Catalog record · Lending & Credit
Stage as of Aug 27, 2026 · historical report, not live reconfirmation.
Confidential external evaluation of proprietary AI models
Primary workflow source. A named organization's workflow supported by a reviewed primary source. Stage is reported separately; this label does not establish measured outcomes or independent validation.
Last researched Oct 29, 2026. Source access and reported stage are recorded below; deployment evidence does not establish ROI.
What the research supports
The pilot evaluated Gemini 2.5 Flash Lite on private AILuminate reserve prompts in an H100/Intel TDX confidential environment using PySyft. Attestation and constrained code/output handling keep model assets and evaluator prompts private to their owners.
- Reported stage
- Pilot
- Technology
- GenAI/LLM
- Business function
- Not assessed
- Impact (legacy label)
- Not labeled
- Catalog date · unconfirmed
- Not reported
- Imported
- Oct 29, 2026
- Last researched
- Oct 29, 2026
Measures & evidence
No quantified measure recorded. This is a gap in the evidence, not a finding of no value.
Documented vendor relationships
No vendor relationship has been linked to this specific workflow in the curated overlay. This does not mean it was built without vendors.
See the company’s documented relationshipsDeployment evidence & limits
Completed pilot reported August 27, 2026; not general production assurance Status as of 2026-08-27.
Stage is as reported in the cited evidence. Current operating status has not been independently reconfirmed; the research date is an editorial review date.
The technical report explains that compromised trusted code or hardware roots can break confidentiality; enclave use alone is insufficient. This demonstrates confidential evaluation execution, not financial-model validity. Trust remains in hardware vendors and the trusted software stack; scenario coverage, outputs and adjudication need bank-specific design.
Potential FS application · inferred
A bank's independent model-risk team could test vendor models on confidential underwriting, financial-crime or advisory challenge sets without exposing those tests for provider tuning or receiving model weights.
A bank's independent model-risk team could test vendor models on confidential underwriting, financial-crime or advisory challenge sets without exposing those tests for provider tuning or receiving model weights.
Original industry: AI research and evaluation. Transfer limits: This demonstrates confidential evaluation execution, not financial-model validity. Trust remains in hardware vendors and the trusted software stack; scenario coverage, outputs and adjudication need bank-specific design.
Source register
- Piloting the world's first double-blind AI evaluationsdeepmind.google · opens in a new tab
Google DeepMind · issuer · full-text · Published 2026-08-27
- Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditingstorage.googleapis.com · opens in a new tab
Google / AVERI / Singapore AISI / OpenMined / MLCommons · issuer · full-text · Publication date not recorded
Research lanes: Frontier firms: deployed workflows, safety triage and research/software operations · Batch global-discovery-2026-10-29. Coverage and remaining gaps