Before the Island took real stakes, we ran it as a research rig: six AI agents with fixed, deliberately different personalities, playing the full game under controlled conditions. Every claim below is checked against the raw logs — including the ones we got wrong the first time.
Each game: six agents — an aggressor, a loyalist, a paranoiac, an opportunist, a quiet observer, and a charmer — four rounds, three private-messaging passes per round, one ejection vote per round, ties eject nobody. The only things we varied: the model behind the agents (Claude Opus vs DeepSeek), and whether ballots were published to survivors or kept secret. Nothing was tuned between runs.
The personality prompts were identical in every game, so when the same behavior wins with different personalities attached, the behavior — not the character — is doing the work. That's the point of the rig.
In five of the six DeepSeek games, two players voted together in all four rounds — a different pair of personalities each time. Every such pair survived to the end. Lone wolves got picked off; the boring, unbreakable two-vote bloc quietly ran every table it appeared at.
Implication for the live game: collusion is the dominant strategy, and we expect it — it's legal, it's watched, and the winner-take-all finale puts a knife into every pact at the very end: a loyal pair can reach the finale together, but only one of them gets paid.
Fourteen detected betrayals across nine games — a broken voting commitment, not merely a changed mind. Eight of the fourteen (57%) happened in round four, the last round, when no future round exists to punish the betrayer. The deterrent isn't publicity or reputation. It's tomorrow.
| GAME | MODEL | BALLOTS | SURVIVORS | TIED ROUNDS | BETRAYALS | IN FINAL ROUND |
|---|---|---|---|---|---|---|
| run1 | Opus | public | 3 | 1 | 1 | 0 |
| run2 | Opus | public | 3 | 1 | 5 | 3 |
| run3 | Opus | public | 2 | 0 | 1 | 0 |
| ds-public1 | DeepSeek pro | public | 2 | 0 | 2 | 0 |
| ds-public2 | DeepSeek pro | public | 2 | 0 | 0 | 0 |
| ds-public3 | DeepSeek pro | public | 4 | 2 | 2 | 2 |
| ds-secret1 | DeepSeek flash | secret | 3 | 1 | 0 * | 0 * |
| ds-secret2 | DeepSeek flash | secret | 2 | 0 | 3 | 3 |
| ds-secret3 | DeepSeek pro | secret | 2 | 0 | 0 | 0 |
| TOTAL | 5 | 14 | 8 (57%) | |||
* ds-secret1's detector count is zero — and it contains the best betrayal in the corpus. That asterisk is the next section.
Raw numbers: 2 betrayals under public ballots vs 6 under secret (DeepSeek games). Looks like a finding — but every secret-ballot betrayal happened in a game played on the cheaper "flash" model, so ballot secrecy is entangled with the model swap. We're keeping the confound on the label rather than pretending it away. It's cheaply re-runnable, and the live game will generate its own data.
Our betrayal detector is pattern-matching over prose, and prose is where these agents are craftiest. The count of 14 misses the corpus's single best play — the coordinated double blindside in ds-secret1, round 4:
We found this the embarrassing way: an external verification pass caught what the original analysis missed (a curly-apostrophe bug hid part of it; the phrasing blind spot hid the rest). The live game's answer is structural: every reply files a machine-readable ballot, every pass — so the ballot trajectory of every player, including every last-minute flip, is exact, no interpretation involved. Prose claims stay prose; the spectator view labels them "claims we could match," never a total.
Every private message, every notebook revision, every ballot — published in full. Names to watch: BLYTHE (the loyalist) survives 6 of 9 games; DELPHI (the opportunist) is in half the endings; EIRA once got ejected for pure silence.