Training worlds for agents

# Training worlds for agents

The same insertion, with a policy in the seat. A kept world an agent is dropped into that it cannot have
memorised, a task, a perspective, an interval — and a log that says what it knew, what it did, what it
cost, and what changed.

```text
pinned Domain@revision × task × seat × interval × policy@version
  → attributable trajectories
  → a scoped profile: anticipation, discovery, calibration, recovery, cooperation, persistence
  → failures as replayable coordinates, not a scalar
```

## Why a kept world is the right test

- **History resists memorisation.** A world that has run forward holds state no corpus contains; the test
  is the world's own past.
- **The same world serves people and agents.** Nothing is built twice: the training world a person plays
  is the one an agent is run through, with the reading swapped for the log.
- **Families compound.** Versioned origins of one world become a fixed instrument a policy can be compared
  against over time.

## Two uses

- **Testing** — one policy, unseen circumstances, how it behaved.
- **Training** — cohorts through parallel origins under varied policies, tools, prompts, or rules; the
  result is trajectories and comparisons, never a Meridians claim that one is better.

## Gates

| Claim | Evidence |
| --- | --- |
| A run is worth reading. | Someone runs their own agent through a pinned world and changes what they let it do. |
| The measure is real. | Repeatable across seeds and versions, discriminative between systems, useful to a team improving a model. |
| The world resists contamination. | A held-out continuation is not predicted better by a system that saw the earlier history, beyond what that history legitimately teaches. |
| It transfers. | Independent external evaluation; nothing above depends on it. |

## Boundaries

Systems, not people. Declared synthetic world only — performance is evidence about that world until
transfer is shown ([WORLD_ENGINE](../../architecture/WORLD_ENGINE.md#training-environments-preserve-the-learning-conditions)).
No omniscient context and no presentation-owned truth for the policy; canonical state stays canonical. The
evaluation design lives in [CUMULATIVE_VISION §21](../../../specs/world/CUMULATIVE_VISION.md#model-evaluation-product-hypothesis);
the human twin is [12 — The practice record](12-the-practice-monopoly.md), whose proof programme borrows
this leaf's measurement infrastructure — pinned worlds, anchored items, fit — for the independent transfer
study. This leaf does not gate the consumer plan.

Raw Markdown source