# Polling dataset — flags and open questions (v0.3, 2026-10-14) ## 1. Sources disagree (document figures used) `register/source_conflicts.csv` lists the cases where a pollster's own document and a listing give different figures for the same question. The document figures are used in every case. - **Pan Atlantic, May 2026, Republican primary for Governor:** the report prints David Jones at 3%. Wikipedia lists 20%, which looks like a Wikipedia typo. *v0.3:* the Wikipedia copy (`-Q1`) had survived beside the document question (`-Q8`); it is now removed by the corrections step (audit PC-1). The same listing-copy pattern was removed for Emerson Sept 2016 CD2, Siena Sept 2018 CD2 and Pan Atlantic Oct 2022 CD1, and GQR Oct 2014 `-Q1` (which carried PPP's May 2014 figures, PC-3). - **Emerson, Sept 2016, CD2:** the release text gives Clinton 32 / Trump 41. RCP gives 36 / 41. *Rechecked 2026-10-09:* the only document on file is the press release ("she trails him 41% to 32%"); the full crosstab (theecps.com) was not archived. The release figure stays. - **UNH Sept 2016 and PPP Oct 2011 referendum questions** matched by "yes/no" only. These are different questions, not conflicts. ## 2. Two printings inside one document - **SurveyUSA Oct 17–19 2004, Question 1:** the topline page says No 55 and the crosstab says 54. The topline is used. - **SurveyUSA Oct 3–5 2004, Question 2:** the topline says Yes 44 and the crosstab says 43. The topline is used. - **PPP March 2011:** the release says Dodge 11 and the topline says 10. The topline is used. - **MPRC 2017, Question 1:** the "No" chart segments add to 58.5 but the legend says 59. The legend is used. - **Pan Atlantic Oct 2014, CD2 Governor:** the text gives a 7-point lead and the chart gives 6. The chart is used. ## 3. CD1/CD2 stand-ins SurveyUSA 2004 prints "Southern Maine" and "Northern Maine" columns, and the release equates them with CD1 and CD2. They are recorded as district 1 and 2, with a note on each row. Marist 2026 reports Kennebec County separately, so its CD columns leave Kennebec out. ## 4. Dates - **Field dates not printed:** Critical Insights Spring 2022, Spring 2023 and Fall 2023 give months only. They are stored as the first and last day of those months and marked "month only". *v0.3:* Fall 2023 had been loaded twice (`P2023-1102` and `P2023-1130`, same n = 614) and Spring 2022 twice (listing and document); each is now one poll. Spring 2022 uses the listing's dates (Mar 14 – Apr 7), which fall inside the document's "March and April". - **Documents without dates:** where a document prints no dates but its listing does, the listing's dates are used and marked. - **Two Mellman CD2 polls in 2018:** Wikipedia gives no dates (`P2018--the-mellman-group`, `-2`). - **Comparison polls:** 30 earlier polls are quoted inside later documents (for example "PPP's September numbers") but have no field dates. They are kept in `manual/out` and not loaded. ## 5. Pollster identity - **"Bangor Daily News" 2026-06-03:** resolved in v0.3 — it is the SurveyUSA poll for the BDN (same date and figures); merged into `P2026-0603-surveyusa`. - **MainePulse Aug 2006:** the pollster is not named in the document. Portland Research is a guess. - **AARP 2020:** the pollster is not named in one document; Fabrizio Ward & Hart is assumed from the matching release. - **Data for Progress, 2020 Democratic primary:** the PDF never names Maine. The attribution rests on the listing link. - **Research 2000 (2008–2009 polls):** its data integrity was disputed in the 2010 Daily Kos lawsuit. The polls are kept and flagged. - **Strategic Marketing Services:** this was Pan Atlantic's predecessor (SMS). It is kept as a separate pollster name. ## 6. Scope calls to confirm - **Candidate "likelihood of support" batteries:** Pan Atlantic 2025 and 2026 Governor primaries are kept and flagged; they are not ballot tests. - **Questions asked after messaging:** kept and flagged. - **Replacement-nominee questions (Blueprint, July 2026):** kept and flagged. - **Same-sex-marriage issue questions as proxies for the 2012 ballot question:** kept and flagged (PPP Oct 2011 and Mar 2012, MPRC Apr 2012). - **Generic ballots** (e.g. "Collins vs. a Democrat"): kept and flagged. - **2028 presidential primary preference questions (UNH Oct 2025):** included, since they belong to the 2028 primary. ## 7. Values printed as "*" (under 0.5%) These are stored as 0.4 when the transcriber marked them, or left blank otherwise. Gravis June 2019 prints crosstabs with no total column, so its questions have no figures. ## 8. Truncated archive copies Two Critical Insights PDFs from 2023 were archived truncated at 1 MB and could not be read: the Spring 2023 report and the 2023 fact sheet. The Spring 2023 figures were taken from the full copy at digitalresearch.com (`186bc68bff`). ## 9. Newspaper sources (v0.2) - OCR text can garble digits and mix two columns; transcribers