# Bilingual and coding review — 2026-09-26

Luna prepared the six question pairs and reviewed them before collection. The lead assistant reviewed the questions, analysis and publication claims afterward. All review is AI-assisted; no native human reviewer or independent human double-coding is claimed.

The pairs hold buyer intent and language-service requirements broadly constant. Two wording differences limit causal interpretation: “central Tokyo” becomes “the central part of Tokyo's 23 wards”; the Tokyo wedding question's “host about 80 people” becomes a reception for about 80 people. This is an exploratory comparison, not a clean isolation of language. The question file remains frozen and unchanged after collection.

Every response has five explicit position records in runs/main-2026-09-26/manual-coding.json. Canonical IDs normalize capitalization and bilingual variants; location-specific venues remain distinct from chain-level mentions. Four duplicate positions and three unresolved/generic positions are marked unusable. We exclude the entire Osaka coworking and Osaka web-agency scenarios from BOTH across-language and within-language name-overlap charts, rather than treating unknown identities as non-matches. Four complete scenarios (24 answers) remain; each supplies nine cross-language comparisons and three within-language comparisons per language.

A named recommendation is a name in the answer, not a independently verified real business or proof that its location/services satisfy the question. This research does not validate every suggested business. Examples checked against the saved answers include the English/Japanese Ritz-Carlton Tokyo alias, THE Tender HOUSE variants, International House of Japan variants and Landmark Square Osaka/WAKON naming. The initial regex-only analysis was superseded because it counted generic entries and missed aliases.

The initial collector's combined sourceUrls field contains both answer text and provider metadata and MUST NOT be described as visible citations. The final analyzer extracts answerTextUrls from the response text and providerMetadataHostnames from the provider response separately. All 36 answers have nonempty sets for both. Empty sets remain unknown in the general overlap helper. URLs show links, not causal provenance or verified supporting citations.

The lead review corrected the initial inconsistent exclusion rule (incomplete shortlists were previously excluded from cross-language metrics but retained in within-language metrics) and added uniqueness, per-scenario completeness and known-cost checks. No provider calls were made during recoding.

Provider-reported new successful-call cost is $0.057808. One initial network attempt has unknown cost and a separate $0.75 reserve. The older studies also retain their own $0.75 reserve; neither is silently released. See ../followup-research-budget.json.