Hugging Face's logo

Datasets:
Oracle4CEG
/
OracleEconLab-Oracle-Incentives-v1

Tasks:
Tabular Classification
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
blockchain
oracle
economics
trustworthy-ai
tabular
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
OracleEconLab-Oracle-Incentives-v1 / README.md
LukeS3's picture
LukeS3
Document scoped Atlas authors, acquisition supplement and Colab acceptance
356fb1a verified
preview code
|
Raw History Blame Contribute Delete
17.8 kB
metadata 
license: cc-by-4.0
language:
  - en
task_categories:
  - tabular-classification
pretty_name: OracleEconLab Oracle Incentives and Accountability v1
size_categories:
  - 100M<n<1B
tags:
  - blockchain
  - oracle
  - economics
  - trustworthy-ai
  - tabular
configs:
  - config_name: abc_dictionary
    data_files: data/abc_dictionary.parquet
  - config_name: uma_real_episode
    data_files: data/uma_real_episode.parquet
  - config_name: uma_real_episode_abc
    data_files: data/uma_real_episode_abc.parquet
  - config_name: uma_public_rpc_episode
    data_files: data/uma_public_rpc_episode.parquet
  - config_name: uma_public_rpc_abc
    data_files: data/uma_public_rpc_abc.parquet
  - config_name: uma_decisions
    data_files: data/uma_decisions.parquet
  - config_name: uma_decision_splits
    data_files: data/uma_decision_splits.parquet
  - config_name: decision_evidence
    data_files: data/decision_evidence.parquet
  - config_name: ai_feature_dictionary
    data_files: data/ai_feature_dictionary.parquet
  - config_name: public_ai_episodes
    data_files: data/public_ai_episodes.parquet
  - config_name: public_ai_decisions
    data_files: data/public_ai_decisions.parquet
  - config_name: public_ai_splits
    data_files: data/public_ai_splits.parquet
  - config_name: public_ai_evidence
    data_files: data/public_ai_evidence.parquet
  - config_name: public_ai_predictions
    data_files: data/public_ai_predictions.parquet
  - config_name: public_ai_metrics
    data_files: data/public_ai_metrics.parquet

OracleEconLab Oracle Incentives and Accountability v1

This review-sized draft turns fixed, public UMA protocol evidence into complete or right-censored economic lifecycles. It is a release candidate for the paper Who Verifies Decentralized Information? Economic Incentives, Accountability, and Trustworthy AI Across Oracle Protocols.

Dataset configurations

Configuration Rows Observation unit Purpose
abc_dictionary 28 economic variable Formulas, units, sources, missing rules and A/B/C family
uma_real_episode 1 cross-chain episode End-to-end Polygon OOV2 to Ethereum DVM validation
uma_real_episode_abc 28 case-measure status Honest mapping of every A/B/C measure to the real case
uma_public_rpc_episode 1 cross-chain episode Clean-room episode reconstructed without curated-ledger input
uma_public_rpc_abc 28 case-measure status Public-source-only A/B/C mapping and explicit unavailable states
uma_decisions 810 challenged proposal-time decision Minimal four-action AI evaluation
uma_decision_splits 810 decision split assignment Frozen chronological train/validation/test partition
decision_evidence 2430 timestamped evidence reference Prediction-level provenance and leakage audit
ai_feature_dictionary 16 decision-time feature Feature meaning, source and availability
public_ai_episodes 64 public-source cross-chain episode Economic lifecycle reconstructed from fixed public receipts
public_ai_decisions 64 proposal-time decision Credential-free public-source AI demonstration inputs
public_ai_splits 64 split assignment Label-free chronological 44/10/10 partition
public_ai_evidence 192 timestamped evidence reference Decision-time traceability for every demonstration prediction
public_ai_predictions 70 model/action output Four-action predictions, confidence, evidence and protocol endpoint
public_ai_metrics 7 model evaluation Calibration, coverage, abstention and stylized economic cost

A/B/C economic measure families

  • A — Verification concentration (Who verifies?): HHI, top-k share, effective number of monitors and threshold monitor count.
  • B — Incentives and frictions (Why verify?): reward/bond, capital-days, fees, Gas, verification cost and scenario-based expected private utility.
  • C — Accountability outcomes (What happens?): dispute, adjudication, transfer, realized payoff, delay and right censoring.

A/B/C are measure families. They are not the project's Sample A/B/C data cohorts. Identity, decision-time, cross-chain-link and provenance fields remain supporting fields rather than being mislabeled as economic outcomes.

Construction and provenance

The fixed cutoff is 2026-06-30T23:59:59Z. The real case begins with canonical Polygon request, proposal, dispute and settlement records, links them to the Ethereum DVM with a Grade-A deterministic match, reconciles token flows, and separates returned principal from realized reward. Every processed episode or decision contains transaction/evidence references and a deterministic provenance identifier. File integrity is recorded in checksums.csv.

