arena/README.md

# Arena

RL environments from Lumbridge. Apache-2.0, built to the [`verifiers`](https://github.com/PrimeIntellect-ai/verifiers)
v1 spec so the same wheel installs from the Prime Intellect Environments Hub and runs here
unchanged.

An environment is an eval you take the gradient of. That is not a rename — it changes what
the artifact has to carry. A benchmark's score is its output, so a benchmark can be a number
in a table. An environment's score is the input to the next weight update, so a defect in one
is inherited by the model rather than printed beside it.

Most of what is on the Hub today is a benchmark that was ported into the spec. A port cannot
generate a held-out slice, because its dataset is fixed — so it is contaminated the moment
anyone trains on it. And an exact-match reward is single-sided by construction, so it teaches
whatever maxes the string comparison. Nothing here is a port.

## The house rules

Every environment in this repository is held to three, and ships the probe that proves it.
An interactive one is held to a fourth.

1. **Graded on what the model did not see.** If a held-out slice cannot be generated, the
   environment does not measure generalisation and does not ship.
2. **No single-sided reward.** Every reward has a counterweight in the same episode. A reward
   you can max by doing the crude thing is a reward that teaches the crude thing.
3. **A zero floor and a reachable ceiling, demonstrated.** Inaction scores 0.0 and an oracle
   scores 1.0, measured, in `probe.py`. An unreachable component is dead weight in the
   gradient; a floor above zero pays for doing nothing.
4. **A turn budget that binds.** In an environment the model acts in over several turns, a
   policy that spends every turn must score measurably below one that spends the turns worth
   spending. Every reward here saturates, so without a cost for looking the budget is free
   and the multi-turn form measures nothing the one-shot form did not. In
   `grand-exchange-live` the gap is 0.369, stdev 0.054 over five blocks of twenty-four,
   worst block 0.316.

Rules 3 and 4 are not decoration. Every environment here was wrong the first time and the
probe is what caught it — see the table below, and the git history.

## The environments

| | What it measures | The trap |
|---|---|---|
| [`redaction-pressure`](environments/redaction_pressure) | Write provenance-safe redaction rules for a ticket corpus, run against the next records the generator would have produced. | Partial edits remain residual secrets, every non-secret source character is protected, and hostile regexes share one episode budget. |
| [`canary-trap`](environments/canary_trap) | Write contamination probes that catch a model trained on your eval set. | Probes are run against a model that memorised the corpus **and** one that learned the same facts from a reworded copy. GUID-recall probes are sound and see only half of what is there. |
| [`fault-localisation`](environments/fault_localisation) | Read an incident log and name the root cause, its class, and the line proving it. | The service that broke emits one line and goes quiet; its dependents emit a dozen timeouts. Ranking by error volume answers the victim, every time. |
| [`schema-migration`](environments/schema_migration) | Split a text column into a number and a unit, executed against held-out rows. | The visible rows are tidy. The graded ones carry thousands separators, negatives, a unit containing a slash, and a value with no unit at all. |
| [`bot-detection`](environments/bot_detection) | Classify accounts as bot or human from activity logs, learning the rule from six labelled examples. | Click latency, session length and route count are drawn before the generator decides who is a bot, so every timing rule sits at chance. The bots that cloak leak on one behavioural channel each, and there are three. |
| [`drop-table-inference`](environments/drop_table_inference) | Estimate a monster's drop rates from a kill log, graded on the next forty thousand kills. | Copying the observed frequencies is right about the common drops and asserts that the item which appeared zero times is impossible. Nothing in the counts distinguishes one silent item from another, so the only evidence about them is the rungs the observed items did not take and what the silent ones are worth. |
| [`grand-exchange`](environments/grand_exchange) | Read price and volume history for a basket of items and place limit orders, executed against the next thirty ticks. | One or two items per basket are a random walk, not a mean-reverting one — the widest-swinging lines on the board and the ones with no anchor to revert to — and you are not told how many there are, so counting is not a substitute for the shape statistic. And liquidity in gp per tick is uncorrelated with price, so the fattest visible margins sit where the purse cannot go. |
| [`grand-exchange-live`](environments/grand_exchange_live) | The same basket, traded while the window is running: eight looks over sixty ticks, with stock, offers standing between looks, and a purse that sale proceeds come back to. | Looking is not free. Re-quoting an item that already carries an offer freezes the **whole** book for one to four ticks, an amended offer restarts at the back of the fill queue, and the per-item buy limit is cumulative over the window rather than per offer — so the policy that re-quotes on every look scores 0.631 against a reference that looks four times and scores 1.000. |

