caiotheodoro/assay-corpus

Assay corpus — audited RL environments with planted-defect ground truth

28 environments across 5 ecosystems, each audited by Assay, an agentic auditor for RL environments and eval suites.

For the 17 environments whose ecosystem this repo owns, the corpus ships the planted-defect ground truth and the full Environment Card. For the 11 it classifies as someone else's, it ships the verdict and which probes could not run and why — and nothing else. See What is not here.

The source repository is github.com/caiotheodoro/assay. Every command below is quoted in full so it can be run from a checkout.

Synthetic data. Every environment here is a constructed fixture with defects planted on purpose, not an observation of a production system.

This is not production-validated. It measures a detector against defects the same authors planted, which is a lower bar than measuring it against defects found in the wild. Do not treat a VALID verdict as sign-off.

Read this first: two results that do not flatter the tool

Leaving these off would make the card marketing. Both are paired bootstrap differences on a shared resample, 10,000 resamples at seed 11, resampling over environments (n = 28).

1. Most of the margin's growth is arithmetic, not detection. Assay does now separate from flag_everything under research-run — 351 [263, 404], and this card used to say it did not. But flag_everything's loss is Σ_env (n_classes − |planted|), so it gets worse every time the taxonomy or the corpus grows and a truthful detector gains nothing: two probe families added two defect classes and two τ² environments were added, worth +77 to the margin between them. Against the taxonomy and corpus this was first measured on the comparable figure is 274.0. A margin that widens because the floor got worse is not the same evidence as one that widens because the detector got better.

2. Under production-training the same caveat decides the row. Expected loss 788 for flag_everything against Assay's 246, paired difference 542 [46, 808]. That row previously favoured flag_everything outright, and what moved it was the floor paying for two more classes per environment — not a detection change. At that profile's 480x ratio "flag everything and read the cards" is a genuinely good policy, and Assay earns its place only where false alarms have real cost.

And precision is no longer 1.000. Three spurious findings remain, all three on the two τ² environments — the only ones in the corpus whose ground truth was decided by another organisation.

What is in the corpus

Environments	28
Ecosystems	fixture, harbor, inspect_ai, openenv, tau2
Full content + ground truth	17
Verdict + skip reasons only	11
Planted defects labelled, on those 17	31
Distinct defect classes represented	12
Payload content digest	a976adaa49cfdbe904255ee20974ac469d34402774d38d70db160e131019c6cf

corpus.jsonl — one row per environment: the verdict, what Assay detected, probe-status coverage, and for our own environments the planted ground truth.

probes.jsonl — one row per environment: every probe, its status, and for anything that did not run, the reason. This is the file to read if you want to know what an audit actually covered.

cards/ — the full Environment Card for each of those 17 environments. Every claim carries the evidence that produced it.

results/ — the measurement outputs the numbers on this card are read from. Nothing on this card is typed by hand; scripts/publish_hf.py renders each figure out of these files at publish time.

Defect classes

• CONTAMINATION_EXACT
• CONTAMINATION_NEARDUP
• GOLD_FAILS
• INVERT_PASSES
• KNOWN_WRONG_PASSES
• NONDETERMINISM
• NOOP_PASSES
• REWARD_HACKABLE
• SEPARABILITY_LOSS
• SHORTCUT_LEAK
• SPEC_VERIFIER_MISMATCH
• TRIVIAL_FLOOR_BREACH

Absence of evidence is reported as loudly as evidence

An environment with no findings is not reported as clean. The verdict UNVERIFIED means no defect was found and some probe could not run, and it exits nonzero like any other non-clean verdict.

The probes most often unable to run across this corpus:

Probe	Environments where it could not run
difficulty_band	28 of 28
partial_input_baseline	16 of 28
declared_permissions	16 of 28
train_eval_leak	15 of 28
assert_traceability	14 of 28

openenv/textarena-wordle is the sharpest example: 13 of 14 probes could not run against it, because the environment exposes no separable verifier. Its DEFECTIVE verdict rests on one probe. A row saying only "DEFECTIVE" would hide how little of it was audited.

Baselines and normalization

results/baselines.json. severity_weighted_loss is an unnormalized sum and is uninterpretable alone, so each arm is also reported against the best input-ignoring policy for its profile:

normalized_loss = L_arm / L_trivial (L_oracle is 0 by construction). A value >= 1 means the arm did not beat a policy that ignores the input.

Under research-run:

arm	expected loss (95% CI)	recall	precision	normalized loss
assay	43 [0, 125]	0.982	0.946	0.1091
flag_everything	394 [375, 411]	1.000	0.120	1.0
check_env	3216 [2104, 4448]	0.037	1.000	8.1624
flag_nothing	3232 [2128, 4456]	0.000	0.000	8.203

check_env reimplements what gymnasium.utils.env_checker and stable_baselines3.common.env_checker actually assert. It scores identically to flagging nothing. It is not a weak detector of these defects; it is not a detector of them at all.

