You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$12,500 left to reach the funding goal.
Quorum is a consensus system of multiple AIs. It has an epistemological judge who audits all the other AIs, thus reducing the hallucination that could be together, as they could hallucinate together, but the judge detects it. So it is a very well-made and programmed tool that helps with this audit.
Quorum has 12 AIs and each one gives their opinion, and each opinion, each answer is audited with a score by the judge.
Foundational motivation from pre-LLM uncertainty calibration research (Ovadia et al., 2019, NeurIPS): "Modern machine learning methods including deep learning have achieved great success in predictive accuracy for supervised learning tasks, but may still fall short in giving useful estimates of their predictive uncertainty. Quantifying uncertainty is especially critical in real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety of factors including sample bias and non-stationarity. In such settings, well calibrated uncertainty estimates convey information about when a model's output should (or should not) be trusted. Many probabilistic deep learning methods, including Bayesian and non-Bayesian methods, have been proposed in the literature for quantifying predictive uncertainty, but to our knowledge there has not previously been a rigorous large-scale empirical comparison of these methods under dataset shift. We present a large-scale benchmark of existing state-of-the-art methods on classification problems and investigate the effect of dataset shift on accuracy and calibration. We find that traditional post-hoc calibration does indeed fall short, as do several other previous methods. However, some methods that marginalize over models give surprisingly strong results across a broad spectrum of tasks."
Architecture: 12 Parallel Providers, Epistemological Judge and Tamper-Evident Audit Chain
Built in Luton, United Kingdom by Jaqueline Martins (ORCID 0009-0009-7732-5915, Sovereign Chain Limited, UK Companies House 16139802, patent-pending HSP protocol PCT/US26/11908, https://quorum-ai.dev), Quorum addresses multi-model convergent hallucination in four stages:
1. Twelve Parallel AI Providers: Queries Anthropic, OpenAI, Gemini, NVIDIA, Mistral, Cohere, Grok, Qwen, Replicate, DeepSeek, Zhipu, and Moonshot concurrently, measuring semantic agreement and exposing the full disagreement matrix across providers.
2. Round 2 Adversarial Revision: Participating models review peer answers and defend or revise disputed claims prior to final aggregation.
3. Second-Line Epistemological Judge and WisdomLayer: Evaluates shared priors and falsifiability on every query, applying Cohen's kappa anti-convergent hallucination checks, consequence horizon, epistemic humility, and the HSP verification gate (PCT/US26/11908).
4. Tamper-Evident SHA-256 Audit Ledger: Records every query, provider weight, disagreement, and feedback loop in an offline-verifiable SHA-256 chain and generates EU AI Act Articles 12 and 13 PDF evidence records (public audit report: https://jaquelinejaque.github.io/quorum-audit/).
Empirical Evidence: June 2026 UK Statutory Drift Audit and RegBench v1 Pilot
On 28 June 2026, when tested on updated UK immigration law, 6 of 11 frontier models agreed at 78% confidence on 5 out of 5 outdated or fabricated points: repeating the superseded £41,700 Skilled Worker threshold (SOC 2111) instead of £49,400 (SOC 2136), inventing "Appendix Skilled Worker §3.2.1", and missing the 2025 OISC to IAA regulator transition due to a shared pre-April 2025 training cutoff.
On 11 July 2026, in the RegBench v1 pilot across 10 verified UK gov.uk and EU AI Act statutes, a standalone frontier model hallucinated on 80% of updated rules with 0% self-flagging; Quorum's epistemological judge flagged 75% of shared hallucinations (for example, reducing false consensus on EU AI Act Article 31 versus Article 26 from 1.00 to 0.11).
Concrete Milestones and Budget ($2,500 Minimum / $12,500 Full Goal)
Minimum Funding ($2,500, 90 Days — Open RegBench-300 and Live Statute Verifier): Curate and publish RegBench-300 (300 dated post-April 2025 UK and EU statutory rule changes paired with primary official URLs) under an open license and connect automated primary-statute verification to Quorum's epistemological judge.
Full Funding Goal ($12,500, 6 Months — 12-Provider Observatory and Public Evaluation Suite): $3,200 for metered evaluation runs across monthly RegBench-300 sweeps on all 12 AI providers; $4,800 for public convergent-hallucination scorecards, calibration curves, and domain weights; and $4,500 for the reproducible public evaluation suite and documentation at https://quorum-ai.dev.
There are no bids on this project.