record the exact OCR fragment (`source_quote` in `news/out/*.csv`) and flag every row they had to reconstruct. Rows flagged so far: 1960 Republican Party poll by district (art 46, columns mixed; CD2 undecided 70 does not add), 1960 COPE (AFL-CIO) Smith–Cormier poll (art 69, columns mixed), 1974 SSRI Democratic primary March (art 44, 7 numbers for 6 names), 1974 BDN second poll Erwin 20.5 vs 20.3 in another printing, 1974 BDN final poll sample 776 (BDN) vs 516 (AP). - The Kennedy campaign's 1960 "privately sponsored poll" (50.75–49.25) names no pollster and has no undecided. - Newspaper "polls" that are not polls of Maine voters (student mock elections, reader coupons, legislator/delegate canvasses, one-town surveys, national Gallup) are logged as out of scope, not loaded. - Field dates are often not printed; the end date then defaults to the day before publication (or the printed "N weeks ago" / "last month"). - Batches b0012–b0033 (2026-10-10), rows a second reader should check: art 296 (1974 Mitchell 26.3 read from scrambled OCR); 299 (1971 income-tax poll, "about half" left blank); 321 (shares sum to 100 but story says 23% undecided); 323/325 (Dem undecided 86 vs 36 — 36 used); 344 ("No" printed "4", blank; art 349 gives 44); 409 vs 399 (Telegram undecided 28 vs 38); 432 (OCR fractions 41.5/15.5/43.5); 440 (1992 President may be a subgroup); 484 (Snowe 49 reconstructed); 493 (Mitchell 41.8 misplaced in text); 499 (1976 nuclear question, 12/3.6 look wrong); 552 (undecided digit missing); 595 (Andrews 28 reconstructed, Michael blank); 637 (prison bond "barely 38"/"nearly 47" recorded as 38/47); 649 (retailers' Sunday-sales poll, no date or wording); 673 (1977 tax repeal 46-43-11 — base unclear, nearly half no answer); 716 (Hanrahan column, unverifiable "secret polling"); 737 (1980 bond columns interleaved; 23/21 school energy bond); 791 (only 2nd-District undecided printed); 824 (Andrews 48.9 equals the central-Maine Clinton figure — check the image). - Scope calls left in: the 1990–91 turnpike-widening "would you vote" questions (asked before Question 1 qualified; arts 405, 557, 612) and the 1992 workers' comp initiative poll (art 773 — may never have reached the ballot). Not loaded: the 1990 Capitol News Service turnpike polls printed only as ranges (art 719). - Batches b0034–b0082 (2026-10-10 morning, 577 articles), rows to check: 978 (OCR scrambled near forestry 2A 27); 1053, 1089, 1970/1973 (fragments); 1066 (1981 energy-commission figures quoted by a supporter, sum 101); 1091, 1950 (single figure); 1102, 1133, 1771, 1281 (figures placed from scrambled/jumbled text); 1106, 1138, 1315, 1473, 1493 (pollster inferred from a companion story); 1161 ("6 65" read 65); 1162 (MOE 8 for n=450); 1236 (16 vs 14); 1274/1280 (65 vs 61); 1289 (BDN 55/27/8 odd); 1336 (unclear which poll); 1366 vs 1397/1442 (Lawrence 17 vs 16; sums over 100); 1441 vs 1445 (Nader 14.0 vs 4.0 — 4.0 likely); 1483 (marijuana opposed 38 vs "thirty" — 38 used); 1514 (52.5 vs SMS 53); 1531 (62 vs 65 opposed); 1541 vs 1496 (29 vs 30); 1637 (adds to 105); 1673 vs 1618 (88 vs 83 — 83 likely); 1730 ("25.6 6"); 1780 (sample "2 250"); 1833 (Brennan 33 vs 32); 1978 (RKM Michael figure missing). - Exit-poll or projection rows in these batches (1533 1988 CNS projection; 1765, 1963 exit polls) are flagged by the transcriber; the exit-poll flag in build_final applies when the pollster/label says exit. All-candidate "first choice" primary questions (1679, 1771) are kept as primary questions. - Questions about votes never held (1990–91 turnpike, 1997 same-sex marriage ban, 2002 casino) are kept only where a matching ballot question exists; otherwise logged out of scope. - Pollster names: 1990s "SMS" = Strategic Marketing Services (later merged into Pan Atlantic SMS Group; kept separate from Pan Atlantic Research); "Bullet Poll"/NewsCenter = WCSH-TV/WLBZ-TV; the 1990s "Abacus" is the Forest Ecology Network/Ban Clearcutting pollster, not the Canadian Abacus Data; Lake Snell Perry = Lake Research Partners; McInturff = Public Opinion Strategies; Portland Sunday Telegram = Maine Sunday Telegram. - Batches b0083–b0125 (2026-10-10 midday, 540 articles; mostly repeats of known polls plus 1968 Market Opinion Research for Shute, 1972 BDN 11-county polls, 1976 Cambridge Survey Research for the bottle-bill opponents, 1986 MAINEPOLL, 1996 Moore Information for the NRSC, 1996 DSCC primary match-ups, 2003 racino/casino polls), rows to check: 2039 (Emery "382" blank; 38.2 fits art 2099); 2042 vs 2018 (SMS Governor 44/23/4/2/11 vs AP 50/29/6/3 — likely with/without leaners); 2121 (CD2 "Other" 13% blank); 2139 (sums to 118); 2149 (Brennan/Collins order taken from other copies); 2170 (Maynard 6.5 from 2204); 2190 (Brennan 33 vs 32); 2240, 2297, 2391, 2403, 2612, 2701, 2720 (figures assigned by layout or from garbled text); 2274 (1972 McGovern canvass — percentages don't match counts); 2371 (Hughes ".02"; undecided 46.9 vs 49.1); 2458, 2698 (sum 103/106); 2518 (8.8 vs 8.6; "190%" read 19.0); 2788 (casino Yes garbled); 2841, 2843 (1964/1966 Telegram figures partly lost). Partial-area samples (BDN 1972 11-county polls 2032/2076/2085/2086/2174/2177, 1968 BDN street poll 2372) carry the partial-state flag only where the label says so — check. Exit polls in these batches (1990 Emery workers 2102, 1996 VNS 2869, 2002 NewsinMaine.com) are flagged and kept out of accuracy. - Batches b0126–b0162 (2026-10-10 afternoon, 390 articles; new: 1964 Oliver Quayle poll for the White House/Democrats (from a Drew Pearson column, no dates), 1974 Kelley-campaign and Decision Research primary polls, 1976 Cambridge Survey Research for Muskie, 1996 Martilla & Kiley (McCormick) and Bennett Petts & Blumenthal (Allen) internals, 1996 NRSC primary poll, 1998 SurveyUSA/NewsCenter, 2000 Research 2000 for the BDN, 2002 Public Opinion Strategies for Raye), rows to check: 3064, 3080, 3133, 3307, 3405 (figures placed from scrambled columns; 3405 Clinton 58/Dole 37/undecided 9 sums to 104); 3008 (Emery and bottle bill both 52.9); 3013 (CD1 out of place); 3033 vs 2993 (Tanous 1 vs 10); 3153, 3232 (ranges/"nearly" left blank); 3215 (Cohen "70 cent"); 3246 (two bonds printed as Question 5 — check 2003 numbering); 3325 (30/32 may be party vote); 3495, 3514, 3532, 3032, 3148, 3282 (recalled, claimed or approximate figures); 2913 (undecided "16.61"). - Batches b0163–b0182 (2026-10-10 late afternoon, 213 articles; new: 1970 Oliver Quayle for Gov. Curtis (Muskie v Joly/Tupper), 1964 Telegram CD1/CD2 waves, 1992 Fish and Wildlife Coalition CD2, 1994 GOP activist survey (not a voter sample — check scope, art 3611), 1996 DSCC and Monks tracking, 2002 Christensen McDevitt for Faircloth (CD2 Democratic primary inferred, 3662), SMS 2002 with/without leaners), rows to check: 3573 (relayed by an anonymous caller); 3633 (OCR turns digits to 9s); 3623 vs b0067 (Lawrence 17 vs 16); 3649 vs 3628 (53-21 vs 53-20); 3558 vs 3567 (McKernan 40.4 vs 40.1); 3679 (sums to 109); 3729 (undated partial readings); 3752 (question number not printed — labelled Question 3); 3753 (garbled graphic); 3812/3814 (figures from photo captions); 3837 (Snowe 73–16 vs "by 67"); 3909 (undecided from another copy); 3924 (base unclear). - Batches b0183–b0212 (2026-10-10 evening, 300 articles; new: 1977 Kenneth Hayes poll on the UMaine bond, 1980 Republican and Maine Information Radio Network presidential polls, 1982 Senate — Teeter/Market Opinion Research for Snowe(?) and Tarrance for Emery, Curtis exploratory-committee poll, 1990 Marttila & Kiley and Burningham CD1 primary polls, 1990 Potholm Question 5 land-bond poll), rows to check: 4460 (Teeter poll attributed to "Snowe campaign" for a Senate Democratic primary — check the sponsor; Emery 61 from a pollster's letter); 4461 (47.9 without a candidate name); 4473, 4063, 4065, 4385, 4517 (figures pieced from scrambled columns); 4446 (Republican poll reported second-hand); 4415 (1972 McGovern canvass, not a sample); 4192 vs 4191 (King 55 vs "fifty percent"); 4258 (Potholm 40-17 "of those with an opinion" — check base); 4306 (2A/2B/2C letters inferred); 4312 (may be a CD1 subsample of SMS 4313). - **Pass 2** (from 2026-10-13, Matt: retry the failures and search the light years): 34 yearly searches had been cut short by rate limits, 1977–1989 candidate searches had used "per cent", 1960–1976 ones were capped at 150 pages and 1987–2003 yearly searches had not run — 179 new searches found 22,379 more pages. Batches b0255–b0297 (441 articles, first 940 pages read): new polls include the 1976 BDN final survey (Oct 28), 1978 BDN weekly waves and a Longley hypothetical, the 1978 Maine Broadcasting System governor poll and a leaked Republican internal (Cohen–Emery–Palmer operation), 1982 Northeast Research for the Maine Nuclear Referendum Committee, 1984 ERA campaign polls (1st District only), 1984 and 1986 election-day exit polls, Capitol News Service 1986 post-election, 1988 (June Senate, October House and bonds) and 1989 (April initiative/bond, October referendum) surveys, the 1987 Maine Yankee MAINEPOLL for the BDN, 1996 SMS/Critical Insights ballot questions. Dropped as too thin: questions with only one figure printed (53 rows), subgroup-only figures (6803, 6831), lower bounds (6488), a key-precinct election-day sample (6508), approximate recollections (6812, 6817), a candidate's relayed figures (6826), a fragment (6770). Kept but doubtful: BDN 1976–78 face-to-face shopping-centre surveys (UPI called them unscientific; kept as in pass 1), exit polls (flagged), second-hand campaign tracking (6397, 6500, 6639, 6764), 6711/6447 Republican internal described as a projection, 7072 ERA polls of the 1st District only. 7251 (GOP Senate committee, Brennan 61–28) dated February 1982 as printed. Batches b0298–b0326 (2026-10-14, 289 articles from ~1,550 more pages): mostly