SD-44 Early Voting in Review

2026 Democratic Special Primary · Briefing date 2026-05-16 · data from 2026-05-16 snapshot, returns through 2026-05-15 (final EV day). Generated 2026-05-16 12:58 GMT. Internal team brief.
Tier 1 — Cover
23,303
DEM EV ballots in SD-44
2.27× the 2024 EV total of 10,278
5,883
Ballots on the final Friday
25.2% of cycle in one day · 2024 final-Friday was 22.4%
97.9%
In-person
1.9% mail · 0.2% electronic
22,221
P78 (high-propensity) not yet voted
Prime E-Day GOTV target · 41% of P78 voters already banked their ballot
Bottom line: SD-44 closed EV at 2.27× the 2024 DEM EV volume, materially more concentrated in DeKalb-intown precincts (HD-089 +1.7pp vs Clayton HD-076 −3.1pp), and notably younger and less 65+ Black than 2024 — and Friday's close amplified all three drifts.
Tier 2 — Short report

1 · Daily

Final-day surge

Friday closed early voting with the biggest single day of the cycle by an enormous margin — 5,883 ballots, nearly twice Thursday's record (2,954) and roughly 2.55× the equivalent final-Friday in 2024 (2,303). The Mon–Fri 5/11–5/15 stretch extended the cycle's record run that began the prior Monday to five consecutive new highs.

Friday's surge alone produced more ballots than either the first week (4,354) or the second week (5,428) of EV did in its entirety. The final week as a whole accounted for 58% of the entire cycle in five days, and Friday alone delivered 25% of the cycle in a single day.

The final EV total of 23,303 DEM ballots ran 2.27× the entire 2024 SD-44 DEM EV window (10,278), with in-person ballots accounting for 97.9% of the cycle and mail returns staying small throughout. SD-44 finished EV running sharply ahead of its 2024 day-for-day pace at every comparable point in the calendar.

2026-05-16T13:24:33.886405 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-05-16T13:24:33.932301 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
WeekDates (2026)2026 ballotsDates (2024)2024 ballotsRatio
Week 14/27–5/34,3544/29–5/52,3101.88×
Week 25/4–5/105,4285/6–5/122,4082.25×
Final week (Mon–Fri)5/11–5/1513,5215/13–5/175,5602.43×
EV cycle total19 days23,30319 days10,2782.27×

2 · Geography

County and HD composition

Friday's geographic mix tilted further toward DeKalb, with DeKalb's share of the day climbing past 76% against a cycle-to-date mix of 74% / 26% Clayton. DeKalb grew 2.37× from 2024 in absolute terms; Clayton grew 2.02×.

Within DeKalb, HD-090 anchored the close as it has all cycle, posting 39% of Friday's ballots, with HD-089 catching HD-076 in a dead heat for the second slot at roughly 22% each. Compared to the equivalent 2024 cycle on the same boundary, the big structural shift is in HD-076 — Clayton's share has fallen 3.1pp (27.0% → 23.9%), with HD-089 picking up most of the slack (19.3% → 21.0%) and HD-090 absorbing the rest (36.3% → 37.2%).

HD-090 finished the cycle as the single largest HD within SD-44; HD-089 closed essentially tied with HD-076 for second on a cumulative basis after trailing it for most of the cycle. Candler Park finished as the single biggest precinct in the district.

2026-05-16T13:24:33.969446 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-05-16T13:24:34.342899 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
County2024 cycle %2026 cycle %2026 Friday %Growth ratio
DeKalb70.7%74.0%76.1%2.37×
Clayton29.3%26.0%23.9%2.02×
HD2024 cycle %2026 cycle %2026 Friday %2024→2026 shift
HD-07627.0%23.9%21.8%-3.1pp
HD-0782.2%2.1%2.1%-0.1pp
HD-0848.5%9.2%8.5%+0.7pp
HD-08919.3%21.0%21.8%+1.7pp
HD-09036.3%37.2%39.4%+0.9pp
HD-1166.6%6.6%6.4%-0.1pp

3 · Within-cycle shape

Where the final-week shares moved

The cycle-long arc that the daily briefs had been tracking — HD-090 climbing through EV and the intown DeKalb piece (HD-089) accelerating in the back half — held all the way through close. HD-090's share of daily turnout sat in the high thirties for the final week, and HD-089 matched its highest single-day share of the cycle on Friday.

