The Bishop's Stortford projectThree agents, one shared list, and three ways to merge it
Real features from OpenStreetMap, simulated agents and edits, and a result every learner can check.
OpenStreetMap records 424 benches, bus stops and post boxes in a rectangle around Bishop's Stortford: 216 benches, 170 bus stops and 38 post boxes. The 170 stops carry only 94 different names, so many stops share a name with another stop. The learner splits the rectangle into three overlapping strips and gives one to each of three simulated agents, A, B and C. In the first round each agent finds 80 per cent of the features in its strip, and after syncing the team knows about 352.
In the second round the agents work offline. A adds everything it missed. B decides benches are out of scope, deletes every bench it knows about, and adds its other missed features. C decides that stops sharing a name are duplicates, deletes the extras it knows about, and adds new stops. C's rule is questionable, because sharing a name does not make two stops the same stop, and that is part of the lesson: an agent can apply a bad rule perfectly. Then all three sync, in each of the six possible orders.
Merging the three agents' lists, one simulated run (seed 1), our Python code| Merge rule | Items after merging | Same result in every sync order? | Benches left | Same-named stops left |
|---|
| Whole list, last sync wins | 191 to 366 | No | depends on order | depends on order |
| Keep everything anyone added | 393 | Yes | 189 | 71 |
| Observed-remove set (a CRDT) | 172 | Yes | 18 | 21 |
| Rules agreed before starting | 133 | not a merge | 0 | 0 |
Letting the last sync win throws away whole lists: depending only on who connected last, the team kept anywhere from 191 to 366 items, and the worst order dropped 15 items that belong on the agreed list. Keeping everything never loses an addition, but deletions are impossible, so 189 benches survive B's decision. The observed-remove set does what its name says: each addition carries a unique tag, and a deletion removes only the tags the deleting agent had actually seen. Every copy reached the same 172 items in every order. The 18 benches and 21 extra stops that remain are ones another agent added while B or C was offline, so the deleter never saw them; the rule is that an addition you did not know about is not cancelled. Across 100 random runs the observed-remove result never depended on order, and it always ended 27 to 51 items away from the list the team would have made with shared rules.
Ages 8 to 11
Two players edit the same paper list in secret, then merge by three different rules. Which rule loses the fewest changes?
Ages 11 to 15
Code a last-write-wins list and a keep-everything set in Python, and find an order where each goes wrong.
Ages 15 and up
Build the observed-remove set with unique tags, prove every sync order agrees, then run the three-agent survey.
Sources and limits
Features are from OpenStreetMap under the Open Database Licence, as mapped on 30 September 2026. The agents, their strips, their 80 per cent success rate and their rules are invented by us, and the numbers come from our simulation, not from any real agent system. CRDTs are described by Shapiro and colleagues (2011).
