# DEVIATIONS — area-search-map-census-2026

1. 2026-09-30, after collection, during the pre-registered hand check of ambiguous place names: the pre-registered rule (v1, substring match) counted foreign namesakes and word fragments as Singapore places: "orchard park" / "port orchard" (US), "queenstown tasmania / eastern cape", "kallangur" (Queensland), "changing", "clementine". v1 gave 319 hits of 620. Raw suggestions were kept; NO new collection was made.
2. A first correction (v2: explicit Singapore evidence required for every place) gave 157 hits and was checked by hand: 30 of 30 random v2 hits were genuine, but about two thirds of a random sample of 30 cells that v2 dropped were genuine local completions (for example "pet clinic tampines"). v2 is therefore a floor.
3. Final rule (v3, chosen after seeing v1 and v2 results, so it is not pre-registered): whole-word place match + trade stem + no foreign marker for every place; for the 8 ambiguous place names (Orchard, Woodlands, Queenstown, Marine Parade, Clementi, Changi, Kallang, Bishan) also explicit Singapore evidence. v3 gives 265 hits of 620. Published CSV = v3. rescore_v2.py and rescore_v3.py are in the folder; raw.json holds every suggestion.
4. Hand check of v3 with a fresh random seed: 30 of 30 random hits were genuine Singapore completions; of 30 random non-hits that had any suggestions, about 3 looked like genuine local completions dropped by the ambiguity rule (mostly ambiguous names). So v3 is close to the truth and slightly conservative for ambiguous places. It is not proof.
5. 212 of 620 queries returned no suggestion at all. Those count as non-hits ("no completion returned"), not as "no demand".
6. 2026-09-30, after collection: CLAIM resolved with the measured figure (265 of 620, rule v3). ORIGINAL pre-registered wording: "Only SingRank can say in how many of 20 service trades each of 31 Singapore towns and districts shows up as a place people add to their search, counted from Google Autocomplete on one date with one fixed protocol." The measured question is the same; the headline ratio was chosen after seeing the counts.
7. No dated, hashed copy of 01b-signature.md was made before collection (the signature skill was built the same day). The plan file was written before the first request and has not been edited except item 6.
8. Post-hoc descriptive counts added for the article, not pre-registered: the three trade groups (six trades, seven mid, seven low), empty responses by group, hit rate among searches that returned any completion, and regional totals excluding the 8 ambiguous place names. The earlier draft's counts of hits with mall/MRT/plaza-type words (97 of 265) and street/compass-type words (59 of 265) were REMOVED on 2026-09-30 because the word lists were never saved in a script and the numbers cannot be reproduced.
9. Places are grouped into five regions by SingRank's regional guides (Central incl. the Central Area, East, North, North-East, West). The URA Master Plan 2025 regional-plans page names the regions but does not list every planning area, and the older Master Plan 2019 regional pages returned 404 on 2026-09-30, so the grouping is ours and is disclosed in the CSV region column.
10. 2026-09-30, editor-in-chief review: the bottom trade group was first labelled 'town name almost never added' (13 of 217). 171 of those 217 searches (79%) returned NO completion at all, and 171 of all 212 empty responses in the run sit in that group. The label and the advice built on it treated empties as 'no town added'. Relabelled 'Google mostly offered nothing'. Among searches with any completion the hit rate is 148 of 172 (86%) top, 104 of 190 (55%) middle, 13 of 46 (28%) bottom. We do not know whether the empties reflect low query frequency, the endpoint or the request origin (Indonesian address); the top group had 14 empties in the same run, so throttling looks unlikely but is not proved. Title, headline claim wording and tier-3 advice changed; the CSV hit column did not.
11. The wording 'pre-registered the counting rule' in an earlier draft was wrong. The rule was written down before collecting, but no dated hashed copy exists (see item 7).
12. The Singapore-evidence list for the eight ambiguous names (item 3) is loose: it accepts road, rd, central, st 22 and similar. 22 completions pass only on such weak words. A few are foreign (dentist woodlands road camberley, beauty salon woodlands road st helens, locksmith peach orchard road), some are genuine (orchard road). No hit changes, because each affected cell also has other genuine Singapore completions. The CSV column valid_suggestions is therefore inflated for Woodlands and Orchard and slightly for the others. It was left unchanged; the hit column is the reported measure.