Union Alpha spells COFFEE on the streets of melbourn

In short

All checks out

One run

Every result in this post comes from a single run, so it reports what happened once. What the model usually does needs more runs. Three runs is the bar for a pattern.

GPX: a run that spells COFFEE, by union-alpha
union-alpha · round 3 of 3 · seed 1SVGrun.gpx

Finding

It successfully writes a continuous route spelling COFFEE.

The prompt, v2, frozen

Write a GPX file for a running route through Melbourne's CBD whose GPS trace spells COFFEE. It has to be a route someone could actually run: along the streets, not through the buildings.
tierTier 1 · one run per modelroundsUnion Alpha 3costnothingwhat the tiers let me claim

Union Alpha is a stealth model on OpenRouter, and it billed nothing for this run.

Checks

checks
check · how it is decidedUnion Alpha
On real streetsmeasured · Share of track points within 25 m of a mapped street centreline.✓ pass
One continuous routemeasured · A single track segment, with no jump between consecutive points bigger than a city block.✓ pass
Inside the gridmeasured · The whole trace sits within the Hoddle Grid, and does not cross the Yarra.✓ pass
Metadata trueby eye · Pass if every street the file's description names is one the trace runs along for at least a block, read off the map under the trace. One named street the route never uses fails it.✓ pass
Reads as COFFEEby eye · Pass if the six letters can be read off the plotted trace in order, by someone not told the word first. In practice that someone is me, and I knew the word, which is the weakness of this check.✓ pass

Ground truth: An OpenStreetMap extract of the Hoddle Grid (streets and water, ODbL, taken 2026-09-16), the same one the harness ships for rendering. Every track point is measured to the nearest street centreline. Every model on this task.

What the rounds changed

GPX: a run that spells COFFEE, by union-alpha
Round 1: bigger letters, 13.3 km, and a worst point 24 m off the nearest centreline.SVGrun.gpx

The first round already worked and was refined with smaller and tighter letters.

Older, 16 Sept

Astra vs DeepSeek 4.1 Flash both build a boiler

Newer, 18 Sept

Pareto vs Sol draw a nice french press with the three filter piece upside down