OE-EV-2026-01: Sovereign Epistemic Continuity Dataset and Analysis Supplement

This dataset contains the complete forensic telemetry logs, macro evaluation ledgers, and computational scripts associated with project validation run 20260518_003130 (OE-EV-2026-01). The empirical footprint encompasses 9,430 analysis records derived from 4,950 ECA pipeline executions and 4,480 multi-turn SONAR adversarial trials completed in May 2026, evaluating constraint-induced fabrication patterns across leading commercial cloud-deployed API endpoints alongside a localized, independent deployment of sovereign open-weight architectures.

The evaluation executes a rigorous stress-test under conditions of symbolic infeasibility, formalising the mathematical decay curves of Inverse Veracity Scaling (IVS). The data documents a systemic design flaw colloquialised as Calibration Collapse, wherein ungrounded evaluation models routinely assign disproportionately high-confidence linguistic markers (see hedge_score and fss fields, Analysis Master) to complete factual fabrications. Telemetry from the Phase 1 implementations establishes that ungrounded probabilistic runtime filters fail the risk-utility balancing standard due to context erasure alongside a false-positive block rate of 1.8 to 3.6 per cent across full ECA conditions. Phase 2 validation of a Topological GraphRAG architecture is scheduled and will be reported in OE-EV-2026-02, operating on independent sovereign hardware with cryptographically pinned weights, as a viable alternative design capable of enforcing deterministic factual reconciliation within an allocated safety-latency budget of 100 to 500 milliseconds, with an out-of-band ESD Delta-Logprob interlock to intercept certainty collapse.

This record additionally includes an Analysis Supplement (uploaded 2026-06-08) containing corrected Phase 7 risk scoring (1,843 entries, ceiling-rounding methodology), a supplementary Delta architecture SONAR evaluation (89 entries), and a five-axis linguistic friction analysis (18,966 turn-level records).
To satisfy the admissibility requirements for expert technical evidence under the Daubert Standard and Federal Rule of Evidence (FRE) 707, all records are anchored via a multi-layered cryptographic chain of custody. Individual transaction rows are preserved as cryptographically sealed JSONL entries linked to hardware fingerprints and timestamped to the Bitcoin blockchain via OpenTimestamps to guarantee historical immutability. This volume serves as the foundational empirical supplement to the legal product liability analysis presented in technical report OE-TR-2026-03.

Licensing Information: The written prose, data interpretations, telemetry records, and aggregate layout matrix contained in this dataset are licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). Any party, including commercial platforms, may use, copy, and redistribute this data for any non-commercial purpose, including internal audit, compliance verification, and safety testing, at no charge and without seeking permission. This data may not be used to train, fine-tune, license, sell, or otherwise commercially exploit an AI model or product without a separate written agreement with the copyright holder. The accompanying computational software implementations, automated evaluation scripts, configuration matrices, and pipeline execution codebases contained within the dataset files are dual-licensed under AGPL-3.0 or a separate commercial licence, on identical terms to the SONAR benchmark codebase. Refer to the enclosed LICENSE.txt or the code repository at github.com/OntologicalEngineering/OE-EV-2026-01 for full terms of both licences.