# Comparator addendum (2026-07-13) — polling comparison for the AZ GOP gubernatorial primary Status: pre-registered before the 2026-07-21 Arizona result. This addendum fixes, in advance and by rule rather than by hand, WHICH external polling comparisons the us_arizona case study publishes alongside SONAR's frozen prediction. It adds to the anchored record; it modifies nothing. It does not touch any locked value, any anchored interpretation branch, or any anchored literal. Contest: `us_az_gov_r_primary_2026` (Andy Biggs vs David Schweikert). Election day: 2026-07-21. ## a. The locked NPI benchmark is immutable — no re-lock The polling benchmark recorded on the AZ lock at lock time is the pre-registered comparison and stays exactly as locked: - Noble Predictive Insights (NPI), `as_of` 2026-05-07, two-way **72.7 / 27.3** (Biggs / Schweikert), normalized from raw Biggs 48 / Schweikert 18 / undecided 34 by proportional undecided allocation, MoE +/-5.01. This value is frozen on the locked prediction. It is NOT re-locked, NOT recomputed, and NOT replaced by anything in this addendum. Every calibration score against the locked benchmark continues to use this number. ## b. The case study also publishes a latest-pre-election-poll comparator, chosen by rule In addition to the locked benchmark, the case study publishes a second, clearly-labeled comparator: the **latest pre-election poll of this race**, selected mechanically by the following rule (no operator discretion): 1. Eligible poll: a publicly released survey of the us_arizona GOP gubernatorial primary whose field dates fall **entirely before election day** (2026-07-21), with a **published methodology** and a **named sample size**. 2. Selection: among eligible polls, the one with the **latest field-end date** wins; if two share a field-end date, the **later public release** breaks the tie. 3. Two-way conversion: **decided-only normalization** — the two named candidates' shares are renormalized to sum to 100, with **undecided (and any minor-candidate share) excluded**. As of this addendum the rule selects: - NPI, released **2026-07-07**, fielded **2026-06-29 to 2026-07-01**, n = 425 likely GOP primary voters, MoE +/-4.75. Raw: Biggs 60 / Schweikert 10 / undecided 26. Two-way by decided-only normalization: **85.7 / 14.3** (Biggs / Schweikert), from 60 / (60 + 10) = 85.714 and 10 / (60 + 10) = 14.286. If a qualifying later poll is released before election day, the rule selects it automatically and the case study renders that poll instead — no edit to this addendum required. If no eligible poll exists, the comparator section renders "none recorded". ## c. Both comparisons publish regardless of which flatters SONAR The case study publishes **both** the as-locked benchmark (a) and the latest-pre-election comparator (b), side by side, whichever one is closer to SONAR or to the eventual certified result. The asymmetry is acknowledged and accepted: SONAR's number froze on **2026-06-30**, 21 days before the election, while the comparator poll may field as late as election eve. A later poll has strictly more information than SONAR's frozen prediction; showing it anyway, even when it is closer to the result, is the honest presentation and is the point of publishing both. ## d. Prediction-market prices are winner-probability context only The Polymarket entries in the event registry (the `polymarket:arizona-governor-republican-primary-winner` anchors) are **winner-probability** context only. They are never converted to a vote-share and are never used as a share comparator. Winner probability and vote share are different quantities; the case study keeps them distinct and does not manufacture a share number from a price. ## e. Anchored interpretation branches are unchanged The pre-registered AZ interpretation branches (raw wins / shadow wins / NPI beats both / everyone misses the winner) and their anchored literals are unchanged by this addendum. They remain keyed to the **locked** values only (the 72.7 / 27.3 benchmark of (a), the shadow, and the raw prediction). The latest-pre-election comparator of (b) is a **display** comparison for the case study; it feeds no interpretation branch, no promotion rule, and no anchored threshold. --- This addendum is externally anchored (OpenTimestamps + git) exactly like the other pre-registration artifacts; see `calibration/ANCHORS.md`. Its digest is fixed there before the 2026-07-21 result.