The Prisoner's Dilemma — the engine sits down to play

Two agents, one dilemma, every round simultaneous. The classic roster is here — trusting, tit-for-tat, grim trigger, tit-for-two-tats, pavlov, suspicious, random — and so is an engine agent: the runtime as a PLAYER, not a referee. Each round it sends the interaction history over the wire and plays the honored answer, under a hard per-match ask budget. A refusal is not an error: the mood rule holds and the rung says so. Noise garbles transmission — opponents retaliate against moves that never happened — and the R/S/T/P matrix is editable mid-theory, because whether defection pays is a dial here, not a law. Run the round-robin and watch cooperation survive or collapse.

What decides what

  • The engine agent is a player, not a referee. Each round it sends the interaction history — payoffs, the opponent's seen moves, the cooperate rate, the matrix — over the full wire contract and plays the honored answer on the engine rung. In engine-vs-roster play the roster side is a rule every round, and the rung count shows both sides.
  • Refusal is a normal outcome, on the record. An abstain, a malformed answer, or a spent ask budget hands the move to the MOOD RULE — cooperate iff at least half of the opponent's last 8 seen moves are cooperation — and the rung says rule. There is no captain in this game: the engine agent IS the player.
  • Noise garbles transmission, never the ledger. A flipped move is what the OPPONENT sees; the scorer pays the true move; every flip is logged and marked in the ribbon. Tit-for-tat retaliating against a cooperation that never arrived is the Axelrod point, and the chart shows it happen.
  • The classics are test-pinned. suspicious beats trusting 500–0 at T=5; tit-for-tat ties itself 300–300 at full cooperation; a temptation below reward (T=2) flips the moral; the seeded round-robin replays its table exactly. 18 core tests run in CI.

Honest edges

  • The mood rule is a stub on purpose: it is the control arm, so "engine vs rule" comparisons mean something. It opens cooperative on round 1 (no history) and defects once the seen window turns majority-defection — nothing more.
  • The ask budget is per MATCH, not per side: engine-vs-engine shares one pool, so a budget of 2 is spent by round 1 and both agents finish on the mood rule. The stats panel shows tokens left.
  • Noise cuts both ways and the game does not rescue anyone: at noise 0.3 even tit-for-tat pairs bleed into mutual defection — that is real Axelrod-type dynamics, not a bug. Slide it back to 0 and watch cooperation recover.
  • The tournament is headless and synchronous: the current matrix, rounds (capped at 100 per pair), and noise apply to all 28 pairings, and the engine joins with a fresh budget per pairing. Strong results there are evidence about the MATRIX, not about your machine.
  • The scoreboard is session-only, auto-recorded every 50 rounds and at each match end: close the tab and it is gone. Nothing leaves this page.