The main reconstruction commands live in the GitHub real-release-v1.5 branch. public_evidence/ includes the five fixed receipts, canonical headers, case specification, source checksums and verified contract-semantic rules used by the clean-room entry. This draft contains review-sized processed evidence; it does not claim that every protocol ledger can be regenerated without archive infrastructure.

The executed notebooks/uma_minimal_trustworthy_ai_demo.ipynb loads the frozen tables, reruns the leakage audit and chronological evaluation, and writes the review-sized artifacts under demo_outputs/. Its reported performance is not edited or filtered after execution.

The independent public_ai_evidence/ path closes the public-source chain for a fixed 64-episode demonstration: a label-free transaction registry, 280 canonical Polygon/Ethereum receipt and header bundles, decoded economic episodes, decision-time features, model outputs and QC. Live mode uses credential-free public endpoints; offline mode deterministically replays the frozen snapshot. The model result is intentionally reported even when weak. It is a pipeline demonstration rather than evidence of production challenge quality.

Trustworthy-AI task

At the OOV2 proposal timestamp, the benchmark emits one of Accept, Investigate, Challenge, or Abstain. The endpoint is whether the later UMA protocol resolution rejects the proposal. Only decision-time fields may be used as model inputs. Columns whose names end in _outcome_only, the protocol label, dispute result and settlement evidence are evaluation-only and must not enter the model.

The 810 observations condition on proposals that were actually challenged, Grade-A linked, price-consistent and flow-exact. The task is therefore a conditional protocol-outcome evaluation, not a causal policy for all proposals and not an independent truth classifier.

Intended uses

  • audit economic variable construction and source traceability;
  • reproduce the real UMA episode and its payoff decomposition;
  • evaluate calibration, abstention, leakage and robustness on the conditional decision cohort;
  • study A/B/C constructs within their declared observation units.

Uses that are not validated

  • production deployment of automated challenges;
  • causal claims that rewards or penalties improve truthfulness;
  • treating protocol acceptance or DVM resolution as independent factual truth;
  • identity, geography or intent inference from blockchain addresses;
  • aggregation of heterogeneous token amounts without asset-preserving conversion.

Biases and limitations

The decision cohort has selective labels because it includes only actually challenged UMA requests. It over-represents mechanisms and market adapters for which exact cross-chain and token-flow evidence is recoverable. Address-level concentration is not entity-level concentration. Off-chain investigation labor is unavailable and remains a declared scenario. Independent factual truth is available only for a small external-rule subset outside the core protocol label.

Personal and sensitive information

The data contain public pseudonymous contract/account addresses, transaction hashes and block metadata. They contain no asserted mapping from an address to a natural person and must not be used to infer human identity. RPC endpoint URLs, credentials and private keys are not included.

Licensing

Original curated tables and documentation are CC BY 4.0. Upstream contracts, APIs and third-party records remain governed by their own terms. See DATA_LICENSE.md and CITATION.cff.

Pinned research delivery added 10 October 2026

The existing root-level data/ and public-evidence directories remain the previously released demonstration resource. Two separately named archives now support the revised agentic-audit research:

  • atlas/selected-inputs/2026-10-10/: four fixed curated UMA source tables used by research cohort, feature and label reconstruction. This is a selected Atlas snapshot, not the complete 67-file candidate Atlas or a claim that its known semantic problems are resolved.
  • research/agentic-audit/2026-10-10/: experiment-specific 810-case cohort and chronological splits, frozen evidence, original Qwen response records, additional experiments, archived failed/interrupted/superseded attempts, 73 reference CSVs, 18 numeric LaTeX tables, exact prompts and eight original figures. Historical GPT records are separate provenance and are excluded from Qwen comparisons.

The active supplemental scope is 483 validation runs, 320 matched B2 runs (repeats 3 and 4), and 480 A1 guard runs. Optional low-temperature experiments were cancelled. Original Qwen main and corruption panels contain 1,600 and 320 case-runs. Raw model outputs are machine-generated records, not endorsements of their contents or independent factual truth. Read each manifest and the research limitations.

Research code: https://github.com/Oracle4CEG/Oracle-Agentic-Audit (delivery through a personal-fork pull request). Atlas construction: https://github.com/Oracle4CEG/Oracle4SD-Atlas. The research manifests/public-release.json pins the exact dataset commit and SHA-256 of each archive. python -m experiments.download downloads anonymously and verifies all files; python -m analysis.replay --retrain refits CPU baselines and replays saved LLM records without a GPU or inference service. Hosted Colab acceptance remains a separate check; no hosted execution is claimed by uploading this dataset.

No model weights, credentials or confidential manuscripts are included. The existing data license applies to original curated data; archived source/code and third-party content retain their original terms. Research code now has author-confirmed MIT terms. The archive does not certify a public upstream model checkpoint revision.