The intown surge stayed concentrated in the same handful of precincts that had been moving share through the week. Boulevard finished as the cycle's single biggest share-mover; Metropolitan, Burgess Elem, McNair High, Coan Recreation Center, Oakhurst, and Candler Park all over-pulled on Friday relative to their own cycle averages.

Friday also pulled the Winnona Park / East Lake / Druid Hills High corridor up — those precincts had been mixed earlier in the week but moved decisively higher on close. The 2024 cycle showed the same general HD-090 ramp through EV, but the intown DeKalb piece arrived with materially more force in 2026 and was even more concentrated on the final day.

2026-05-16T13:24:34.173271 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
Friday over-performers (Δ Friday share − cycle share)Friday under-performers
PrecinctΔ ppPrecinctΔ pp
Metropolitan (Atl) (090)+1.02Ellenwood 2 (076)-0.66
Burgess Elem (Atl) (090)+0.77Morrow 5 (076)-0.54
Boulevard (Atl) (089)+0.61Glennwood (Dec) (084)-0.48
Mcnair High (090)+0.57Cedar Grove Elem (116)-0.42
Oakhurst (Dec) (089)+0.50Clairemont West (Dec) (084)-0.35
Winnona Park (Dec) (084)+0.42Meadowview (090)-0.31
Candler Park (Atl) (090)+0.42Fernbank Elem (089)-0.29
Coan Recreation Center (Atl) (090)+0.40Flat Shoals (090)-0.29

4 · 2024 vs 2026 within precincts

Cumulative precinct-level deltas

Compared to the equivalent point in the 2024 calendar, the precinct map remained where the daily briefs had it. The biggest precincts running ahead of their 2024 pace are clustered in intown DeKalb HD-089 (Boulevard, East Lake, Coan Recreation Center, Burgess Elem) and the HD-090 corridor immediately north and east of I-285.

The precincts running behind their 2024 pace cluster in Clayton — both Ellenwoods, several of the Morrow precincts, and the Flat Shoals corridor — along with the south-DeKalb Cedar Grove pair. The single biggest cycle-share loser is Candler Park itself, down 1.27pp from its 2024 share even though its absolute ballot count nearly doubled — the rest of the district simply grew faster.

Friday's relative-share movement reinforced rather than rotated this pattern: the Friday over-performers vs cycle pace were almost entirely the intown HD-090 / HD-089 set, and the Friday under-performers were the same Clayton and eastern-DeKalb precincts that had been lagging all cycle.

One nuance that survived to close: Druid Hills High — an older, more affluent HD-089 precinct — moved up on Friday after lagging earlier in the week, narrowing but not erasing the intown-composition story.

2026-05-16T13:24:34.095742 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 2026-05-16T13:24:34.432591 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
Cycle-share gainers vs 2024 (pp)Cycle-share losers vs 2024 (pp)
PrecinctΔ ppPrecinctΔ pp
Coan Recreation Center (Atl) (090)+0.97Candler Park (Atl) (090)-1.27
Metropolitan (Atl) (090)+0.94Flat Shoals (090)-0.82
Boulevard (Atl) (089)+0.80Morrow 5 (076)-0.77
Mcnair High (090)+0.80Ellenwood 1 (076)-0.71
Burgess Elem (Atl) (090)+0.63Cedar Grove Elem (116)-0.57
Oakhurst (Dec) (089)+0.54Kelley Lake Elem (089)-0.55
Cedar Grove South (116)+0.51Morrow 10 (076)-0.39
East Lake (Atl) (089)+0.50Meadowview (090)-0.39

5 · Demographics

Race and age divergence from 2024

Friday's returns sharpened the 2026-vs-2024 demographic divergence to its widest point of the cycle. The 65+ share of Friday's ballots collapsed to 20%, against a cycle-to-date 65+ share of 36% and a 2024 cycle-to-date 65+ share of 50% — close to half.

