01 / fixed rule
Carry every habit
One learned actuator map follows the task everywhere.
body —
Loading…
—
Try this first
Let the old body return.
Press Show outcome. The portable core remembers a separate map when each of four bodies returns. Now choose Same controls: one shared habit becomes useful, and separating memories wastes time. Choose 12 · one use: no body returns, so portable memory and starting blank become the same rule.
Continuity is not one thing.
Every tick requests north, east, south or west. The controller chooses physical control 1–4; the body's hidden bijection decides what actually happens. A correct command advances the recipe. Four correct commands complete one toy assembly.
The three adaptive controllers share the same feedback update. They differ only in what crosses a handoff. Carry every habit keeps one global actuator map. Start blank deletes it. Portable core retains the task and stores a map under the advertised capability fingerprint. The frozen and random policies do not learn.
Remembering the boundary helped—until it did not.
| Setting | Carry all | Blank | Portable | Frozen | Random |
|---|---|---|---|---|---|
| Default recurring bodies | 72.797 | 72.938 | 87.938 | 2.281 | 24.484 |
| Same mapping | 94.156 | 73.141 | 88.344 | 96 | 23.578 |
| Rapid handoff | 49.922 | 50.984 | 88.094 | 3.047 | 24.156 |
| One-use bodies | 72.875 | 72.406 | 72.406 | 3.813 | 24.078 |
In the default batch, portable-minus-blank averages +15 commands. With a handoff every four ticks, it is +37.109. With one-use bodies, the difference is exactly zero. With identical mappings, carrying the shared habit beats the portable core by 5.813; the frozen manual is perfect because its assumption happens to be true.
A fresh label can cost more—and still be safer.
Turn on firmware drift at tick 48, then compare fresh and stale labels. Stale labels produce 7.547 mean errors from obsolete recalled maps. Yet stale-label throughput is 80.406 correct commands versus 80.125 with fresh labels: repairing an old map happens to cost slightly less than four complete recalibrations in this toy.
That does not make stale capability metadata good. Every wrong command has equal, reversible cost here. A hazardous actuator could make one stale move unacceptable. Versioning and safe probing are separate design choices; a throughput score cannot settle safety.
What the controls mean
- Seed
- A repeatable recipe and set of actuator mappings, not a difficulty rank.
- Handoff
- Action ticks before the next body takes over. All policies get the same action count.
- Bodies
- Four identities recur; twelve bodies appear once each in the default 96-tick run.
- Control difference
- Same controls, one swapped pair, or a full seeded permutation. Every mapping remains one-to-one.
- Firmware drift
- At tick 48 every hidden mapping changes. No controller is told the mapping itself.
- Capability label
- Fresh publishes a new versioned fingerprint. Stale reuses the old label while feedback remains available.
Where the self stops.
“Portable self” means scoped controller state in a deterministic toy. It is not consciousness, personality, biological memory, a robot benchmark or evidence about children, companions, animals or human identity. There is no perception noise, latency, injury, safety envelope, language memory or semantic transfer. Commands are abstract and reversible.
All 640 matched seeded worlds and 3,200 policy outcomes are retained across ten settings, including identical mappings, one-use bodies, rapid/slow handoffs and fresh/stale drift. No seed was removed. The private source scan supplied a question—not instructions, public facts or prose—and is not included here.
Method & parameter guide ↓Source package ↓All batch results ↓Read the model
Previously: The Receipt Gap · Explore the eight small worlds.