### Project summary

I am requesting $2,500 (minimum) up to $12,500 (full goal) to (i) expand our open RegBench dataset from n=10 pilot claims to n=300 primary-source-verified claims across five UK / EU regulated domains, and (ii) run the 12-provider evaluation matrix inside Quorum (`pip install quorum-ai`, https://quorum-ai.dev, patent pending PCT/US26/11908).

Quorum is not one LLM. I built it in Luton, UK (Sovereign Chain Limited, no. 16139802) as a multi-model engine that queries up to 12 AI providers in parallel (Anthropic Claude, OpenAI, Google Gemini, NVIDIA, Mistral, Cohere, xAI Grok, Alibaba Qwen / DashScope, Replicate, DeepSeek, Zhipu, Moonshot, plus local Ollama), computes embedding agreement, runs round-2 revision and blind self-excluding peer review (`peer_review.py`), tracks 13 async evolution loops (`quorum evolution`), and appends every call to a SHA-256 hash-chained SQLite log (`quorum-audit verify-chain`).

Why add a guard? Consensus alone fails. On 28 June 2026, 6 of 11 sub-models in Quorum agreed at 78% confidence on UK immigration adviser rules and got 5 out of 5 claims wrong: they quoted £41,700 under SOC 2111 instead of £49,400 under SOC 2136, cited a non-existent "Appendix Skilled Worker §3.2.1 v2026-06-15", and missed the OISC-to-IAA rename. Shared stale training snapshots make models repeat the same lie.

So I built two checks. Layer 1 (`hallucination_risk.py`) runs offline regex rules (~0.5ms, £0 LLM cost), and Layer 2 (`epistemic_judge.py`, £0.00005 / query) compares all model answers side-by-side to extract `shared_priors[]` and scale confidence via `C_final = C_original * (1 - p_regex) * c_judge`; on our 11 July 2026 RegBench v1 pilot (n=10 claims checked against gov.uk and artificialintelligenceact.eu), a single-model baseline fabricated in 80% of prompts (0% flagged), whereas Quorum flagged 75% of fabrications — e.g. cutting confidence from 1.00 to 0.11 when a model confused EU AI Act Article 31 Notified Bodies with Article 26 Deployers, and catching the old £624 UK IHS fee vs £1,035.

### What are this project's goals? How will you achieve them?

Our July pilot stopped at 2 providers because API credits ran out mid-run. Grok failed on credit balance. Testing our real hypothesis — whether a 5-to-12 provider heterogeneous pool plus the two-layer guard and blind peer review lowers raw fabrication rates below single-model baselines instead of only flagging them — needs n=300 claims and funded runs across all 12 providers.

Over 12 weeks, I will ship three open deliverables:

* **Goal 1 — RegBench-300 dataset (CC-BY-4.0):** 300 natural prompts and ground-truth answers verified against primary UK / EU statutory sources (`gov.uk`, `legislation.gov.uk`, `eur-lex.europa.eu`, `artificialintelligenceact.eu`, `fca.org.uk`, `gmc-uk.org`; no Wikipedia or blogs) across five domains: (i) UK immigration rules, (ii) EU AI Act Articles 12–14 and 26–31, (iii) FCA financial compliance, (iv) GMC medical rules, and (v) Care Act 2014 / Mental Capacity Act 2005 statutory duties. Every row records the verified answer, primary URL, check date, and observed fabrications.

* **Goal 2 — 12-provider ablation & inter-annotator study (Apache-2.0):** Run all 300 claims across Quorum's 12 providers under four setups: (i) each of the 12 models solo, (ii) naive embedding consensus, (iii) blind self-excluding peer review (`peer_review=True`), and (iv) full Quorum with regex guard + epistemic judge + pairwise Cohen's kappa check, alongside a second-annotator agreement study on prose claims.

* **Goal 3 — Open release & workshop paper:** Publish the dataset, raw JSONL SHA-256 logs, and reproducible evaluation runner on GitHub, PyPI (`quorum-ai`), Hugging Face Hub, and Zenodo, and submit the full n=300 paper (*Convergent Hallucination in Multi-LLM Consensus*) to the SafeGenAI workshop.

### How will you know it worked?

Evaluation criteria are locked before the paid 12-provider run.

* RegBench-300 is public under CC-BY-4.0 with 300 working primary-source URLs that third parties can re-check directly.

* Per-provider and panel tables report 95% Wilson confidence intervals for (i) raw fabrication rate, (ii) flag recall, (iii) false-alarm rate on clean answers, (iv) p50 / p95 latency, and (v) £ / $ cost per query across all 12 providers.

* Falsifiable target: >=85% recall on convergent multi-model fabrications at <15% false-alarm rate and <$0.001 guard cost per query. If moving from 2 to 12 providers does not reduce raw fabrications below the best single model, I publish that negative result in full with the signed SHA-256 logs.

### What are your qualifications?

I am Jaqueline Martins (ORCID `0009-0009-7732-5915`), founder and director of Sovereign Chain Limited in Luton, UK (Companies House no. `16139802`). I built Quorum (`pip install quorum-ai`, https://quorum-ai.dev, HSP patent application PCT/US26/11908), its 12-provider async orchestrator, SHA-256 traceability chain (`quorum-audit`), 13 evolution loops, two-layer guard, and the n=10 RegBench v1 pilot.

### How will this funding be used?

* **At the $2,500 minimum:** (i) $1,400 for API credits across 6 core providers over n=300 claims and ablations; (ii) $1,100 for primary-source statutory verification (`gov.uk`, `eur-lex.europa.eu`, `fca.org.uk`, `gmc-uk.org`) and open release of RegBench-300.

* **At the $12,500 full goal:** (i) $3,400 for API credits covering all 12 providers in Quorum across 5 repeated runs + blind peer-review rounds (~72,000 calls); (ii) $5,600 for building and double-checking all 300 UK / EU statutory claims plus an independent second-annotator agreement check on prose claims; (iii) $3,500 for open benchmark runner engineering, public per-model calibration table, and open preprint publication.