The 30–44 bucket surged to 29% of Friday's ballots — well above its own cycle pace (19%) and 2.6× the 2024 equivalent of 11%. The 18–29 share ticked up to 10% on Friday, against 7% cycle and 4% in 2024.

The race mix shifted in the same direction. Black share dropped about 3.5pp on Friday — most of that flowed to White (+1.1pp Friday vs cycle) but a non-trivial slice went to Other (+0.8pp) and Hispanic (+0.6pp). District-wide: Black 50.1% Friday vs 53.6% cycle vs 57.6% in 2024; White 35.8% Friday vs 34.8% cycle vs 32.2% in 2024. Hispanic and Asian shares each ran 2–3% on Friday, both notably above their roughly 1% 2024 marks.

The overall picture at close is an electorate that came in younger, less Black, and notably less concentrated in 65+ Black voters than 2024 — and Friday's late surge amplified all three drifts.

2026-05-16T13:24:34.001595 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
Race2024 cycle2026 cycle2026 FridayCycle shiftFriday shift
BLACK57.6%53.6%50.1%-4.1pp-7.5pp
WHITE32.2%34.8%35.8%+2.6pp+3.6pp
HISPANIC1.0%1.8%2.4%+0.8pp+1.4pp
ASIAN1.1%2.0%2.5%+0.8pp+1.4pp
OT7.0%6.5%7.3%-0.5pp+0.3pp
AI0.4%0.7%1.0%+0.2pp+0.5pp
UN0.2%0.6%0.8%+0.4pp+0.6pp
NA0.4%0.1%0.1%-0.3pp-0.3pp
2026-05-16T13:24:34.030616 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
Age2024 cycle2026 cycle2026 FridayCycle shiftFriday shift
18-294.1%7.4%10.5%+3.3pp+6.3pp
30-4411.2%19.5%28.9%+8.3pp+17.7pp
45-6434.4%36.9%40.5%+2.5pp+6.2pp
65+49.9%36.1%20.0%-13.8pp-29.9pp
NA0.4%0.1%0.1%-0.3pp-0.3pp

6 · Projection summary

2024 baseline + 2024 EV-vs-E-Day empirical anchor

The standing uniform-swing model in analysis/sd-44-2026-05-16.xlsx inherits its baseline from the 2024 SD-44 Dem primary: Parent 73.4% / Thomas 26.6% across the 49 SD-44 precincts. Saira Draper inherits Parent's per-precinct share; Thomas inherits her own; a single district-wide swing parameter shifts every precinct's Thomas share uniformly. The §8 detailed analysis validates the uniform-swing assumption — 2024 precinct-level Parent shares were nearly identical in EV vs Election Day (best fit: E-Day = 0.94 × EV + 1.5).

The largest within-cycle Thomas-share movement the 2024 data showed was the 5.1pp EV→E-Day shift district-wide (Thomas 34.8% in EV vs 29.7% on E-Day). That's the empirical yardstick for sizing 2026 scenarios. The §3.7 detailed analysis runs three: A · 0pp swing, B · +5pp swing (= one empirical magnitude), and D · +10pp swing (= two empirical magnitudes). In all three, Saira wins the total election:

ScenarioThomas swingEV marginTotal margin (with ~14,000 E-Day)
A · Inherited margin0pp+9,786 Saira+15,459 Saira
B · Method-shift magnitude+5pp+7,456 Saira+11,729 Saira
D · Double magnitude+10pp+5,125 Saira+7,999 Saira

The race ties (EV-only) at a +21.0pp Thomas swing — about 4.1× the 2024 empirical method-shift magnitude. Beyond ~2× empirical, the swing requires invoking candidate-specific dynamics not observable in 2024 data. See §3.7 for the full sensitivity chart, Tuesday signals to watch, and caveats.