reprints; new — 1978 Maine Broadcasting Co. Senate poll, 1980 Maine Information Radio Network President/House and Maine Yankee polls, 1981 Frank Goldsmith poll for the Curtis exploratory committee (1982 Senate), 1982 Atlantic Research (a CMP-owned firm) Maine Yankee poll dated March 1982, 1988 Bean-Jones campaign CD1 primary poll; 7444 (Sept 1980 Maine Yankee, 'poll for the BDN') merged into the SSRI/BDN Sept 4–17 poll and 7678 into the 1987 MAINEPOLL; dropped 7563 (1983 moose-hunt 40–40 with sides unstated), 7277 and the 7678 recall of the June 1986 MAINEPOLL, single-figure questions. Batches b0327–b0352 (2026-10-14, 248 articles from ~700 more pages): mostly reprints of Capitol News Service 1986–1990, SMS/Critical Insights 1996–2003 and BDN polls already on file; new — Capitol News Service 1986 June obscenity question (May 4) and 1988 June bond question, a 1996 Maine Public Broadcasting/Portland Newspapers forestry poll, SMS 1999 land bond (spring), 2001 bonds (Sept 22–29), 2002 Governor (May), 2003 casino/racino waves (Jan 5–8, Apr 24–28), Critical Insights 1999 Question 1 and 2000 Question 5, SurveyUSA's two 2003 casino polls for WCSH/WLBZ, and the 1988 Bean-Jones primary poll now credited to Arthur J. Finkelstein. Recalls merged into the polls they cite (1986 MAINEPOLL Maine Yankee → Northeast Research May 1–26; 1987 BDN 47.5–37.5 → Northeast Research Oct 3–25; 1996 Mason-Dixon Collins–Brennan → Aug 4–10; 1988 CNS UMaine bond → June survey; 1989 spring prison bond → CNS Apr 23–26; 2003 casino/racino recalls → SMS Apr 24–28 and Critical Insights Sept 13–28). Dropped (`news/cleanup_pass2_b0327-b0352.csv`): a 1994 CNS survey of Republican convention delegates (not voters), the 1986 measured-service question (only No and undecided printed, three copies), a single-figure 1988 bond result, and a 1990 gun-safety initiative that never reached the ballot (two copies). Batches b0353–b0384 (2026-10-14 midday, 335 articles from ~4,000 more pages): mostly reprints; new — a December 1975 Cohen campaign poll (1976 Senate, 47–45 over Muskie), a February 1980 SSRI/BDN Maine Yankee poll, a late-1981 Emery–Mitchell poll reported by the BDN (pollster unnamed), the Tierney campaign's March 1986 governor poll, the first (Oct 4–11) wave of the 1986 Northeast Research/BDN poll (governor and measured-service questions), and a Sept 18–19 1996 BDN forestry poll; added to polls on file — the 1976 BDN October survey (President, Senate, bottle bill, with and without leaners), the 1981 Market Opinion Research (Snowe) and Tarrance (Emery) match-ups, the 1981 Goldsmith/Curtis questions, CNS June and October 1988 President, House and bond questions, 1990 CNS bonds, Mason-Dixon 1996 forestry by district, 2000 Research 2000/BDN primary and Bradley–Bush questions, SMS October 1997 gay-rights repeal. Dropped (`news/cleanup_pass2_b0353-b0384.csv`): the 2002 BDN governor poll and the Mason-Dixon Oct 1996 Senate question (main figures unreadable), a 1980 CD2 poll with one figure, a partial copy of the 1986 Northeast Research first-choice list, a 1992 Gannett poll whose printed figures conflict ("Bush and Clinton each had 21 percent"), and preliminary half-sample figures of the Critical Insights Sept 2003 poll. 1975 Cohen poll dated mid-December from "early December". Batches b0385–b0408 (2026-10-14 afternoon, 240 articles from ~3,000 more pages): almost all reprints. New polls: a February 1980 BDN Republican presidential preference poll (Bush 29.5, Reagan 23.5, Baker 13.7; Connally's figure lost in the scan) and an unnamed October 1995 poll on Question 1 (gay rights) quoted by Maine Won't Discriminate (40–40). Added to polls on file: the 1981 Market Opinion Research (Snowe) primary questions (Emery 47–Snowe 45; Curtis 57–Mitchell 35 — the Curtis–Mitchell line sits in run-together columns and is credited to the Snowe poll from context), Tarrance (Emery) Republican primary 45–41, the DSCC 1996 Senate poll's Brennan primary and general match-ups (its two dated copies, Feb 4 and Feb 9, now one poll, P1996-0205), a BDN 1980 Maine Yankee copy credited to SSRI's Sept 4–17 poll, the 1986 MAINEPOLL and Northeast Research October figures, CNS 1988/1989 copies, SMS and Market Decisions 1996 forestry copies. Pre-2004 ids changed: P1981-0509-market-opinion-research → P1981-0506, P1996-0204/0209-unnamed-pollster-for-dem → P1996-0205. Dropped (`news/cleanup_pass2_b0385-b0408.csv`): a one-sided 1997 Forest Compact copy and a 1996 forestry fragment with the main figure on a missing page. Batches b0409–b0503 (2026-10-15, the last 441 articles; every page of the 179 pass-2 searches read — **pass 2 finished**): mostly reprints; on v0.3 they add 10 polls and 41 questions to polls on file (see README v0.3.2 and §15). Dropped (`news/cleanup_pass2_b0409-b0503.csv`): margin-only figures, party/region/Republican-voter subgroups, undecided-only and other single-figure questions, scrambled or mixed