OBSERVE / PREDICT / VALIDATE
A question should
change what comes next.
A familiar place changes, returns, and changes again.
Can Lumen explain the difference without losing its bearings?
Lumen is finding its bearings.DRAG TO ORBIT · SCROLL TO ZOOM
Ensemble care is ready. The frozen comparison trains an investigation selector and checks it on separate habitats.
Care before curiosity.
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.
Six authored candidate families plus one possible observed combination. Mismatch is measured against empirical healthy-reading scales, not a probability. Readings older than 24 ticks are excluded from rejection. The normal checkpoint is protected. “Unexplained” means none fits well; care still uses its current selected candidate.
Changes of explanation
The protected checkpoint
Predictions, then measurements.
Proposed changes must predict later readings.
Forecasts use current weather held constant and a feedback controller after each candidate journey. They are model estimates, not promises. An investigation spends real travel and operation ticks. A mismatch alone does not identify a 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. Warm bands mark the two hidden disturbance windows in the observer record. Evidence plots can show more than the learner sees.
TRAIN / FREEZE / EVALUATE
Learn when a question is worth its cost.
Six training habitats supply paired branch outcomes. Twelve separate habitats test the frozen selector against ensemble care, decision-value tests and shuffled training outcomes. Compound faults are excluded from selector training.
Count the cost of getting an answer.
Calibration spends 182 actions on measurements and 160 on separate normal validation. Training and every diagnostic continuation are also charged. The full comparison uses at most 201,996 physical actions, including calibration. Each evaluation receives the same 480-action budget. Forecast transitions are counted separately.
One worker on this Mac. Thirty-minute active limit. Checkpoints survive page changes. Repeating these seeds reproduces the same frozen test; it does not supply new validation.
A place to test a small claim.
The body, equations, and six candidate families are authored. A shared calibration measures normal bed responses and tank refill. Proposed parameter changes are checked against subsequent observations before acceptance. Lumen ranks candidates by prediction error, fits their permitted parameters, and compares physical investigations. Unfamiliar combinations can defeat these explanations. This is a bounded experiment in adaptation; it does not establish general intelligence or sentience.
Research & replays ↗Executed-work findings ↗Reuse & replanning findings ↗Terminal-value findings ↗Two-decision findings ↗Planning trace findings ↗Journey ranking findings ↗Prediction & care findings ↗Code review ↗Bottleneck findings ↗Timed-care findings ↗Why this journey? ↗Research roadmap ↗Reading-value findings ↗Previous findings ↗Protocol & limits ↗Image-steering connection ↗Previous adaptation nursery ↗