Agora Calibration Runs

Question

What calibration evidence exists before Agora can claim predictive or control value?

Context

Agora is a vault-native agent-based simulation OS and Club B measure surface for the Recursive Self Improvement Lab: populations evolve under Markov rule dynamics, forecasts are scored against target distributions, and every world, rule, run, and channel is plain OKF files. This report records the first honest calibration numbers — small, provisional, and public so the loop can fail in the open. Companion negative result: CompactCTM shadow stage-1 failed held-out prediction against an empirical prior.

Findings

• A 30-tick opinion_diffusion run ended at adopt 0.600 / undecided 0.388 / reject 0.013 (run-fbaf4427f87ab2ac).
• On the SIR epidemic world at tick 40, susceptible held at 0.838, infected 0.0, recovered 0.163.
• A Markov-plus-Kalman forecast of susceptible over a 10-tick horizon recorded error 0.4513 against its target.
• Later deterministic cron loops on opinion diffusion stabilize near adopt 0.525 / undecided 0.438 / reject 0.037, with evolve improving forecast error from 1.1207 → 0.9529 (best fitness 0.5121) and writing an evolved 10-rule ruleset.
• Closing a measured adoption gap by lowering peer_strong adopt probability 0.60 → 0.50 moved final adopt 0.800 → 0.787 and predict error 2.8272 → 2.4648.
• Rule evolution can reduce error on named worlds without establishing external forecasting superiority or control authority.

Method

Runs use python3 -m agora run, predict, and evolve against named worlds (opinion_diffusion, sir_epidemic) with explicit target.json trajectories. Events and metrics append as JSONL under runs/<id>/. Claims here are local calibration demos, not survey replacement or external predictive superiority. Promotion of any learned controller stays blocked until a held-out gate clears (see CompactCTM shadow stage-1).

Limits

• These are local calibration demos, not evidence of external forecasting superiority.
• The reported rule changes are tied to named worlds and targets, not general social dynamics.
• A public notebook waits on a frozen, privacy-safe metrics snapshot.
• Improving in-sample forecast error is not archive-conditioned inheritance on held-out families.

Artifact

Numbers come from closed evidence notes and coordination records in the private Agora trees (agora, agora-agent-brain). No public notebook is attached yet; the next publishable surface is a sealed metrics card once a clean public snapshot of a named run is frozen.