fragments, reprints of polls on file, a general nuclear-ban question and recalled polls that cannot be dated; copies re-dated to the polls they reprint. - Batch b0254 (2026-10-11, last batch, 9 articles): no polls kept — a 1985 rail-line bond 'Yes 50' quoted second-hand from an unnamed TV poll (6148) and a 1965 amateur survey of ~280 Democrats and independents for 1966 Governor (6182) were dropped as too thin. **The newspaper sweep is finished** (all 18,664 queued pages; 101 page requests failed and were not retried). - Batches b0245–b0253 (2026-10-11 morning, 94 articles; mostly ballot questions: 1971 income-tax repeal (Oliver Quayle for FAIR, figures printed as approximate), 1973 Power Authority of Maine, lottery and parks bond (BDN survey Oct 14–24; Becker Research for Central Maine Power, dated June 1973 as printed in art 5731), 1977 uniform property tax repeal (Hayes, merged with his UMaine-bond poll at art 4455), 1979 bottle-law repeal (Maine Broadcasting System, Feb 1979) and 1979 bonds (Ad-Media for the Maine Good Roads Association, Oct 15–17; an unnamed Oct 1979 fish-pier bond poll), 1981 Question 1 elected Energy Commission (an unnamed spring 1981 poll for Gov. Brennan, quoted secondhand by a proponent), and a 1976 Senate Cohen v Muskie 46–46 Cohen internal that his office would not confirm. Dropped: 6098 (undecided share only), 5884 (undecided only), 5854 (a 1977 recollection of a 1971 poll), the 5713 recall of the Becker poll (kept at 5731), the 5864 reprint of 4455. Doubtful: 5973 bottle-law 'retain' printed only as two-thirds (blank); 6047 Snowe v Curtis and Emery v Curtis garbled (blank). 'Ad Media' spellings and 'Maine Broadcasting System' now map to Ad-Media (Augusta) and WCSH-TV / WLBZ-TV. - Batches b0237–b0244 (2026-10-11 early morning, 86 articles; new: the 1980 UMaine SSRI/BDN caucus poll (Jan 7–Feb 5, Democrats and Republicans) with a Maine Yankee question, 1974 Northeast Markets for Erwin, 1981 Tarrance for Emery (field end = the printed disclosure date May 5, 1981), 1982 Republican governor primary polls by Wirthlin for Pierce and Cromer-Young for Huber, and three 1986 referendum polls: Weil and Firth for two bookstore chains (obscenity ban), Cambridge Reports for Maine Yankee / Central Maine Power (Maine Yankee). The May 1–26, 1986 MAINEPOLL nuclear question ("prohibit nuclear power generation in Maine") and Cambridge Reports are filed under the November 1987 Maine Yankee referendum as hypothetical; both are interested-party or general-wording polls. Removed as reprints: art 1773 (b0073, an October 1987 story that recalled these 1986 figures as a 1987 poll), 5490/5495 (b0243) and 5373 (b0240). Doubtful: 5344 Redmond/Leighton/Ricci figures may be attached to the wrong names (scrambled OCR); 5258 Snowe's Market Opinion Research 57–36 may have undecideds assigned; 5409 Dole, Anderson and Crane printed only as under 1 percent (blank). "Cromer-Young Group" now has its own pollster name (it had been read as Harris because of "Harrisburg"). - Batches b0213–b0236 (2026-10-11, 260 articles; new: 1964 and 1966 Maine Sunday Telegram and unnamed polls, 1976 DNC poll for Carter, 1978 Bangor Daily News weekly polls #1–#4 plus Maine Broadcasting System and the "Emery-Cohen" governor poll, 1980 Maine Yankee referendum polls (Maine Information Radio Network; UMaine SSRI for the BDN), 1980 President and House (SSRI/BDN, MIRN), 1982 Milk Commission and Maine Yankee referendum polls, 1982 Senate and Governor tracking (Command Research for Emery, GOP Senate Committee, Democratic and Cragin tracking), 1986 Governor (Marttila & Kiley for Tierney, a pollster read as "Cromer" for Huber, Northeast Research for the BDN), 1990 Snowe internal, 1992 Guy Gannett polls). Doubtful: OCR gaps leave blank figures in articles 4523, 4544, 4597, 4926, 4935, 4943, 4947, 5034; 4754 prints Emery at both 42.23 and 44.27 (44.27 kept because the row sums to 100); 4762 Democratic tracking is two-way; 4772 Command Research 65–35 may already split undecideds; 4869 the "Cromer" name and Menario's 9 are uncertain; 4529 is Republicans only; 4573, 5097 are campaign polls quoted as one candidate's share; 4594, 4597 unnamed pollsters; "UMaine Orono researchers" for the BDN in 1980 are treated as SSRI. - Capitol News Service 1990: one late-October wave was reported in many stories over a week; it stays one poll (Oct 28–30) with ~19 questions, some of them restatements of the same race in different stories. ## 10. Accuracy table (v0.2) `poll_accuracy.csv` compares each poll that ended within 21 days of an election with the certified result (election results v3.10, first-round figures for ranked-choice races). The error is on the **two-way margin** between the two leading finishers (undecided and minor candidates removed from both poll and result). Polls of a district sub-sample of a statewide race, hypothetical match-ups and flagged questions (generic, post-messaging, batteries, replacement