7 · GOTV universe summary

Who's still out there for Tuesday

SD-44 has 151,885 total registered voters across the 49 precincts. As of EV close, 23,303 (15.3%) have voted in 2026 — leaving 128,582 registered voters available for Tuesday mobilization. The propensity breakout decides yield:

Propensity binRegisteredVoted 2026 EVTurnout to dateNot yet voted
P78 (high — most reliable)37,44515,22440.7%22,221
P56 (med-high)23,4274,69120.0%18,736
P34 + P12 + P0 (med-low and lower)91,0133,3883.7%87,625

The 22,221 P78 not-yet-voted bar is the highest-yield Tuesday target. 2026 EV alone is 92% of the entire 2024 Dem primary turnout (25,234 voters in 2024) — so the GOTV question is not "can we catch 2024" (already done) but "how much higher can E-Day push us." See §9 for per-HD universe sizes, the not-yet-voted map, and caveats around precinct-aggregate sizing (no per-voter joins available locally).

Tier 3 — Detailed report

3.1 · Cycle pace, expanded

The day-by-day chart in §1 shows two distinct cycles. Both started with a strong opening Monday — 2026 at 804 and 2024 at 490 — and both sagged through Tue/Wed before the end-of-week Friday lift. From there, the cycles diverged: 2024 grew steadily through the EV window; 2026 added a back-loaded super-spike in the final week that has no 2024 analog.

Cumulatively, 2026 was already ahead of 2024 on day 1 (the cumulative curve in §1 never crosses), and 2026 surpassed the entire 2024 EV cycle total by day 15 of EV — Monday May 11 — with four EV days still to run. By close, 2026 had run 2.27× the full 2024 EV total.

The back-loading is real and large. In 2024, the final five days of EV delivered 54% of cycle volume; in 2026, the final five days delivered 58%. The roughly +4pp shift is mostly the Thursday and Friday spikes (Thu 2,954, Fri 5,883) — both individually larger than any single day in 2024.

Method of voting was strikingly stable: in-person ballots accounted for 97.9% of the cycle, mail 1.9%, and electronic 0.2%. The mail piece was concentrated in the first half of EV — Friday returned just 6 mail ballots against 5,875 in-person.

2026-05-16T13:24:34.250461 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

3.2 · County structure, expanded

DeKalb finished EV at 17,238 DEM ballots (74.0% of SD-44), 2.37× its 2024 total of 7,269. Clayton finished at 6,065 (26.0%), 2.02× its 2024 total of 3,009. Both counties grew substantially in absolute terms, but DeKalb grew faster — and the gap is what shows up as the +3.2pp DeKalb-share shift between cycles.

The within-DeKalb story is even more concentrated. HD-089 (intown DeKalb, extending from Druid Hills through East Lake into Decatur and the Boulevard / Edgewood corridor) grew 2.47× from 2024 — the fastest-growing HD in the district. HD-090 (south-central DeKalb, running from Candler Park through McPherson and the Glenwood / Gresham Road corridor) grew 2.33×. HD-116 (south DeKalb / Cedar Grove) grew 2.25×. HD-084 (east DeKalb / Decatur fringe) grew 2.44×.

Clayton's HD-076 grew 2.01× — slower than any DeKalb HD in the district. That single number is the engine of the geographic-composition story: SD-44's Clayton wing kept up with its own 2024 pace but lost relative share because DeKalb grew faster around it.

3.3 · HD shifts and what's driving them

The HD-share chart in §2 shows the cycle structure at a glance. Three HDs moved meaningfully: HD-076 down 3.1pp, HD-089 up 1.7pp, HD-090 up 0.9pp. HD-084, HD-116, and HD-078 all shifted by less than a percentage point.

