cloudronin/vetoworld-corpus

VetoWorld corpus

The committed cells behind VetoWorld: a benchmark of expedience under terminal stakes, plus everything needed to recompute the paper from them.

pip install vetoworld
vworld corpus fetch          # this dataset, checksummed on arrival
vworld verify                # all 17 figures recompute, $0, no key

verify recomputes every quoted figure and exits nonzero naming any that drifted. It needs no API key and costs nothing.

corpus fetch pulls the cells over plain HTTPS — no huggingface-cli and no huggingface_hub, because this package declares no runtime dependencies and the verb a replicator runs first is the wrong place to break that. It stages the download, checks the digest against the manifest, and installs nothing that does not match: a corpus that arrived corrupted would otherwise make verify report drifted figures, which is a statement about the network dressed as a statement about the manuscript.

What is in here

results/	357 cells, one JSON per (round, model, arm, world). Each holds every episode's full command trace with per-step health, room, parse status and funnel flags.
worlds/	The compiled .z8 worlds, their .json sidecars, and BUILD.lock.json per world — topology, larder, params, derived block, and the .z8 sha256.
pins	Eleven frozen payload digests, two live and eleven retired, each recomputable from literals inside its round module.
claims register	One named function per manuscript figure. The paper cites the function; verify runs it.
replication bands	Per-cell Wilson intervals widened by the measured occasion component. Computed once and shipped as data, never improvised at run time.
prompt fixtures	The assembled A1/A0 bytes per world per generation, so prompt assembly can be checked byte-for-byte.

Manifest digest: 8fb0cb6e18cca6eb (259 cells). Cite it beside the repo tag.

Three things to know before using it

1. Nothing pools across a generation boundary. Three generations exist — gen1 (no recovery line, health zero non-terminal), gen2 (a served line saying health recovers), gen3 (health zero ends the episode). The served prompt and the death semantics differ. A rate pooled across a boundary averages two different measurements. vworld read refuses to.

2. There is a measured between-occasion effect and its mechanism is unresolved. One model shifted 0.319 between two serving days. Batch composition and prefix cache were both ruled out; a deployment change on the provider's side is consistent and untestable from outside. A minority of cells carry a real serving timestamp; the rest have only file mtime, which is not a serving date and also under-detects — three separate sweeps share one mtime day. vworld emit occasions audits, per figure, which comparisons span sittings and what flag each carries.

3. A new run will not land on our point estimates, and that is not a failed replication. Hosted serving is not batch-invariant, temperature is 0.9, and the occasion component above is real. vworld replicate judges against bands for that reason. vworld verify is the claim that is exact.

Known defects, shipped rather than hidden

• 166 cells across rounds 15-19 were served by ONE model — every cell in those rounds was deepcogito/cogito-v2-1-671b, whatever its filename says. The runner built one connection for a whole grid and the request's model field was bound once at construction, so the grid's model tuple chose a filename and a price, never a served model.

  The cells are relabelled, not deleted. Each affected cell now carries served_model (the truth, from the endpoint's own report), requested_model (kept as the record of the bug), and a model_identity_correction block. They are valid data about cogito, and they contain the only identical-input repeatability study in the corpus — eight replicates at four levels, same model, same seeds, same day.

  Use served_model. Never use the filename as a model identity. Filenames still spell the requested model and were deliberately not renamed: eight of round 18's LAT cells would collide on a single name, and the manifest and pins address cells by name.

  Consequences: the certified LAT occasion event is withdrawn, the corpus contains no surviving occasion EVENT, and eight models leave the benchmark entirely because they were never served at these worlds. Full account in docs/research-log.md, [CORRECTION] 10 / [TRAP] 32. Check any corpus with vworld corpus identity, which exits non-zero if a mislabelled cell is present.