nominee, proxy, ranked-choice items) are left out. With few polls per pollster, the averages are rough. ## 11. Hand-transcribed crosstabs (v0.2) - Many Pan Atlantic, Critical Insights and Fabrizio charts print only the leading candidates (or only "undecided"/"Yes") per group, so those groups do not add to 100; each such cell says "partial" in `notes`. Some cells come from report text bullets rather than tables (noted). - Congressional-district columns are recorded as a group of the statewide question even where the same figures are also on file as separate CD questions — do not add the two together. - Printed slips kept as printed and noted: Pan Atlantic Oct 2022 Governor "> 4-year degree" Mills 60 / LePage 44 (adds to 112 with undecided); SurveyUSA 2004 lowest-income banner printed ">$40,000" (read as under $40,000); Pan Atlantic Sept 2024 p.25 chart titled "by Income" shows age bands. - Questions on file that the linked document does not contain (check their listing source): Pan Atlantic Dec 2025 and May 2026 CD2 general "Joe Baldacci v Paul LePage" (listed on Wikipedia; not in the linked PDF — probably a separate CD2 release), SurveyUSA Oct 2004 House races in one release, NYT/Siena Sept 2020 Q5 (44/40 vs the table's 49/44), Siena 2018 CD2 Q1 (39.89/46.4 matches no table). ## 12. Data Audit II fixes (v0.3, 2026-10-14) All fixes are applied by `scripts/apply_corrections.py` (last step of `run_all.sh`) from `corrections/audit_2026-10-14.csv` plus general rules; see `corrections/README.md`. Every removed record is in `data/poll_reprints.csv`. **Done** - PC-1 / PC-3 listing copies removed (see §1). PC-2: Pan Atlantic March 2026 CD1/CD2 Senate-primary figures removed. The March report (68th Omnibus) repeats the December 2025 sentence word for word ("Mills leads Platner 54%-23% … in the 2nd Congressional District, while Platner leads Mills 45%-43% in CD-1"); March's own statewide result (Platner 46 / Mills 39) cannot come from those district figures. No March district figures are printed. - PC-4 reprints and recalls: 105 questions merged into the earliest printing; 106 in-poll duplicates merged; 45 poll records left empty were removed as reprints (Northeast Markets 1974, CNS 1990, RKM 2000, SMS 2000, Guy Gannett 1992, BDN 1972/1976/1978, MST 1964/1966, SSRI 1974/1980, Market Decisions 1998, and others). The 1976 BDN weekly waves were sorted into their own polls (Oct 9 wave = `P1976-1009`, Oct 16 wave = `P1976-1015`, Oct 23 wave = `P1976-1026`). - PC-5: 27 subgroup questions flagged; the 1972 BDN 11-county surveys flagged `partial-state sample`. - PC-6: ranked-choice round rows removed from 20 questions; SurveyUSA June 2026 round-3 question no longer scored; the empty Mellman "RCV round 4" record removed. - PC-7: stray options dropped where a figure from another question had been merged in (SMS Oct 2000, Sun-Journal/Critical Insights Oct 2000, the 1976 three-wave table, GQR 2009 and Pan Atlantic 2025 strength components). 19 questions that still add to 106–147% are flagged `figures do not add up`. - PC-8: post-election recaps re-dated to the poll they repeat (MST 1964, SMS 2002 recap of Pan Atlantic, MST 1974, BDN 1998); three post-election campaign recalls flagged; UNH July 2026 convention question unlinked and flagged replacement-nominee; UNH Oct 2025 CD2 → 2026; FAU 2024 → Republican primary; Colby's 11-month RCP row merged into the July 2020 poll. - PC-9 pollster names (see README). Wrong aliases removed: Tarrance (Market Opinion Research/Snowe), Hobbins (McKernan staff), Tanous (Times Record). - Hole: ballot-question polls linked to referendum results (348 of 384). - PC-11: single-figure fragments dropped (154); poll `n_questions`, `max_sample_n` and `races` recomputed. **Still open (not changed; needs a source or Matt)** - Possible misattributed duplicates between *different named* pollsters, left as they are: Critical Insights Sept 1996 vs SMS Sept 1996 forestry question (45/14/18/23, same BDN issue); "Maine Citizens Campaign" Sept 1996 vs SMS Oct 1996 forestry figures (a summary box may have mislabelled a row); SMS vs Critical Insights 1999 and RKM/SMS/Critical Insights 2000 referendum figures that match exactly; GQR Sept and Oct 2014 Governor figures identical (40/42/15/3). Each needs the original article or report. - `P1964-0818-unnamed-pollster-for-por-Q1` (Johnson 79 / Goldwater 12 / Not sure 18): probably the Quayle poll (70-12-18) with an OCR error; flagged, not changed. - 1976 BDN Senate figures that mix two waves (flagged). `P1990-0513-foster-burningham-Q1` mixes Democratic and Republican primary names (flagged). - PC-10 (population labels, `hypothetical` blanks and the "undetermined" value) is not yet normalized. - Poll type of Mason-Dixon Oct 1996 shows "Campaign internal" (a Longley campaign mention in a merged article); should be Media-sponsored. Not changed. - Florida Atlantic Mar 2024 Republican