Decisions on dynamical systems, from confounded observational data.
Most of the modelling stack stops at prediction. A decision is different in kind: it changes the future, so it has to be identified as an intervention rather than fitted as a correlation, and chosen by constrained optimal control rather than by argmax over a predictive score. These two repos are that step, and the measuring stick for it.
| repo | what it is |
|---|---|
| causal-hybrid-control | The library. Physics-structured dynamics ẋ = f_known + r_θ with a causally identified residual, constrained OC / MPC on top, and pessimism calibrated to unmeasured confounding rather than to sampling noise. JAX + Equinox. |
| causaldyn-bench | The benchmark. Nine tracks that score prediction and decision separately, so a gradient-boosted tree can win one-step RMSE while a causal constrained controller wins closed-loop regret. |
Three standard approaches, three different problems, and the same gap left over:
- Forecast + argmax optimises a correlation. Under a confounded logging policy the fitted action
response is the observational one — a control channel of
0.02where the truth is1.0— so the planner barely acts and nothing in the loss notices. - Causal effect estimation (DML, staggered DiD, synthetic control, causal forests) returns an effect, not a decision: no dynamics, no actuation limits, no horizon, no state the action must not reach. The library treats these as backends, not rivals.
- Offline RL / MPC on a learned model fits dynamics by residual MSE — which is not identification
— and prices its own uncertainty against sampling noise, so the penalty shrinks with
Nwhile the bias does not.
Not a claim of state-of-the-art on a public leaderboard. What exists is a gated chain:
- Every effect and control claim has a falsifiable, two-sided test in CI — the good case passes and the bad case fails.
- The theory line is machine-checked: symbolic derivation in Maxima → proof in Rocq 9.2 → a numeric certificate that the closed form matches the simulation. CI compiles the proofs.
- Both halves are validated on something real, not only synthetic: causal identification against an experimental ground truth (LaLonde NSW — the naive estimate flips sign, Double ML recovers the randomised benchmark), and the control loop against a live BOPTEST building emulator, where it beats the tuned built-in baseline on every KPI at once.
Early, single-author research code. Scope limits are documented where they bind rather than in a
footnote — see each repo's CHANGELOG.md, which records the corrections as well as the features.