A narrated, AI-generated walkthrough of the whole finding — the same story as the three acts below, condensed to under two minutes. Good starting point if you're new here; everything it references is backed by the numbers and proofs on this site.
One synthetic patient, three synchronized views: a causal graph that lights up when treatment information flows into the state, a state vector that flips cells between clean and treatment-confounded, and a predictability gauge that moves with it — all three driven by the same mechanism proven in Act 2 below, not a staged animation.
Open the live visualizer →Six state designs, tested on the same task: predict the next vasopressor decision from the current state. The x-axis groups them by physiological purity — does the state still contain a raw, treatment-derived SOFA feature that fails Theorem 1's non-identifiability check? The y-axis is each variant's actual, already-verified action-recoverability AUROC — no new computation here, same numbers as the Evidence page.
F is the point that matters here. It sits in the same purity bucket as D — same construction logic, no raw dose feature either — yet it has the highest action-recoverability AUROC of all six variants, above even the untouched full state (A). Purity and low recoverability visibly are not the same axis: D and E's designer might reasonably expect "no raw treatment feature" to mean "hard to predict the next action from," and for D and E that expectation holds — but F shows it isn't a guarantee. Two purely temporal history features (duration on vasopressor, time since the last dose-tier change) reconstruct the next action better than physiology does, without reintroducing a single raw SOFA-embedded dose term. Full write-up: Evidence, Experiment 8.
Everything in Act 2 is proven or directly measured. H3 asked: does the AUROC gap in Experiment 2 actually reflect the mechanisms proven here (Theorems 1–2), or is it substantially explained by ordinary treatment persistence — vasopressor courses just don't change from one decision point to the next very often, independent of any SOFA-construction quirk? The pre-registered paired-bootstrap test has now run: Delta = 0.147, 95% CI [0.140, 0.154] — Prediction A. Full results: see Status.
That result doesn't settle which mechanism produces Delta. Act 2's Variant F result (Experiment 8) was already corroborating evidence for the persistence explanation, and exploratory work in review (a synthetic ground-truth simulation, not pre-registered) goes further: the same decision rule fires whenever a state encodes a currently-running treatment under realistic persistence, whether or not a SOFA-specific construction effect is present. So Prediction A, on its own, cannot separate the SOFA-construction mechanism from ordinary treatment persistence — that's what the exploratory work points to, and it's the actual open question this project is built to answer next.
Deliberately not built: dual RL agents, live Q-value comparisons, or any policy-level visualization. No RL policy has been trained anywhere in this project — Experiment 8 used classifiers to test action-recoverability and mortality-predictive validity, not a trained policy. A demo element implying otherwise would misrepresent the project's actual state, regardless of how compelling it would look. Those become buildable once a policy-sensitivity study (the deferred Experiment 4) actually runs.