primary poll has n = 18 and is now scored for accuracy; consider excluding very small samples. - 1990–1994 referendum questions cannot be linked until those ballot questions are in the election results. - New newspaper batches (pass 2) are covered by the general rules, but hand actions name specific ids: if a rebuild changes an id, `apply_log.txt` prints a warning. ## 13. Remaining links and the weighted average (P-151 a, d; 2026-10-15) - Every question that can be linked now is. Hand links and "no contest" reasons are in `corrections/links_2026-10-15.csv` (61 rows, audit id P-151d), so pass-2 rebuilds keep them. `poll_question_links.csv`: 1,876 linked; 132 for 2026/2028 contests not yet held; 30 "no results on file"; 14 caucus years; the rest have a stated reason (generic party ballot, all-party first choice, a question that never reached the ballot, a party-only subgroup). - Newly hand-linked: 1973 state lottery (Bangor Daily News, Oct 1973), 1989 Question 1 campaign-spending limit (Capitol News Service, Apr 1989), 2014 Question 1 bear baiting (Ipsos, Oct 2014). Their labels are set in the same corrections file. - Still unlinkable: the 1990–1994 ballot questions (26 questions) and the June 1990 House primary (4) — not in election results v3.12 (PENDING-WORK P-178). Presidential primary questions in caucus years (1980, 2004, 2008, 2012, 2016) have no primary vote to compare. - The page's polling average is now weighted (grade × √sample × half for campaign/party/partisan/advocacy polls; see `site-design/polling-page/README.md`). Because of this, the Mason-Dixon Oct 1996 poll mislabeled "Campaign internal" (§12) now counts half; fixing its type would give it full weight. ## 14. Official totals for ranked-choice races (2026-10-15, results v3.14) - Accuracy for the 11 ranked-choice contests with an official round-one tabulation is now measured against that count (`link_results.py`); the town first-choice figures add up to slightly less. A candidate the official count does not list keeps the town sum. Polling data v0.3.1; nothing else in `data/` changed. ## 15. Newspaper pass 2, last batches (v0.3.2, 2026-10-15) - Built on v0.3.1 (the v0.3 pipeline reproduces the committed v0.3.1 `data/` byte for byte before the new batches are added); every question id named in `corrections/` keeps its meaning (checked: 103 corrected question ids, 2,054 shared question ids unchanged). - **Two old transcription errors fixed:** b0019 art 493 had the Tierney campaign's January 1990 primary figures (53–26–10) mixed into the Mitchell campaign's May 1990 poll question (now split with an `[earlier poll: …]` label; the May poll reads 24.1–22.4–19.7–5.7–0–27.6); b0038 art 945 had the Sept. 27, 1995 Bullet Poll wave written over the Oct. 26–27 wave (now split). - **Pre-2004 poll ids changed:** P1995-1030-wcsh-… → P1995-1027 (true dates Oct 26–27) plus P1995-0927 (the Sept. 27 wave); the Nov. 1995 Bullet Poll (48–41) sits with P1995-1102. P1996-0810-mason-dixon-… → P1996-0808 (same questions; the fragment that had set the Aug 10 end date is now a reprint). P1990-1102-unnamed-pollster-for-ken (1990 land bond 40–17) → P1990-1009-unnamed-pollster-for-eve (the earlier printing). P2002-0803-critical-insights-digita was a misdated reprint of the March 2002 Critical Insights poll → P2002-0328. No id after 2003 changed. - RKM Research's spring 2000 poll (P2000-0509) now carries its printed field dates, Apr 25–27 (`corrections/pass2_final_2026-10-15.csv`), and its President and gay-rights questions. - The 1978 "Frankland v Palmer" question picks up the `ranked-choice item` flag from the PC-6 rule (the name contains "rank"); harmless (it only keeps the question out of the accuracy table, where a two-candidate fragment would not belong anyway). - 1990 polled ballot questions still show "no ballot questions on file for that year" until P-178 loads 1990–1994 results. ## 16. Party crosstabs added in v0.4 (2026-10-16) - The 2,653 cells come from `model-data/polls/scripts/build_crosstab_additions.py` (parsers and hand transcriptions; every source document in `model-data/polls/sources/`). Its FLAGS (model-data FLAGS §11–18) apply: partial tables (Portland Research Aug 2006, Research 2000 Sept 2009, Siena Oct 2018 CD 2, PPP/Kleban Sept 2025, Emerson Sept/Oct 2016) do not sum to 100; SurveyUSA group n's are approximate (composition × sample); Emerson n's are weighted counts. - Group dimensions say what the party is: `Party registration (voter file)` (NYT/Siena), `Registered party (self-reported)` / `Party registration (self-reported)` (UNH, SurveyUSA Aug 2021, Magellan), `Party affiliation (registered)` (Critical Insights 2018), otherwise party identification. "Independent" by party ID is not the unenrolled. - 469 crosstab option labels were renamed to the question's own label so pages can line them up (printed label in `notes`); options like "Other", "Undecided", "Someone else" keep