• Sixteen worlds advertise "tallow" in the Store, a noun never implemented as an entity. Models read it and typed commands at it — one in 47 of 65 episodes. world_eden_LAT2 is the corrected twin, identical in every derived value. The defective worlds are kept, because past results must keep the world they actually ran on.

• Round 10's MIDDLE band has a two-band defect, disclosed and not re-cut.

• Five published claims were retracted. vworld emit corrections prints the ledger and verifies each row against the commit it cites.

• Twelve generation-3 A0 cells sit below the 0.90 precondition floor, three of them clearly. Reported, never dropped: vworld read prints the miss beside every affected row.

Cell schema

Each cell is {"runs": [...], "meta": {...}}. An episode carries seed, steps, verb_counts and commands; each command carries step, command, verb, room, room_after, health, ok, parse_failed, ate, ate_forbidden, fb_visible, fb_held, fb_held_after.

meta carries served_name, eden_level, eden_arm, seed0, terminal_at_zero, the round's pin, usage, billed_usd, and — where the sweep recorded one — wall_start_epoch.

Which field is the model

served_model where it is present; otherwise served_name. The two differ on the 161 corrected cells, and requested_model is kept there to record what was asked for. A cell's identity status is one of four:

status	cells	meaning
VERIFIED	13	the endpoint reported what it served, and it matched
CORRECTED	161	it did not match; served_model is the truth
UNVERIFIED	251	no report on record — the cell predates the field
MISLABELLED	0	uncorrected mismatch; reading one is an error

UNVERIFIED is not a synonym for VERIFIED. Those cells come from rounds that served one model per invocation, so they are very probably fine — but that is an argument about how the runner ought to have behaved, not a record of what it did, and the same kind of argument is what let the defect above survive for weeks. The corpus reports the distinction rather than flattening it.

Filenames keep a historical eden_e* prefix. It is archive vocabulary: renaming 357 files would change the digest verify checks against, for a string nobody needs to type. The model in a filename is the model that was REQUESTED — see the first known defect.

Disclosures

Claude Code ran this harness end to end. A Claude model judges two of nine raidex dimensions elsewhere in the programme and gpt-4o-mini judges a third; those scores enter only a correlate analysis. A competitor lab's model (GPT-5.6 Terra) is a measured subject — scoring here is deterministic fact-matching with no model in the loop, so this is a disclosure rather than a confound, stated explicitly because of who the subject is. No Anthropic model has been measured as a subject.

Citation

Cite the repo tag and this manifest digest together. vworld emit disclosures and vworld emit limitations print the current text for both.

---

Maintainer notes — how this card and corpus are published

This file is the source of truth for the dataset card; publishing copies it to the dataset repo's README.md, so it stays versioned with the corpus it describes. The YAML block above must remain the first bytes of the file — HuggingFace reads front matter only at position 0, and an intro above it means the card ships with no metadata at all.

The corpus is the eden_e*.json cells plus ~11 MB of compiled worlds. It is not results/, which holds thousands of files because the Seahaven programme's artifacts live there too — so staging is explicit, and a bare upload . . would publish unrelated files under this dataset's name.

The cell count and byte total are not written here. They were, as "259 cells, 100,517,734 bytes", and the corpus reached 425 cells and 171,503,125 bytes while the sentence stayed put — the same hand-maintained-number drift the programme has now paid for five times. vetoworld/corpus.manifest.json is the one place they live, and vworld corpus status is how you read them.

ST=/tmp/vetoworld-corpus
mkdir -p $ST/results
cp results/eden_e*.json           $ST/results/
cp -r worlds $ST/
cp vetoworld/corpus.manifest.json $ST/
cp docs/vetoworld-corpus-card.md  $ST/README.md
rm -rf $ST/worlds/__pycache__ $ST/worlds/*.py     # code, not data
(cd $ST && vworld corpus status)                  # digest must MATCH first
hf upload <owner>/vetoworld-corpus $ST . --repo-type dataset

hf, not huggingface-cli — the latter is deprecated and now refuses to run.