LOCAL READINGS / CHANGING CONDITIONS
When the world
changes its mind.
Remember a place. Predict a journey. Notice a mismatch.
Can Lumen recover when familiar conditions change?
Lumen is finding its bearings.DRAG TO ORBIT · SCROLL TO ZOOM
A control sends Lumen to that mechanism. Weather advances on every travel and operation tick. Pausing stops physics; an intervention pauses automatic decisions.
01 / BEFORE THE NEXT STEP
Compare the journeys.
What is being learned?
Forecasts use current weather held constant and a feedback controller after each candidate journey. They are model estimates, not promises. A mismatch requests investigation; it does not identify the cause.
02 / THE COST OF NOT BEING THERE
What stayed in memory?
Fern bed · A
Reed bed · B
Solid: actual moisture in the recorded evidence. Dashed: Lumen’s estimate. Shaded: the bed’s viable range. Evidence plots can show more than the learner sees.
03 / THE SAME TEST FOR EVERY CLAIM
Can adaptation earn its place?
Five independently seeded habitats. A familiar habitat is changed without announcing the change to the learner. Every evaluation starts from the same trained checkpoint for that habitat.
Prediction, recovery, and retention
Six controls · two shared training seasons per habitat · eight evaluation seasons per control · four common action histories. 103,680 environment actions in a complete run. Model forecasts are counted separately.
Inspect disturbances, paired uncertainty, and common prediction tests
Runs continue on this Mac while you switch pages. Five independent habitat seeds are the units of replication. The two full-information controls are diagnostic references, not achievable ceilings. They receive information denied to the learner. Evaluation seeds become exposed after the first run.
Recovery has two obstacles.
In a separate follow-up on twenty exposed cases, both water exhaustion and degraded retained models mattered. With physical state reset and learning disabled, the pre-evaluation models kept both beds viable for 87.3% of a fresh 128-tick trial; post-evaluation models averaged 73.6%. After shared supply shortages, that fell to 53.3%.
A fresh reset changes the physical state and time horizon. This is a diagnostic of retained models, not another independent validation. Three sampled full-length cases had insufficient water even under optimistic assumptions.
A place to test a small claim.
The body and the equations are authored. Lumen learns delay, evaporation, delivery, and supply estimates from its available history, then compares simulated journeys. Earlier models and prediction failures remain inspectable. This is a bounded experiment in adaptation; it does not establish general intelligence or sentience.