LIVE INFERENCE / LOCAL TRAINING
Change the conditions.
Question the prediction.
Verifying model…
02 / COMPARE THE SAME INPUT
| C* | 01 / ENCOUNTER | 02 / BUFFER | 03 / WITHDRAW |
|---|---|---|---|
| Original grid | |||
| Scaled interpolation | |||
| Residual ML / saved candidate | |||
| Saved direct solver | |||
| Interpolation absolute error | |||
| Residual ML absolute error |
C* is a reference-normalised tracer quantity, not measured odour or comfort. A reference appears only at an exact saved input. Sampled maxima can miss between-sample peaks.
03 / FIT A RESIDUAL MODEL HERE
Fit the same data.
Compare a stronger baseline.
Refit the selected context’s correction from zero kernel coefficients around its training mean: 967 fixed training cases, checked against 128 separate gate-validation cases. Residual labels come from five folds of training-only interpolation. The correction weight stays at 0.25; other contexts keep their saved weights.
Saved kernels loaded · scaled interpolation selected before the final test.
Training solves a regularised kernel system using training targets only. The 2,048 final-test references are never fed to the optimiser. No new data, persistent learning or sensor connection.
FRESH FINAL TEST / 2,048 CASES
Lower overall error.
Room 03 is still unresolved.
Against interpolation with the same 3,868 training labels and the same axis-scale search, residual ML reduces average log error by 13.9% in general conditions and 11.6% on P10. Both paired intervals are below zero. Room 03 mean MAE falls only 1.0% and 0.3%; both intervals include zero. The general worst-five-operating-point guard also fails. Scaled interpolation remains the default.
See the complete comparison and remaining limits ↗Domain + reproducibility
Four contexts: two temperature profiles × two door events. Wind −6…6 Pa; supply 4,320…5,040 m³/h; lower general-extract share 25…75%; interzone bypass 10…90 m³/h. The lower share applies to general extraction, not the 360 m³/h local extract. This remains an in-domain, uncalibrated simulator benchmark.
One continuous lower Room 01 source; each 20-minute run starts at closed-door steady state. The film uses a separate ten-minute release and ten-minute clearing protocol. Nine targets: three room means, three lower-layer means and three door-change-sampled maxima. No measured scent intensity, comfort, plume velocities or sensor feed.
Both arms use 3,868 labels: all 3,356 previously opened v2/v3 cases and 512 new fit cases. Models are chosen on 512 new selection cases. A separate 512-case gate can only accept or reject that choice. Another 2,048 final cases are generated after model and evaluation-code freeze. All four contexts of a base operating point remain together. New points are at least 0.06 cell widths from earlier points; this deliberate spacing differs from v3’s 0.10 and is not extrapolation validation.
The baseline is exact cubic RBF interpolation of log₁₀(C* + 10⁻⁶), with axis lengths [0.5,4,4,2] and an absolute-wind feature. A Matérn 5/2 kernel learns five-fold out-of-fold residuals using lengths [2,4,4,2] and regularisation 10⁻⁶. Its signed correction is multiplied by 0.25 before being added to interpolation; only the final concentration is clipped at zero. Both arms had access to the same four axis-scale choices.
Browser fitting factorises the 967-case residual kernel matrix incrementally. Only training residual labels update coefficients; the 128 gate cases per context monitor error without early stopping or selection. Final-test references are never optimizer inputs. Partial coefficients are provisional. Completion requires agreement with the saved model; failure restores saved coefficients. No new data or persistent learning.
Selecting a checkpoint retains its exact stored input and context, even if native sliders round their displayed values. Manually changing a slider, temperature, door event or choosing Custom clears the solver reference. Changing the displayed metric or predictor preserves the checkpoint.