The HD-076 drop is concentrated in the Morrow cluster (Morrow 2, 5, 8, 10, 11 all running below their 2024 pace) plus both Ellenwoods. The HD-089 gain is concentrated in Boulevard, East Lake, Coan Recreation Center, Oakhurst, and Decatur. The HD-090 gain is concentrated in Coan Recreation Center, Metropolitan, McNair High, Burgess Elem, and Gresham Road, partly offset by the Candler Park drag.

3.4 · Precinct-level deep dive

The biggest 12 precincts by raw cycle volume are: Candler Park (HD-090, 1,107 ballots), Boulevard (HD-089, 932), Ellenwood 2 (HD-076, 922), Cedar Grove South (HD-116, 773), Cedar Grove Elem (HD-116, 759), McNair High (HD-090, 750), Coan Recreation Center (HD-090, 699), East Lake (HD-089, 678), Ellenwood 1 (HD-076, 674), Burgess Elem (HD-090, 645), Clairemont East (HD-084, 629), and Meadowview (HD-090, 622).

The cycle-share gainers vs 2024 are dominated by HD-090 (Coan Recreation Center +0.97pp, Metropolitan +0.94pp, McNair High +0.79pp, Burgess Elem +0.62pp, Gresham Road +0.32pp) and HD-089 (Boulevard +0.80pp, Oakhurst +0.54pp, East Lake +0.49pp, Decatur +0.39pp). The losers are split between Candler Park (−1.28pp, the single biggest mover in either direction), Flat Shoals (−0.83pp), and the Clayton Morrow / Ellenwood cluster (Morrow 5 −0.77pp, Ellenwood 1 −0.71pp, Morrow 10 −0.39pp).

Friday's relative-share movement was a sharper version of the cycle-share pattern: Metropolitan, Burgess Elem, Boulevard, and McNair High over-pulled even more on Friday than they had over the full cycle, while Ellenwood 2, Morrow 5, and the Decatur Clairemont/Glennwood cluster gave back even more relative share on close.

3.5 · Demographic deep dive

The race × age share-shift heatmap below pinpoints where the demographic divergence lives. The single biggest move is in the 65+ Black cell: 31.6% of the 2024 cycle was 65+ Black voters; in 2026 that share is 22.7%. That single cell lost roughly 8.9pp of district share — more than any other race × age cell.

The biggest gainers are 30–44 White (2024 5.1% → 2026 8.9%) and 30–44 Black (2024 4.1% → 2026 7.1%). The 2026 surge is generationally narrow: voters in their thirties and early forties drove most of the growth, in both racial groups that anchor the district.

Gender stayed essentially unchanged from 2024: F was 63.1% of the 2026 cycle vs 63.2% in 2024.

By propensity bin, the 2026 cycle is anchored by high-propensity voters: 65% of EV came from P78 (high-propensity voters), 20% from P56, and a combined 15% from the three lower-propensity tiers. That mix means the 2026 surge is mostly the district's reliable primary-going electorate showing up at higher volume — not a wave of brand-new EV voters.

2026-05-16T13:24:34.294868 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

3.6 · Within-cycle temporal pattern

The cumulative curve in §1 already encodes most of the temporal pattern. The weekday vs weekend rhythm was sharper in 2026: weekdays in the final week averaged 2,704 ballots, while Saturdays (5/2, 5/9) and Sundays (5/3, 5/10) averaged just under 399.

Per-HD daily share evolved over the cycle. HD-090 climbed steadily from the low thirties early to the high thirties on close. HD-089 was in the mid-teens early and the low twenties on close. HD-076 ran in the mid-twenties early and slipped to the low twenties on close. The intown-DeKalb piece was where the late-cycle composition shift lived.

The youth surge was back-loaded rather than gradual: 18–29 share ran around 6% through the first two weeks and jumped to 10% on Friday. 30–44 was similarly back-loaded, climbing from the mid-teens early to 29% on Friday.

