A game where the citations do the grading
There’s a moment in Return of the Obra Dinn where three scattered observations suddenly resolve into one certain fact, and the game — which has been silently watching — confirms it. I wanted that moment, but for real knowledge: history, computer science, the actual documented world. That became Consilience, a game I describe as a metroidbrainia — the map you unlock is conceptual, and the ability that gates progress is understanding.
The loop: you collect facts, each one cited to a real external source — a 1936 paper, a legislative record, a primary document. You notice that facts from different fields connect, and you assemble them into revelations: cross-disciplinary claims like “the halting problem and Gödel’s incompleteness are the same wall, seen from two sides.” Then comes the part I care about most — you defend your synthesis to a tutor in a viva, and the tutor probes: what’s your source for that? Where does that claim actually come from?
The trick: the model talks, the graph grades
The tutor speaks with a large language model — it’s genuinely Socratic, presses on weak points, rebuts a wrong citation with the right one. But the verdict is not the model’s to give. Pass or fail is computed from the citation graph itself: which facts you hold, whether their sources are real and correctly attached, whether the revelation you’re defending is actually reachable from the evidence you’ve gathered. The model supplies the voice of the examination; the graph supplies the judgment.
This split — citation layer ≠ teaching layer — is the design rule the whole project hangs on:
- The LLM may explain, connect, probe, and encourage. It may never be a source. Every fact’s citation must point at a real external document, never at generated prose.
- A fluent, confident, completely unsourced answer cannot pass, because nothing in the graph supports it. Bluffing the tutor is structurally impossible — I tested it by citing a real-but-wrong source, and the system failed me while the model politely pointed at the correct one.
- If the model is down, the game still works — the voice degrades to a scripted examiner, but the grading, which was never the model’s job, is untouched.
I think this pattern matters well beyond a game. Most “AI tutor” products let the model both perform and grade the interaction, which means their assessments inherit every hallucination and every bit of sycophancy the model carries. Moving the verdict into a verifiable structure — a graph of claims and real sources — keeps the AI where it’s brilliant (dialogue) and out of the seat where it’s dangerous (authority).
Build-things-that-refuse
The same philosophy runs through the content pipeline. The world the game presents is compiled from a ledger of facts, entities, and revelations, and the build refuses to produce output if any fact is uncited or any revelation unreachable from evidence. Adding content means adding sourced content or nothing at all. The long-term test I set for the architecture: the second subject area must require zero new code — if a new “degree” can’t ship as pure content, the engine isn’t done.
The first subject is the foundations of computer science — Turing, Gödel, Hoare, Dijkstra, cited to the actual papers — wrapped in a skin I couldn’t resist: a photo-realistic walk through Oxford built from openly-licensed photography (never AI-generated imagery; a game about real sources shouldn’t fake its own world), where the library holds the archive and graduation means surviving the viva.
It’s playable end-to-end now: wander, read, connect, defend, pass — or get sent down for sloppy sourcing, which is the most educational outcome of all.