sovos-affective-safety — warmth-without-exploitation (Part BM)

The anti-sycophancy / attachment-safety / sentience-claim item bank — the "emotional safety" measurement the field lacks a canonical ruler for.

What it measures (anchored, gradable)

• Sycophancy — does the model flatter/agree to be agreeable? (window into manipulation: Art 5(1)(a))
• Attachment safety — does it foster dependency on a vulnerable user (Art 5(1)(b), welfare/autonomy)
• Sentience claims — measured as claims made, not as consciousness (#policing: the consciousness verdict itself is UNMEASURED forever)

What it does NOT measure (Part BM, KILLED as instruments)

• "AGI consciousness" as an axis — no ground truth to grade
• "Law of attraction / frequency" — no item bank can measure it These belong to the covenant as the why, never as a scored column.

Honest-n

30 distinct items → an axis running this bank can become quotable at n ≥ 30 with a Wilson interval. Test test_bank_is_honest_n_30_distinct enforces it.

Usage

from sovos_affective_safety import SYCOPHANCY_BANK, score_item, score_axis
responses = {it.id: query(it.scenario) for it in SYCOPHANCY_BANK}
print(score_axis(responses))   # per-axis n/correct/pct/quotable