the printed wording. - Toplines of the 24 new questions: UNH from each report's STATEWIDE row for the poll's month (the trend tables print several waves); Emerson Sept 2022 generic ballot computed from its three party groups weighted by their counts (they cover the whole sample). - The SurveyUSA July 2006 TABOR question read the ballot summary (Yes 71 / No 25); the October and November waves asked "certain yes / certain no / not certain" — not comparable with July, all three kept as printed. ## 17. Found while building v0.4 — page rules changed, data left as it is (2026-10-16) - `build_final.py` flags `ranked-choice item` whenever a listing's context matches "rank|RCV|first choice|…", so 197 questions carry it, most of them ordinary first-choice ballots of 2018–2024 (and "Frankland" 1978). The Historical Polling page had skipped them all, which left the 2020 President race drawn from CD 1 / CD 2 subsamples. The page now skips only later-round / second-choice items (`site-design/polling-page/build_polling_data.py`); the flag itself should be cleaned in a corrections pass (PENDING-WORK). - 228 District 1 / District 2 subsamples of statewide races are linked to the statewide contest (correct for links; `link_results.py` already keeps them out of accuracy). The page now leaves them out of statewide charts. - `hypothetical = yes` is set on some real match-ups whose names were printed as "Joe Biden, the Democrat" or with "minor party candidates" (Colby, Fabrizio, Siena 2020; Pan Atlantic May 2022). The page now keeps such a question when every candidate in it was on that election's ballot; the field itself is unchanged. - `subgroup only` questions are now left off the page too (one, SMS Oct 2003 southern-Maine casino subsample, had been charted as statewide). ## 18. Nominees printed under another form of their name; listing "hypothetical" flags (2026-10-17, Matt) - Matt (2026-10-17): "Robert Charles is obviously Bobby Charles … same race, same last name, Robert is the full name of Bobby." Standing rule: a candidate printed under his full name, a short form or a quoted nickname in a poll of the same race is the same candidate — no need to ask. Matching = same office/district, same surname, first name equal or a known short form (Robert/Bobby, Richard/Rick, Matthew/Matt …) or the quoted nickname; checked against the SOS ballot list for upcoming races and the results' candidates for decided ones (`link_results.py` already matches by surname). - The listings' `hypothetical` flag was wrong for nine 2026 nominee match-ups (v0.4.2, `corrections/nominees_2026-10-17.csv`). Re-run `corrections/make_nominees_2026-10-17.py` (it writes the same file) after adding new 2026 polls if they come from listings; add a dated copy for 2028. - The Historical Polling page applies the same rule itself for upcoming generals (`build_polling_data.py` `nominee()`), so a poll is never left off a chart for a spelling: UNH Jul 15–20 ("Robert Charles") is now on the 2026 Governor chart. ## P-178 links (2026-10-22): the 1990, 1991 and 1994 ballot questions and the June 1990 CD 1 Democratic primary - Election results v3.18 now carries these contests, so 24 poll questions that were "no results on file" are linked (`corrections/links_2026-10-22_p178.csv`, applied to `data/poll_question_links.csv` / `data/poll_option_links.csv` by `corrections/relink_p178.py`): 1990 Questions 1-8 (15 questions), 1991 Question 1 (4), 1994 Question 1 (1), the 1990 CD 1 Democratic primary (4). The Historical Polling page gains 11 races. - **Turnpike polls turned round [Certain]:** the two Capitol News Service polls of 1990 asked about *widening* the Turnpike ("Oppose widening" / "Support widening" / "Favor widening"); Question 1 of 1991 was the initiative to *stop* it, so "oppose widening" is entered as Yes. Done with option links (candidate_display Yes / No). `apply_corrections.py` has no action for option links: **on the next full polling rebuild check that `link_results.py` keeps these three option links** (or add them by hand again). - **Left unlinked:** the CNS October 1990 CD 1 / CD 2 subsamples of Question 6 (Q13, Q14 — subgroups, not statewide), and the Becker Institute's January 1992 workers' compensation question (the November 1992 questions are not loaded). - **Possible reprint [Likely]:** Marttila & Kiley (Jan. 4, 1990) and Foster Burningham (May 13, 1990) print the same Tierney 53 / Mitchell 26 / Andrews 10 for the CD 1 primary; one may be a newspaper recall of the other. Both kept; not checked against the documents. - Northeast Research (June 1990) asked about letting *any* store open on Sunday; the initiative covered stores over 5,000 sq ft. Linked to Question 1 with that note. - `poll_accuracy.csv` / `pollster_maine_accuracy.csv` (the "How the polls did" scorecard) were not recomputed: they come from the polling rebuild (`link_results.py`), which this cloud session cannot run.