Recorded research acceptance and code license

The public inputs remain pinned to dataset commit ee197fc6fd6b8d980e7c7dda4feae966efe5d28c. The acceptance directory research/agentic-audit/2026-10-10/acceptance/github-notebook/ contains the actually executed notebook, environment, anonymous-download receipt, replay receipt and per-table comparison. All six code cells completed in a fresh GitHub Actions environment; 73 scientific CSV comparisons and 18 numeric LaTeX tables plus exact prompts passed, and eight figures were regenerated. The receipt records elapsed time and figure hash equality separately. Separate delivery notes record the earlier local network interruptions.

This is a GitHub Actions fresh-notebook acceptance, not a hosted Google Colab run. GitHub Actions independently passed the recorded CPU replay. No new LLM inference was performed. The commit-pinned notebook is available in research PR #2; final hosted Colab and independent scientific review remain pending.

The author confirmed MIT for original research code on 10 October 2026. See the research repository's LICENSE and DATA_LICENSE.md. Dataset and third-party terms are unchanged. This acceptance addition does not change the bytes or identities of the previously published Atlas-input and research-record archives.

Atlas reconstruction inputs — 12 October 2026

The original 15 viewer tables above remain a small review subset. Full Atlas offline reconstruction now has a separate frozen component: 64 native/curated input Parquet files, 1,807,346,651 bytes. Download the native snapshot. Five pinned public source/registry archives are also retained separately.

Executable code and commands verify file hashes before reconstructing common, economic and application layers. The source is a repair branch; execution reports specify validated scope. Historical source snapshots are retained, not silently relabeled as corrected outputs.

The root Croissant file describes only the original 15 review tables. Its JSON-LD, Boolean types, immutable file links and RAI descriptions were repaired and all 15 record sets were loaded with mlcroissant 1.1.0. See review_metadata_validation.json and review_provenance.json. These checks do not establish complete historical person/agent provenance or full-Atlas coverage. This revision has no dataset DOI.

Validated full Atlas outputs — 12 October 2026

Download the repaired Atlas: 95 canonical Parquet tables, 17,341,941,711 Parquet bytes, plus one decision-prediction Parquet and execution/metadata files. The common layer has 105,588,120 records. Layers overlap and must not be added as independent observations. Data revision: 57ab1e2409652291c75f02c9e066c52011d8b4ad. Scientific code: 6317a1b66bc6f9653b21650a7d170dc762891a31; the preserved lineage records the earlier native build and explicit repairs.

The reconstruction passes 50 release checks; 27 regression tests and all nine local notebook cells also pass. Seven financial models complete 21 fits over three tasks; three decision baselines use 810 frozen observations. Financial F1/F2 use an explicitly marked random split fallback, F3 is chronological, and unavailable metrics remain null. High pooled scores are not evidence of cross-protocol generalization, independent factual truth or profitable decisions.

The release includes table/field documentation, source and transformation provenance, before/after semantic checks, timings, memory usage and full logs. All 1,789 fields have type/definition coverage; native source-reference descriptions do not establish complete historical collection or person/agent provenance. Missing timestamps and censoring remain explicit. The old 71-file inventory and its audit failures are retained separately.

For installation, selected downloads and full reconstruction, use the Chinese quick start or full guide. The downloader verifies immutable URLs, sizes and SHA-256. Frozen replay requires no wallet, paid RPC or inference service.

The root viewer configurations and root Croissant still cover the original 15 review tables; full Atlas tables use the separately linked catalog and manifests. Archived source-registry trees preserve their upstream licenses and are not relicensed by the Dataset Card's CC BY designation. No hosted Google Colab pass, independent scientific approval or dataset DOI is claimed.

Atlas authors and acquisition supplement — 12 October 2026

For the Atlas release, the requester supplied this author order: Huaiyu Jia; Siguang Li; Lin William Cong; Luyao Zhang. Jia and Li: The Hong Kong University of Science and Technology (hkust-gz.edu.cn); Cong: Nanyang Technological University (ntu.edu.sg); Zhang: Duke University (duke.edu). These affiliations are retained as supplied; no ORCIDs were inferred. This statement applies to the Atlas release, not every other study hosted in this repository.

Download the versioned code and acquisition supplement, archived from code commit 38e63d2ed503b38b97d2fddd484f7871b9dda7de. It includes all 64 input-table provenance bindings, 39 retained historical manifests, 52 available producer identities and a fresh 688-file raw hash audit. It records conservative build-level lineage for all 95 canonical outputs. One previously mismatched historical manifest has been recovered from a fixed public revision and matched to its expected digest. Historical collection operators and executed collector commits remain unknown where original receipts are missing.

The current delivery passes 35 local regression tests; the pinned tutorial execution commit runs its original 28 tests. Colab acceptance instructions require an executed notebook, evidence ZIP and operator statement. Local execution is not a hosted Colab pass. The four-author DOI registration metadata is prepared, but no DOI has been assigned.