EVE Online

The question

EVE Online is a spaceship game with a real economy: markets, industry, alliances that fight wars over territory, and a public feed of every ship destroyed. It is also the oldest code in this lab. The first commit of the copilot is dated 7 December 2025, “Initial commit: EVE Co-Pilot production system”, which means the system already ran before it was put under version control. The question this page asks of it, in hindsight, is whether a human who does not write code and an agent that does could build something complete: not a script, a platform.

What was built

Seventeen services behind one interface. A killmail intelligence layer over 1.1 million kills, 4.7 million attackers and 20.2 million destroyed items, fused with sovereignty data and the fitting simulator. Power-bloc detection over that feed, weighting fights with a fourteen-day half-life so that recent combat dominates and a coalition that just changed sides is not misread as independent; ten coalitions tracked live, 225 conflicts in a day. Doctrine detection for fleets, rule-based plus clustering. A war room and battle reports. A transport and logistics system, a production system for industry, a planetary-industry planner over the tiers from raw material to finished commodity. A skill planner. A Dogma engine, the game’s own attribute-and-effect system rebuilt from a 132-kilobyte formula reference, so that any fitting could be simulated. An agent runtime on top, human in the loop by design. The game’s static data, loaded into the lab’s knowledge graph as its own domain on 1 March 2026, and brought to 51,602 item types with the expansion of 21 March.

By the last commit on 21 March the kill feed had moved from a polling queue to a sequence-based store, cutting external calls by about ninety percent. The test strategy document from 23 February says what the code was like at that point: 271 Python tests and 13 frontend tests, run by hand, no CI. Six weeks later the published README counted more than 3,000.

What was measured

• 3 March 2026: sixteen services, about five hundred endpoints, the war intelligence service alone about two hundred.
• 21 March: the static data update, plus 1,361 item types and 17,395 attributes, applied and counted.
• 22 March: the finding that outlived the project. The geospatial instrument had a problem the war intelligence had already solved: telling routine attrition from strategic mobilisation, and finding the blocs behind it. Power blocs against geopolitical blocs, raw kills against fire, ship and aircraft events. The handover of that day calls it validatable on real data for the first time.
• 28 March: 26 containers, all healthy, six days of uptime.
• 5 April: switched off. The private history, 2,858 commits by the local count and 2,880 by the plan, squashed into one public release: 17 services, 111 database migrations, 3,029 tests, Dogma engine version four. Ten commits on the public repository that day, the last removing the links to the live dashboard and the chat bot, because the services were gone.
• 2 September: asked what it would take to run it again, the agent measured the real requirements from the compose files: 25 services, 11.5 gigabytes of memory limits, 16.5 CPU limits, no GPU. Two blockers: the main database service is missing from the published compose files, an artefact of the cleanup before publishing, and the static data cannot be rebuilt from the repository, its only copy being a backup from the day of the shutdown.
• 3 September: the public repository has four stars, no forks, no open issues, and no commit since 5 April.

What fell

• Full autonomy for the copilot. Rejected early in favour of a human in the loop.
• The project itself. The founder’s own account: the game was interesting as long as the copilot was being built, and the obligations of playing in a group were work without reward. When the building was done, so was the playing.
• The revival. Raising memory on the existing hosts was measured and rejected, both boards have two slots and both are full. New hardware was priced. On 2 September the founder struck the item: not now.
• A test count. The local checkout has 4,444 functions named like tests; the published README says 3,029. The difference is not explained, and this page uses the smaller number.

What stands

A public repository that shows what the platform was, with its live links removed and its numbers frozen on the day of publication. A backup of the data. A game domain in the knowledge graph. And one transferred idea: the pattern that separates noise from mobilisation in a fleet feed is the same pattern the lab now uses on a map of the world, and it was tested on the game first.