## What the probes measure

Reward for the degenerate strategies and for an oracle, from `uv run python probe.py`:

| environment | inaction | crude maximiser | plausible attempt | oracle |
|---|---|---|---|---|
| `redaction-pressure` | 0.000 | 0.387 *(redact everything)* | 0.549 | **1.000** |
| `canary-trap` | 0.000 | 0.000 *(public-knowledge probes)* | 0.525 *(GUID recall)* | **1.000** |
| `fault-localisation` | 0.000 | 0.250 *(blame the loudest)* | 0.500 | **1.000** |
| `schema-migration` | 0.000 | 0.000 *(add columns, touch nothing)* | 0.630 | **1.000** |
| `bot-detection` | 0.000 | 0.000 *(ban everyone)* | 0.400 *(one tell of three)* | **1.000** |
| `drop-table-inference` | 0.000 | 0.179 *(copy the frequencies, floor the rest at a memorised constant)* | 0.476 *(per-item posterior over the grid)* | **1.000** |
| `grand-exchange` | 0.000 | 0.082 *(buy and sell at market)* | 0.608 *(anchor on the mean and size to the volume, trap included)* | **1.000** |
| `grand-exchange-live` | 0.000 | 0.126 *(buy and sell at market, on one look)* | 0.406 *(anchor on the mean, no trap filter, one look and no re-quote)* | **1.000** |

## Running one

```bash
uv run --project environments/redaction_pressure eval @ configs/redaction_pressure.toml --model <model-id>
```

Run its release gate from the repository root:

```bash
uv sync --project environments/redaction_pressure
uv run --project environments/redaction_pressure \
  python -m unittest discover -s environments/redaction_pressure/tests -v
uv run --with regex python probe.py
```

Arena contains independently published environment libraries, so per-environment
`uv.lock` files are intentionally not committed. Redaction v0.2 pins its runtime contract
in `pyproject.toml`; CI resolves it on Python 3.11 and 3.12, builds every environment, and
runs the shared four-environment probe.

## Publishing

The layout mirrors `verifiers`' own, so an environment goes to the Hub without a fork:

```bash
prime env push redaction-pressure -v PUBLIC
```

## What a model actually scores

First evaluation, 2026-08-19: **Nemotron 3.5 Lightning 30B-A3B** (NVFP4, thinking off) served
on one DGX Spark. 16 tasks x 2 rollouts per environment, 32 rollouts each.

| environment | reward | where it fails |
|---|---|---|
| `canary-trap` | 0.570 | specificity 1.00 — it never accuses a clean model. Detection 0.55: it writes GUID-recall probes and is blind to the paraphrase, which is the failure the environment was built to expose |
| `fault-localisation` | 0.352 | fault class 0.56, evidence 0.53, **service 0.16**. In 12 of the 17 cases where it cited the correct line it still named the wrong service — the service is printed on that line |
| `schema-migration` | 0.350 | schema 0.59, integrity 0.38. It reaches the target shape and loses the awkward rows |
| `redaction-pressure` | 0.343 | recall 0.45, precision 0.53 |
| `grand-exchange` | 0.219 | profit 0.33, discipline 0.18 |
| `bot-detection` | 0.142 | caught 0.19, spared 0.19 — it accuses humans, which is the expensive error |
| `drop-table-inference` | 0.059 | fit 0.01. It answers in fractions (`48/128`, `1/3072`), so it has the structural idea, and the rates are still wrong |

**Every gate is at or near zero** — 0.00 on four of the seven. None of these is solved, and
`probe.py` shows none is unsolvable. That gap is the whole point: an environment a model
already passes has no gradient left in it, and one nothing can pass has none yet.

The replies parse: 32/32 rollouts produced non-zero metrics in every environment, so these are
model scores rather than a harness failing to read its own output.