Reproducing the anticipatory-coordination paper

Target paper: Anticipatory Coordination: Recovering Delay-Lost Performance with Peer-State Prediction.

Versions pinned for the canonical results

Fill the TBD hashes from the provenance.packages block of any cell’s committed experiment.json.

Environment setup

python3.10 -m venv .venv && source .venv/bin/activate
pip install --extra-index-url https://us-central1-python.pkg.dev/arboria-research/python-packages/simple/ \
    gossamer-threaded-intelligence==0.3.1
export ENGINE_MODE=inprocess
export PYTHONHASHSEED=0

The grid

The canonical run is the dcc_p3 batch: task × predictor × delay × seed, with the gossip primitive held fixed at N = 500.

Total: 200 cells. Delay 0 is excluded — anticipation only matters under delay. Fixed parameters match the phase-diagram run (energy_rate=0, fault_prob=0, dt=1, tau_sec=300, steps=1500).

The prediction hook

With a predictor configured, at each step the runner feeds the agent’s delayed peer view through the predictor, extrapolated forward by the delay horizon, and the coordination primitive acts on that estimate; the prediction is scored against the realized state (calibration) and never mutates ground truth. Run locally via scripts/run_paper_experiments.py --plan p3.

Expected outcome

Every predictor recovers delay-lost quality relative to none, with the ranking const-velocity ≳ linear > Kalman > none. On consensus at delay 40, none gives Q = 0.00 while const_vel holds Q = 0.66. The run produced 0 non-finite values.