Anchor (Reference) Statements

What It Is

Anchor statements (the paper's "reference statement sets") are the hand-authored exemplar texts for each level of a Likert scale — e.g. what a "definitely would buy" vs. "definitely would not" reaction sounds like. SSR maps a persona's free-text reaction to a rating by embedding-similarity against these anchors.

The Robustness Trick

  • Build several sets, not one. The paper used six reference-statement sets and averaged the resulting probability mass functions. Averaging across sets is what makes the mapping robust to any single set's quirks.
  • ~5–7 statements per scale level per set.
  • Writing them is "the real intellectual work" of an SSR implementation — a few hours of careful writing plus operator review.
  • No standardized scale language exists for novel audiences (e.g. developers); it must be written from scratch (App. C.1 of the paper has CPG examples to adapt).

How It Applies to Marketing Factory

Anchor sets are reusable factory assets: write several per evaluation dimension (clarity, relevance, intent-to-act), version them, and average across them at scoring time. Because they encode what each rating "means" in the target audience's voice, they are the main thing to revise when calibration-loop data shows the synthetic panel drifting from real outcomes.

Referenced from: ssr-paper-arxiv-2510-08338