come back.