Elicitation validates the text-to-rating mapping on human data, and records the simulation claim as owed

"SSR works" is two claims. The mapping claim: given what a person wrote, SSR recovers the rating that person gave. The simulation claim: simulated personas' purchase-intent distributions for a product match real people's. The architecture expected an acceptance gate against published human purchase-intent data; the paper's 57 surveys belong to a company and are not public, and no public benchmark of concept-test purchase intent with free text is in hand.

Elicitation therefore validates the mapping claim, on public human product reviews — a person wrote the text and gave the star rating, so the ground truth is real — scoring SSR's distributions against those stars with log loss, Brier score and rank correlation, against a baseline that does not read the text. It is run small, on a few hundred reviews balanced across stars, to move fast. The simulation claim is recorded as owed, and the engine's trust level stays uncalibrated (ADR 0008) until a real purchase-intent benchmark exists.

Embeddings come from Amazon Titan Text Embeddings v2 on Bedrock, reached through a local LiteLLM proxy that also serves the chat models, because neither of Bedrock's OpenAI-compatible endpoints serves embeddings and OpenAI's embedding models are not offered on AWS. The paper used OpenAI text-embedding-3-small; whether SSR survives the substitution is measured by the review validation, not assumed, with Cohere Embed v4 as the next candidate.

Considered options

Relying on the paper's reported agreement was rejected: it was tuned on its own evaluation data and used different anchors and a different embedding model. Running a purchase-intent survey with real people is the only direct test of the simulation claim and remains the way to discharge it, but it is a project of its own, not a module gate.

Consequences

Reviews measure satisfaction, not purchase intent, so the validation uses a satisfaction anchor set frozen before it runs, and it validates the mechanism — not the purchase-intent anchors, whose construct the data does not contain. Those rely on the no-human-data checks of ADR 0027. A passing review validation means SSR can read human text into human ratings with this embedding model; it does not mean simulated purchase intent matches people's.