Field study 03 / 06 · Learn, Forget, Evolve
Learn a world.
Lose a world.
A policy that worked yesterday may be precisely the wrong habit tomorrow. The old world will come back to check.
学会一个世界，也可能遗忘一个世界。
Loading the model. Recorded evidence remains available below.
Delivered useful work / units · higher is better
| Treatment | At this tick | At mission end |
|---|
01 / Try to break it
Wait for the first world to return.
Follow A → B → C → A. Compare frozen rules, explicit memory, bounded mutation and random search. A quick recovery is interesting; a recovery that still wins after paying for calibration is more interesting. Inspect the rejected candidates too.
Can a small policy change adapt to a new world without erasing the old one?
02 / The mechanism
What the little world actually does.
Agents operate a small, fixed policy language. Adaptation spends field trials on calibration and selection. Mutation and random search receive matched candidate budgets. Candidates face held-back calibration gates. A separate frozen evaluator audits retained policies on disjoint seeds without controlling promotion; the agent cannot rewrite its reward. A lineage records proposals, rejected candidates, promotions and rollback.
This is policy evolution inside a bounded simulator. Foundation-model weights do not change, and the experiment does not establish recursive intelligence. The useful question is smaller and testable: does experience buy better action in a changed or unseen world, after charging for it?
任务按 A→B→C→A 改变，比较固定规则、明确记忆、有界变异与等预算随机搜索。训练和筛选都收费，保留的校准样本决定候选是否晋级，独立种子评估器随后审计，并记录被拒绝的候选与回滚。这是有界策略演化，不是模型权重自我进化。
03 / Recorded evidence
Keep the inconvenient outcomes.
Simple memory won the clear-season comparison: 1,149.5 delivered units versus 931.0 for bounded mutation. Mutation also spent 6,132 training evaluations, while memory made 192 table updates; separate diagnostic evaluations are counted too. The retention criterion failed when environmental cues became ambiguous.
Bounded mutation plus a retention gate may preserve useful behavior, but a simple context memory can be better, and no policy can identify hidden seasons when their observable cues become indistinguishable.
Paired seeded simulations. Every retained result can be regenerated from the versioned source; these are synthetic experiments, not field validation.
Inspect every world and treatment
| World / treatment | Mean units | n |
|---|---|---|
| Four seasons / Fixed habit | 754.85 | 24 |
| Four seasons / Experience memory | 1,149.55 | 24 |
| Four seasons / Mutation + retention gate | 930.97 | 24 |
| Four seasons / Matched random search | 864.77 | 24 |
| Foggy cues / Fixed habit | 754.85 | 24 |
| Foggy cues / Experience memory | 1,015.28 | 24 |
| Foggy cues / Mutation + retention gate | 738.41 | 24 |
| Foggy cues / Matched random search | 723.04 | 24 |
| Indistinguishable / Fixed habit | 754.85 | 24 |
| Indistinguishable / Experience memory | 732.2 | 24 |
| Indistinguishable / Mutation + retention gate | 613.42 | 24 |
| Indistinguishable / Matched random search | 519.97 | 24 |
| Changed home / Fixed habit | 485.43 | 24 |
| Changed home / Experience memory | 996.29 | 24 |
| Changed home / Mutation + retention gate | 768.52 | 24 |
| Changed home / Matched random search | 799.61 | 24 |
One chosen seed is an illustration. The downloadable batch retains every world and treatment; null means unavailable. A confidence interval describes the sampled synthetic worlds, not the real world.
04 / Inspect the boundary
Enough detail to disagree.
Inputs, budgets and observation boundaries
Every measure in the current run
- A finite contextual foraging task and a 12-parameter linear policy; not open-ended self-improvement or AGI.
- Field calibration deliberately grants every arm the same paid, full-feedback resource samples. Real robots would need to collect those samples.
- Policy candidates are scored on recorded calibration trials; this is not free counterfactual access to unobserved actions.
- The final evaluator uses disjoint seeds and new weather samples, but the same declared family of three ecologies.
- The illustrated height is a resource/learning cue, not simulated three-dimensional physics.
3D sketching adds depth to the explanation. It does not add an unmodeled physical dimension, extra sensor data or unearned intelligence to the experiment.