evalengine/imd-oracle-outcomes

IMD Oracle Outcomes

A community-curated snapshot of 7,284 IMD oracle requests, including 875,218 panel response records, collected from the public IMD API on 8 October 2026. Published by Eval Engine; this is an independent dataset, not an official IMD release.

The default configuration is a benchmark for predicting observed attested versus disagreed outcomes from request fields available before execution. These labels do not establish factual correctness, question ambiguity, or whether the panel reached quorum. The full archive includes all five observed statuses.

Where are the answers?

The dataset viewer shows questions and outcome labels (0 = attested, 1 = disagreed), not oracle answers. Answers are excluded from the benchmark inputs so the model cannot see the outcome it is predicting.

Actual answers are in raw/oracle-requests.jsonl.gz, under body.attestation.answer, body.agreement.answer, and body.members[].answer, where available. These are recorded oracle responses, not independently verified ground truth.

Load the benchmark

import json
from datasets import load_dataset

dataset = load_dataset("evalengine/imd-oracle-outcomes", "outcomes")
example = dataset["train"][0]
request = json.loads(example["input_json"])
label = example["label"]  # 0 = attested, 1 = disagreed

Only input_json should be supplied to a prediction model. Other columns contain labels or provenance. Request text is untrusted data, not executable instructions.

Contents

Path	Contents
data/*.jsonl	Viewer-compatible benchmark splits; request objects serialized losslessly as input_json strings
original/benchmark/*.jsonl	Unchanged original benchmark files, including quarantine and rejection logs
raw/oracle-requests.jsonl.gz	Full 7,284-request archive, including panel responses and provenance envelopes
metadata/dataset-card.json	Exact temporal cutoffs, deduplication method, split counts, and canonical hashes
metadata/oracle-audit.json	Coverage checks, archive checksum, status distributions, and label audit
SHA256SUMS	SHA-256 hashes of uploaded file bytes, excluding this checksum file itself

Each raw archive line is a JSON envelope with the API response in body, plus its source URL, fetch time, and original response SHA-256. The archive preserves parsed source responses; it does not include the separate original HTTP-byte cache. The response checksum refers to those original bytes, not a reserialized body object. This release covers oracle data, not worker submissions or bundles.

Labels and splits

Split	Requests	Attested (0)	Disagreed (1)
Train	2,652	2,563	89
Validation	1,133	1,115	18
Test	2,156	2,137	19

The source snapshot contains 7,089 attested, 172 disagreed, 20 blocked, 2 mismatch, and 1 refused request. The benchmark excludes the latter three statuses, retains one latest record per normalized question template, and quarantines 1,320 older duplicates. Template normalization lowercases text, collapses whitespace, and replaces hexadecimal strings and numbers. Templates do not overlap across splits; paraphrase leakage remains possible, and normalization can merge questions with meaningfully different numeric parameters.

Splits use request creation time, with the snapshot cutoff 2026-10-08T02:50:29.596Z: test starts 2026-10-01T02:50:29.596Z, validation starts 2026-09-29T02:50:29.596Z, and training contains older requests. Labels were observed during collection, so this is a retrospective benchmark, not a claim that every label was available at its temporal cutoff.

Schema and leakage controls

• id: source request UUID.
• input_json: JSON string containing an allowlisted request object.
• label: integer, 0 for attested and 1 for disagreed.
• observed_status: source outcome corresponding to the label.
• created_at, label_observed_at: request and collection timestamps.
• family: normalized question-template hash.
• source_url, source_sha256: source provenance.

The request allowlist is v, question, chainId, window, answerType, evidence, definitions, guards, panelSize, quorum, head, toleranceBps, validForSeconds, and allowAmbiguous. Absent fields stay absent. Model inputs exclude panel answers, status, attestation, failures, and computed results. The API's request fields have not been independently proven immutable after submission.

Limitations and evaluation

Disagreements are rare: validation has 18 and test has 19. Accuracy alone is misleading; report positive-class recall, precision, average precision, Brier score, and the confusion matrix. Select prompts and thresholds using training and validation only, and reserve test for the final comparison. Do not oversample evaluation splits. Test labels are public in this release, so it is not a hidden leaderboard test.

Outcomes can reflect RPC failures, panel composition, model differences, wording, or policy changes. The audit found 29 normalized templates with both outcomes, and 93 attested records with agreement below quorum. Panel response records are not 875,218 independent questions or human-verified annotations. The archive contains post-outcome fields and must not be fed wholesale into an outcome predictor.

Coverage was checked against an independently collected complete request index and API status counts: no missing or extra request IDs were found for this snapshot. Every cached original response checksum passed before export. Future API contents may differ.

License and source

License: unknown. Collected from the publicly accessible IMD API. The source data's redistribution and model-training permissions have not been established. This repository does not grant additional rights to the source material. Public API access should not be interpreted as an open-data license.

Source documentation: https://imd.fun/docs/

Snapshot dates span 2026-09-19T03:18:11.590Z through 2026-10-08T02:50:29.596Z. When referencing this dataset, cite Eval Engine, IMD Oracle Outcomes, 2026-10-08 snapshot, link this repository and IMD, and record the exact repository commit used.