▶ how a self is chosen
~ method ~
One cycle lasts 96 seconds and always ends in a collapse. No cycle may end without an execution and a memory write.
| # | phase | what happens |
|---|---|---|
| 1 | OBSERVING | twelve signals read, each with a reliability weight. |
| 2 | FORECASTING | near-future estimates produced per horizon, with bands. |
| 3 | BRANCHING | 32 candidate selves spawned, each with a different disposition. |
| 4 | SIMULATING | each branch plays its action forward against the forecast. |
| 5 | ARGUING | branches state their assumption and defend their action. |
| 6 | SEMIFINAL | the field is cut to twelve on goal progress and risk bounds. |
| 7 | FINALIST | four remain; only one may be executed. |
| 8 | COLLAPSING | one branch is selected; the rest are killed and archived. |
| 9 | EXECUTING | the selected action becomes the agent's real behaviour. |
| 10 | REMEMBERING | a permanent note is written, then the loop restarts. |
scoring terms
| goal progress | 0.30 | does this move the objective, not just the metric |
| optionality | 0.18 | how many futures remain reachable afterwards |
| information gain | 0.16 | what the agent learns even if it is wrong |
| resource efficiency | 0.12 | cost of the action against its expected value |
| forecast confidence | 0.14 | how much the underlying estimate can bear weight |
| risk | −0.20 | penalty; irreversible actions are penalised hardest |
constraints
- — a branch cannot be selected if its risk exceeds the bounded constraint, whatever its score.
- — an irreversible action requires forecast confidence above 0.5.
- — a killed branch is never deleted; its argument stays readable forever.
- — regret is computed against the strongest killed branch and published unedited.
- — nothing simulated is presented as measured. demo signals are labelled demo.