AOSSIE/openverifiable-enwiki-20260901-20260918-r1-evidence

OpenVerifiableLLM Wikipedia provenance evidence

Development in progress. No production-trained or end-to-end verified model is published here yet. Synthetic test results do not establish Wikipedia training.

This new AOSSIE repository is reserved for publicly reconstructible inputs, checkpoints and reports for the OpenVerifiableLLM Wikipedia base model and its conversational derivative. The governing goal and code are maintained at AOSSIE-Org/OpenVerifiableLLM.

The intended corpus is the complete eligible English article corpus from one completed dated Wikimedia dump, with a separately pinned OpenAssistant/oasst1 conversation phase. Exact files, revisions, digests, extraction policy and initialization will be published before production training. Raw reconstruction and continuous exact replay are required before any verified release claim.

Current scope: publicly downloaded/replayed synthetic development evidence, publisher identity tests, complete raw source archival and a verified public source/preparation commitment. Raw archives are under inventory-addressed raw/ prefixes with per-archive cards and licenses. G01 source identity/availability is complete: all19files were anonymously downloaded and rehashed, and the full Wikipedia compressed archive passed decompression and upstream checksums. The source/preparation statement was signed at GitHub commit f6d4371, verified against an operator-reconstructed publisher policy, archived at HF4b6a102, and verified again after a fresh anonymous download. Statement SHA-256: ec263b5c3914fd1a8bd25e97fa377b4f8416c38a9ef216d27c203252948b9129.

Original complete preparation passed: all 25,859,744 source records were accounted for, including 7,232,582 eligible articles and 7,118,121,139 Wikipedia targets. The conversation training stream has 1,278,137 targets and its separate validation stream has 65,355 targets. Preparation root: 8de1d60065b1e8e084c31c202e47367f28c8b89f11b4e895873f775bc0f7ba68. A fresh full reconstruction from the pinned raw source is running with no prepared stage-cache adoption. G02 and G07 remain pending.

The fifth guarded CUDA diagnostic passed a complete synthetic cycle: eight recorded updates, a full fresh-process replay from regenerated initialization, and a separate four-update resume from the middle checkpoint. All three states matched across all three processes. Full runtime audits remained enabled. The immutable complete archive was anonymously downloaded: all 6,265 files and all nine safe checkpoint states passed complete byte/root checks. Prior failed attempts remain preserved. This tiny cycle supplies no sustained-throughput or complete-Wikipedia replay claim. CI for source 555005f passed 1,063 tests with three existing CUDA-dependent skips.

Both external guards confirmed the pod terminated. At 2026-09-18 12:46:52 UTC, the account reported zero pods, zero volumes and zero hourly charge. Current provider-attributed rows total $0.4331596787524177 for six earlier pods; later settlement remains pending. Five full runtime-attempt ceilings of $0.575006 each plus $0.15 slack remain conservatively reserved until reconciliation. Automatic provider termination remains UNVERIFIED. G02–G10 remain pending. No production registration or model release exists. These checks were performed by the project operator, not an independent third party. The public evidence trace links scoped checks and failed attempts.

Licensing and retention

Components retain their own source licenses. Wikipedia text and attribution, OpenAssistant Apache-2.0 data, project source and operator-authored evidence must be identified separately in LICENSES.md before dataset content is uploaded. The code license is not a blanket license for Wikipedia text.

The target is at least 90 days of verification-input availability and long-term retention of final models/manifests under AOSSIE stewardship, subject to confirmed host policy. This is a retention target, not a guarantee from Hugging Face. No paid storage add-on, new subscription or permanent GPU service is authorized.

Verification claims

Artifact integrity, publisher identity, data reconstruction, continuous replay, and inference reproduction are different checks. Operator-run replay will be labeled as such. Training provenance does not establish the factual accuracy of model answers or identify which article caused an answer.