Charting how AI holds its course.

The constellations

Three fields of sky. One chart.

Every field is named for the way it moves. We chart each one by the same rule: fix a star you trust, then measure everything else from it.

• ai · the first field

  Artificial intelligence

  Systems that plan and act. We chart their objectives, their limits, and the first sign that they are drifting off course.

• ml · the slow field

  Machine learning

  Models that learn slowly from data. We check each star against the sky it was measured in, before we trust its position.

• dl · the deep field

  Deep learning

  Networks with many layers and few landmarks. We check whether they still hold their course once they leave the sky they were trained under.

Navigation

Six bearings, taken in order.

A navigator never trusts one fix. We take six bearings on every model, and only the ones that agree make it onto the chart.

1. 01

   Fix a position

   Before training, write down where the model should be, and how far it may drift before we turn back.

2. 02

   Read the drift

   Training curves and gradients show the direction the model is moving, not only how fast.

3. 03

   Triangulate

   Calibration, robustness and behaviour under shift, measured against each other rather than one headline number.

4. 04

   Sail through weather

   Shifted conditions, rare cases and hostile inputs. A clear night is not the same as a sound course.

5. 05

   Log the coordinates

   Seeds, data versions and settings recorded, so anyone can take the same bearings again.

6. 06

   Re-check the heading

   Reliability drifts after launch. Monitoring and retraining keep the model on the chart.

Observation logs

Every paper is an observation.

Each entry records what we saw, how we measured it, the code and the data, and the readings that did not fit. The first entry is still being observed.