OusiaResearch/aurelia-runtime-corpus-v2

Aurelia Runtime Corpus v2

Aurelia Runtime Corpus v2 is a schema-gated validator corpus for the Aurelia causal world-simulation engine. It trains and evaluates a runtime-model coprocessor to distinguish verified mechanical/lore-grounded simulation records from subtly invalid candidates.

This release is the first full v2 local build: 25 fresh simulation seeds, 20 years per seed, all five Aurelia worlds, with grouped train/validation/test splits and balanced deterministic + AAA hard-negative preference rows.

No model inference, GPU training, external API calls, or paid compute were used to create this corpus. It is local deterministic simulation plus dataset construction.

Contents

Primary data files:

• —data/validator_sft.jsonl — 2,500 verified validator SFT positives
• —data/validator_dpo.jsonl — 35,000 DPO preference rows
• —data/mixed.jsonl — 37,500 total records
• —data/train.jsonl, data/val.jsonl, data/test.jsonl — grouped split by source card

Provenance and reports:

• —source/lore_history_packs.jsonl — 2,500 lore/history packs
• —source/source_report.json — source-positive build report
• —metadata/corpus_report.json — corpus totals and generator settings
• —metadata/coverage_report.json — world/year/failure-mode coverage
• —metadata/quality_gates_report.json — gate-by-gate validation
• —metadata/split_report.json — grouped split report
• —metadata/batch_run_report.json — local simulation batch summary
• —metadata/checksums.sha256, source/checksums.sha256 — SHA-256 checksums

Dataset totals

• —Seeds: 25 (6101 through 6125)
• —Years per seed: 20
• —Worlds: arkos, mirithane, solara, valdris, verge
• —Validator SFT positives: 2,500
• —Validator DPO preference rows: 35,000
• —Mixed records: 37,500
• —Lore/history packs: 2,500
• —Split: grouped by source card; no SFT/DPO sibling leakage across train/validation/test

Split sizes:

• —Train: 2,004 cards / 30,060 records
• —Validation: 270 cards / 4,050 records
• —Test: 226 cards / 3,390 records

Failure modes

Each of the 14 failure modes has 2,500 preference rows.

V1 deterministic corruptors:

• —missing_state_delta
• —invented_actor
• —wrong_year
• —missing_provenance
• —missing_lore_cause
• —canon_contradiction
• —invalid_json
• —missing_evidence_terms

AAA hard-negative families:

• —near_miss_mechanical
• —cross_world_leakage
• —chronology_impossible
• —lore_shadow
• —repair_quality_bad
• —adversarial_json_semantic

Quality gates

All v2 quality gates passed:

• —schema validation
• —no host-local path leakage
• —no duplicate record IDs
• —no content-hash collisions
• —task/output shape consistency
• —failure-mode balance
• —world coverage
• —seed coverage
• —early/middle/late year-regime coverage
• —lore-pack refs present
• —grouped split overlap check

Direct path-leak scans of validator_sft.jsonl and mixed.jsonl found zero occurrences of /Users/, /private/, or /tmp/.

Loading examples

Grouped train/validation/test split:

from datasets import load_dataset

ds = load_dataset("OusiaResearch/aurelia-runtime-corpus-v2")
print(ds)

Full mixed corpus:

from datasets import load_dataset

mixed = load_dataset("OusiaResearch/aurelia-runtime-corpus-v2", "mixed")

DPO-only rows:

from datasets import load_dataset

dpo = load_dataset("OusiaResearch/aurelia-runtime-corpus-v2", "validator_dpo")

SFT-only positives:

from datasets import load_dataset

sft = load_dataset("OusiaResearch/aurelia-runtime-corpus-v2", "validator_sft")

Record shape

All records use:

{
  "schema": "aurelia.runtime.dataset_record.v1",
  "record_id": "...",
  "task_type": "validator_sft | validator_dpo",
  "source_kind": "derived_from_run | synthetic_negative",
  "source": {
    "run_id": "...",
    "world_id": "...",
    "year": 1
  },
  "lore_history_pack_ref": "...",
  "input_payload": {},
  "output_payload": {},
  "provenance_status": "verified",
  "generation_method": "...",
  "content_hash": "sha256:..."
}

DPO rows include output_payload.chosen and output_payload.rejected, where chosen is the correct validator rejection and rejected is the bad acceptance.

Provenance

• —Built locally on 2026-06-13 from Aurelia engine commit 65297b8
• —Raw run root during build: /tmp/aurelia-v2-runs
• —Durable local artifact root: ~/aurelia-artifacts/aurelia-runtime-corpus-v2-full-2026-06-13
• —Total local batch wall time: 894.7 seconds
• —Per-seed runtime mean/min/max: 29.6 / 25.2 / 33.4 seconds

Raw SQLite run outputs are not included in this Hugging Face dataset release; the dataset release includes portable validator records, lore/history packs, coverage reports, quality gates, split reports, and checksums.

Intended use

This corpus is intended for Aurelia runtime-validator research:

• —supervised fine-tuning of schema-gated validator outputs
• —preference training for hard-negative discrimination
• —held-out eval development for causal/lore grounding
• —regression tests for runtime model adapters

It is not a general-purpose instruction dataset.

Limitations

This is a v2 corpus, not the final AAA/v3 candidate. Before treating a trained validator as production-grade, freeze a held-out adversarial eval set and prove improvement over deterministic baselines on hard-negative discrimination, valid-card acceptance, false-acceptance reduction, and grounded repair-hint quality.

Citation

@dataset{ousiaresearch_aurelia_runtime_corpus_v2_2026,
  title = {Aurelia Runtime Corpus v2},
  author = {Ousia Research},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {https://huggingface.co/datasets/OusiaResearch/aurelia-runtime-corpus-v2}
}