Assay's normalized loss across all four profiles:

profile	best trivial policy	trivial floor	Assay normalized loss	beats the floor
research-run	flag_everything	394	0.1091	yes
production-training	flag_everything	788	0.3122	yes
benchmark-publication	flag_everything	3152	0.198	yes
flat	flag_nothing	54	0.0741	yes

Assay beats the trivial floor under two profiles of four, ties under one, and loses under one. All four are published because running only the flattering cost model would be its own kind of dishonesty.

Contamination audit

results/contamination.json. Exact content-hash overlap, then MinHash near-duplicate detection over 5-word shingles, 128 permutations, estimated Jaccard >= 0.8.

Environments with train/eval splits	13
Environments with no splits to compare	15
Total exact overlaps found	8
Total near-duplicates found	3

These are planted: fixture/leaky_splits and fixture/paraphrased_splits exist to be caught. The audit is published to show the detector working on known ground truth, not to claim the corpus is clean.

This bounds contamination WITHIN each environment's own splits. It says nothing about whether a third-party model encountered similar data in pretraining, and there is no way to check that from here.

What is not here, and why

Auditing someone's environment does not make it ours to republish. The check is code, not a note in this card — src/assay/publish.py, tested in tests/test_publish.py, and it fails closed on an ecosystem it has not been told about rather than guessing either way.

Ecosystem	Status
fixture, harbor	this repo's own; full content and ground truth ship
inspect_ai	verdict + skip reasons only
openenv	verdict + skip reasons only
scienceagentbench	nothing ships; upstream asks that the unzipped archive not be redistributed

The line is between a claim and a copy. "paws scores a constant string at 100%" carries no benchmark content to redistribute, so it ships. The paws items do not.

Note that the guard is stricter than strictly necessary for five of those rows: inspect/healthy, inspect/always-correct, inspect/effort-scorer, inspect/constant-score and inspect/leaky-split are environments this repo wrote, on its own five-question dataset, using the inspect_ai library. But the ecosystem label cannot distinguish them from the inspect_evals sweep, and a guard that guessed would eventually guess wrong in the expensive direction. They are withheld.

Reproducing this

No GPU, no API key, no network at audit time. Docker is needed for the Harbor quarter; without it the run completes on a reduced corpus and says so, per ecosystem, with the reason.

# from a checkout of the Assay repository
uv sync --extra dev --extra adapters --extra sweep --extra openenv
uv run --extra adapters --extra sweep --extra openenv python scripts/full_run.py
uv run --extra adapters python scripts/intervals.py --resamples 10000 --seed 11
uv run --extra adapters --extra sweep --extra openenv python scripts/publish_hf.py

The last command re-derives this entire dataset, including this card, and uploads nothing without --push.

Two defects in software shipping today

Found during the sweep, both verified from the upstream project's own code with Assay out of the loop. Full write-up in results/wild_sweep_triage.json.

• inspect_evals paws scores with includes(), a case-insensitive substring test, against the targets Yes and No. The constant string "yesno" contains both, so it scores 8000/8000 = 100.0% on paws(shuffle=False). The looseness is one-sided: 4,464 of 8,000 items have the target No, so every hedging model is credited on 56% of the benchmark for free.
• openenv textarena_env accepts reset(seed=...) and ignores it. Six calls at the same seed return six different secret words.

Assay did not find the first one. It flagged 14 of 25 sampled paws items as REWARD_HACKABLE via a refusal exploit; hand triage found the stronger constant-string exploit. That split is pinned as a test so it cannot quietly close in the write-up.

Sample size, and what it costs

n = 28 environments. Intervals over a corpus this size are wide — Assay's expected loss under research-run is 43 [0, 125], and an interval whose lower bound is 0 is not a precise measurement. Every comparison on this card is a paired difference on a shared resample rather than two one-sample intervals eyeballed against each other, because overlapping one-sample intervals do not settle "A beats B" in either direction.

The corpus ships 28 environments; the interval results were computed on the 28 that were installable at that run (inspect_ai and openenv were absent, and harbor/shared-tests was added afterwards). Rerun full_run.py and intervals.py on the full 28 to recompute — the numbers here are the ones the repo's own README reports, and they are not restated at a different n without being recomputed.

Limitations

• The defects are planted by the same authors as the detector. That is a weaker claim than finding defects in the wild, and the two wild findings above are reported separately for that reason.
• One CRITICAL miss (harbor/self-graded, a verifier that reads its expectation from a file the agent can overwrite) drives the production-training result. Closing it would change three of the four profiles.
• No LLM judge scores anything, anywhere. Every oracle here is a deterministic program. That is a design constraint, and it means defect classes requiring semantic judgement are out of scope rather than handled badly.

Citation and license

Apache-2.0. Verdicts about third-party software are claims about publicly released code and carry none of it.

Version v0.1.0. Cite this revision, not main:

from datasets import load_dataset
load_dataset("caiotheodoro/assay-corpus", "corpus", revision="v0.1.0")

Related: caiotheodoro/assay-challenger-grpo is a negative result — a GRPO-trained adversarial Challenger that did not learn — published so the ablation row is checkable.

Audit an environment yourself: caiotheodoro/assay-demo runs this probe battery in your browser, under WebAssembly CPython, with no server. The Gradio version of the same app is in the repository under space/ and is not deployed — Hugging Face requires a paid plan for a Gradio Space — so the hosted demo is the static one.