SCENT COMMONS / V121 / ARCHITECTURE & MACHINE LEARNING

Compare the saved ML predictions.

Seven recorded examples show what ML predicted and what the original simulation found. Change the setting, room and readout to inspect the comparison. This page reads saved results; live ML inference and new native checks run only in the local research instrument.

90 simulated settings→ML predicts places→5 new proposals checked→ML learns again→40 fresh tests

Choose a recorded comparison

Setting — / 625

Five prospective ML proposals, the reference and one adverse development example. These seven examples are not the separate 40 fresh validation settings. The original study used 625 possible settings; this viewer offers seven saved examples.

Five settings chosen by ML before their simulation:

Curved room

Retained curved-room model section locating N1

N1 is an alternative use of an existing seat. The retained section is a visibility view; amber lines are sightlines, not air paths.

● Lower exposure● Higher exposure◯ N1

Seat map: choose a seat to inspect its predicted and checked values.

Exposure, place by place

— ML prediction— Native checkShading: band with a fixed minimum width
N1 · worst case, both rooms—
Largest protected change—

—

Loading saved results…

—

What the learning has actually shown

Evaluation in progress

The two model versions are frozen before 40 previously unsimulated settings are revealed. The separate Codex evaluation is complete; its aggregate results load below.

A fast preview is a different claim from a better optimiser. V120's search tie is preserved. This instrument has a fixed geometry and a bounded menu; its results do not establish physical smell comfort or fan-energy savings. No new air setting has been adopted.

Data come from original virtual tracer simulations. N1, C1 and F1b are seat / journey locations, not participants. Both rooms retain 72 seats.

Minimum band width applies to 99.74% of fresh outputs. Coverage does not establish calibration of GP uncertainty alone. Read the limits and costs. Recorded metrics and model hashes.