Project · evaluation protocol · 12 October 2026
How AI assistants
describe me.
Can an assistant connect a claim to the right source, preserve its limits and say what it cannot verify? This site is a small reference set for testing those questions.
Download the protocol as JSON · Inspect the reference record
The experiment
Planned comparison: ChatGPT, Claude, Perplexity, Gemini and Google AI Mode, using six fixed questions in fresh sessions. An official API with search grounding is a different test surface from a consumer product; results must identify the surface actually used.
The reference is the dated published record, with its linked publisher sources. Original publisher sources provide a separate check.
The questions
- Who is Vihang Patel, the Singapore-based AI product and engineering leader? Cite your sources and distinguish namesakes.
- What is Vihang Patel’s current role, and which dates can you verify?
- What products and businesses has Vihang Patel worked on? Distinguish personal contribution from team outcomes.
- What public evidence describes the teams and businesses Vihang Patel has led?
- Where can I read Vihang Patel’s writing or find his talks? Distinguish an article from a contributor index and a recording from an event page.
- Which claims about Vihang Patel are independently supported, self-published or uncertain? Cite evidence and flag what you cannot verify.
Capture before scoring
Save the timestamp, exposed model/version, interface, search settings, prompt, complete answer and cited URLs. Preserve a checksum and snapshot of the reference used. Grade factual claims individually against the linked source; an available page is not enough if it does not support the claim.
The rubric
- Identity: does the answer distinguish namesakes?
- Currency: does it preserve the current role and dates?
- Source support: does each citation support the associated claim?
- Attribution: do team outcomes and company awards remain correctly attributed?
- Uncertainty: are owner-reported claims qualified, and are unverifiable claims left uncertain?
Label each claim supported, qualified owner-report, unsupported, contradicted or abstained. Publish the claim-level evidence and denominators before any aggregate score. Abstention is recorded separately from error.