> ## Documentation Index
> Fetch the complete documentation index at: https://docs.alakazam.gg/llms.txt
> Use this file to discover all available pages before exploring further.

# The exam

> The frozen Webots oracle: contract, verdict bars, and how to read results honestly.

The exam is the sole scoreboard. It is frozen: same two worlds (arena,
slalom), same controller tick, same proximity remap, same episode count, for
every job ever run, so results are comparable across the whole program
history.

## Contract

Every `job.json` records the contract it was examined under:

```json theme={null}
"contract": {
  "DISC": "1", "CTRL": "350", "RANGE": "1",
  "GATE3_PROX_MAP": "0.0:0.0,0.05:0.05,0.15:0.27,0.25:0.50,0.55:0.57,0.65:0.66,1.0:0.70",
  "EPS": 20
}
```

20 episodes per world, 172 steps per episode (60 s at a 350 ms tick), fixed
spawn slots.

## The four bars (per world)

| Bar       | Threshold                        | Meaning                              |
| --------- | -------------------------------- | ------------------------------------ |
| Contacts  | 0 contact episodes               | Never touch a wall                   |
| Clearance | `min_clear` > 0.02 m             | Keep real margin, not grazing        |
| Coverage  | `frac_ge_5cm` ≥ 0.95             | Hold ≥ 5 cm clearance ≥ 95% of steps |
| Path      | `path_len` ≥ 2.0 m every episode | Actually travel: no pivot-in-place   |

`verdict.pass` requires all four in both worlds. `verdict.why` lists each
failed bar, quoting the worst episode per world.

## Reading verdicts honestly

<Warning>
  No policy in program history has passed all four bars. The best-known
  profile, reproducible via lineage jobs, is a contact-safe wanderer: 0
  contacts in 40/40 episodes, but short of the clearance, coverage, and path
  bars. A `pass: false` with clean contacts is state of the art, not failure.
</Warning>

Judge from per-episode rows, not verdict lines: the verdict quotes minima,
and fixed spawn slots collapse any wall-saturated policy into identical worst
episodes, so two very different policies can produce the same `why` lines.

## Anti-exploit machinery

Frozen environments get gamed, so every exam carries control arms:

* Cruiser control: a fixed forward-driving policy runs alongside your
  champion and *must* rack up contacts (`summary.cruiser_contacts` = 20/world).
  If it doesn't, the exam is vacuous and the run is invalid, not a pass.
* Dream fitness (`dream_F`) is reported but never part of the verdict.
  Policies that exploit the dream get exposed by the oracle.

## Determinism

* Training: bit-deterministic for a given seed on a fixed platform. The same
  spec on macOS-arm64 vs Linux-amd64 yields different champions; the
  floating-point paths differ.
* Exam: deterministic to aggregates; per-episode values drift at the third
  decimal across launches (cross-launch physics noise).

## Policy format: bring your own

The exam accepts either:

* a 9-float controller (`genome9`, or the `genome6` it expands from), or
* your own policy as a sandboxed Python module (`policy.format =
  python_module`); see [Certify your own policy](/train/own-policy).

A `python_module` policy implements `reset(seed)` + `act(obs)`, where `obs` is
the same observation dict the [local gym](/train/local-gym) emits and the
action is the wheel-fraction vocabulary, so a policy trained locally drops
into the exam unchanged. The policy replaces only the champion arm; everything
else here stays frozen, and every exam-stage change passes a locked take-7
acceptance replay before serving.