Endpoint-audited F-pre research model · v3.0.0
Make the estimate.
Keep the evidence visible.
Explore literature-derived PHBV CO2 mineralization with an exact browser-native random forest, a kinetic reference, and two deliberately visible uncertainty boundaries.
1,031endpoint-audited observations
11independent source studies
17.35exact RF curve-held-out MAE
26.06exact RF new-study MAE
The exact RF did not outperform the global time-only baseline in the cross-study task. Predictions for an unobserved study remain insufficiently reliable for quantitative decisions.
Input profile
Describe the recorded scenario
Ranges describe marginal coverage in the training corpus, not joint applicability.
The exact browser model is loading.
RF methodological estimate
—% mineralization
Global time-only reference
—
Not for ranking, certification, regulatory decisions, safety claims, or replacement of laboratory testing.
Trajectory mode
Inspect time dependence
The non-decreasing option is a display constraint for a cumulative endpoint; it does not alter source observations.
Model comparison
Mineralization over time
RF estimate
Time-only reference
Represented-profile interval
Cross-study stress interval
Read before use
What the interface can—and cannot—tell you
Why the cross-study interval is wide
The 69.32-point radius is derived from the maximum absolute error over all curves and times in each held-out study. It is an intentionally conservative stress diagnostic, not prospective conformal coverage.
Why a simple kinetic line is shown
The global reference, B(t)=64.63(1-exp(-0.03504t)), achieved a new-study MAE of 24.87 percentage points versus 26.06 for the exact F-pre RF. The RF therefore has no demonstrated transport advantage.
Why two curated fields are absent
The public model uses only the F-pre set expected to be available before testing. Source-curated degradation-mechanism and PHA-degrading-microbe labels were excluded because their status is not reliably known at prediction time.
What “inside the range” means
Marginal checks only detect obvious numeric extrapolation and unseen categories. They cannot establish that the full formulation–environment–protocol combination is represented.
Privacy and exactness
Preprocessing and all 300 trees execute locally in your browser. Inputs are not sent to an inference API. Cross-language release tests compare browser predictions with the frozen scikit-learn pipeline to 1e-10.