A leveling agent that gets cheaper as it runs

# Jev

A guide-directed leveling agent for a private TBC 2.4.3 server, built so that **it gets
cheaper to run the longer it runs**.

```
System 1   33-50 ms   scripted     free       move, face, swing, loot, eat
Tracker    250 ms     predicates   free       step done? off route? timed out?
Coach      0.2-2 Hz   policy       free       arm a skill, move the playhead
Teacher    queued     Claude/GLM   expensive  resolve ambiguity, write skills
Learner    nightly    training     free       grade, retrain, promote
```

The teacher never presses a key and is never on the hot path. It answers the small
fraction of moments no rule can settle, and everything it answers becomes training data
for the policy that replaces it. The demo sentence is *tokens falling, steps still
advancing*.

## Read these in order

| | |
|---|---|
| **[ARCHITECTURE.md](ARCHITECTURE.md)** | how it is built, and why it differs from the plan |
| **[DECISIONS.md](DECISIONS.md)** | every call made, with its reason. Read before proposing an alternative |
| [docs/PLAN.md](docs/PLAN.md) | the original vision |
| [data/README.md](data/README.md) | what is on disk and where it came from |

## Quickstart

```bash
uv venv --python 3.12 && uv pip install -e ".[dev]"
.venv/bin/python -m pytest          # no game required
.venv/bin/python tools/gen_addon_fields.py
```

The wire format, the state contract, the situation key, the verifier and the reward
function are all testable with no client, no capture and no network. A count is left out
of this file on purpose — it goes stale in a day and a stale number in a README is a
small lie you tell yourself every time you read it. `STATUS.md` carries the current one.

**Nothing drives the game yet.** The brain is built and simulated; capture and input are
not written. See `