What does not work yet, told before it works.

The problem: The simulator produces coordinates and captures in every session. Until now, comparing a rider’s session with the one they did a year ago meant opening two reports side by side.

The work: In-house software to read, normalise and compare those outputs inside the athlete history, with the interpretation logic that is applied by hand today written out as explicit rules.

Where it stops: It is an internal support layer. It decides no adjustments, replaces no session, and has not yet been used to produce a single recommendation given to a rider.

The problem: Session capture currently runs on Azure Kinect, and Microsoft has discontinued it. A measurement meant to be repeatable five years from now cannot depend on hardware that is no longer made.

The work: Tests with Orbbec depth cameras, the declared successor to Azure Kinect and compatible with its body tracking, checking their output against the current system before anything moves. Horizon 2027-2028.

Where it stops: It is in comparison testing. Until that comparison closes, session capture continues on the system already validated.

The problem: Postural and joint-range assessment consumes a large share of the session, and part of that measurement depends on the hand taking it.

The work: Evaluating the purchase of a VALD HumanTrak system to standardise postural and movement tests, document them and compare them between sessions under the same protocol.

Where it stops: It is not in the laboratory yet. Strength, ranges and asymmetries are measured today with the VALD DynaMo Plus dynamometer, which is operational.

The problem: With enough cases, regularities appear that the eye does not catch: combinations of restriction, asymmetry and geometry that tend to end in the same complaint.

The work: Models trained on in-house data to classify and flag patterns worth a second look, together with language models for summarising long histories.

Where it stops: No model output reaches a report without someone reviewing it. A correlation found in in-house data is not a clinical claim and is not published as one.

The problem: A best-power curve mixes what you can hold with what you simply never tried. Read carelessly, a duration without recent efforts looks like a weakness.

The work: A model that estimates capacity from efforts in normal training, with probability distributions conditional on its assumptions, relative gains between power meters, and abstention where a duration has no support in the data. Versions, assumptions and synthetic tests are documented.

Where it stops: Calibration of its uncertainty on real riders is still pending, and the first real fits show misfits the report itself flags. No validated capacity at eight hours is claimed, it does not replace an assessment when informative efforts are missing, and it is not used to prescribe.

The problem: Course simulation already projects power, pace and time, but treats rider drag as a reference value rather than something measured case by case.

The work: Work on drag coefficient, rolling resistance and frontal area so that a race projection starts from your actual position instead of an average.

Where it stops: A line of exploration. Projections delivered today state the assumptions they were built on.