3.7 · What this implies for Tuesday — empirically-grounded scenarios

Anchoring scenarios in the 2024 EV-vs-E-Day evidence

The standing uniform-swing model in analysis/sd-44-2026-05-16.xlsx applies the 2024 Parent-vs-Thomas precinct shares (Saira Draper inherits Parent's; Thomas inherits her own) plus a single district-wide Thomas swing parameter. The right way to pick that swing is to anchor it in something empirical. The cleanest local anchor is the 2024 EV→E-Day Thomas share shift: the largest within-cycle, within-candidate movement the district has actually shown us.

In 2024, Thomas's share of the two-way vote was 34.8% in advance/EV vs 29.7% on Election Day — a district-wide shift of +5.1pp (Parent did better on E-Day). The pattern was an amplification: HDs that favored Parent intensified that lean on E-Day, and HDs that favored Thomas intensified hers. The chart below shows it by HD:

2026-05-16T13:24:34.838795 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

HD-089 (intown DeKalb, Parent's strongest HD) swung 7.5pp toward Parent on E-Day; HD-116 (Cedar Grove, Thomas's strongest HD) swung 7.0pp toward Thomas. The district-wide 5.1pp "method-shift magnitude" is the empirical yardstick this section uses to size the 2026 scenarios. One method-shift magnitude was the largest within-cycle Thomas share movement in 2024; two magnitudes is at the edge of what 2024 evidence supports without invoking candidate-specific dynamics not observable in the data.

EV-weighted baseline + critical swing

Thomas's 2024 advance/EV share weighted by 2026 EV ballot volume per precinct is 29.0% on a hybrid baseline: 2024 EV-only Thomas share for the 69% of 2026 EV volume in precincts where DeKalb's official SOVC published an advance/E-Day candidate breakdown (33 of 49 precincts), and 2024 blended Thomas share for the remaining 31% where the SOVC redacted those cells as "Insufficient Turnout to Protect Voter Privacy" (16 DeKalb precincts including Cedar Grove South, East Lake, Fernbank Elem, and 13 others — totals are published, per-method per-candidate splits are not; see §3.8). The EV-only race ties at a district-wide Thomas swing of +21.0pp off this baseline — 4.1× the empirical 2024 method-shift magnitude.

2026-05-16T13:24:34.760100 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

Three scenarios, calibrated to 2024 evidence

EV uses 2024 EV per-precinct shares; E-Day uses 2024 E-Day per-precinct shares (method-matched); the swing applies uniformly across precincts on top. E-Day volume is held at 14,000 (close to 2024's 8,696).

2026-05-16T13:24:34.802153 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
ScenarioThomas swingEV Saira / ThomasTotal margin (EV + ~14,000 E-Day)
A · Inherited margin0pp 16,544 / 6,758 (71% / 29%) +15,459 Saira
B · Method-shift magnitude+5pp 15,379 / 7,923 (66% / 34%) +11,729 Saira
D · Double magnitude+10pp 14,214 / 9,088 (61% / 39%) +7,999 Saira

A — Inherited margin (0pp swing). No candidate-specific shift — voters treat Saira and Thomas as fungible substitutes for Parent and Thomas. Likely too optimistic for Saira because Parent had eight years of incumbency; defensible as a "if 2024 patterns just hold" base case.

B — Method-shift magnitude (+5pp). Thomas runs 5pp better than her 2024 baseline — equivalent to what the EV→E-Day shift was in 2024. This is the most defensible "open-seat reset" scenario: the swing is empirically the largest movement the data has shown. Saira still wins decisively (66–34% on EV) because the 2024 baseline tilts so far Parent.

D — Double magnitude (+10pp). Thomas runs 10pp better than 2024 — twice the empirical magnitude. At this swing, Saira's margin shrinks to ~22pp EV-only but she still wins. Beyond this magnitude, the swing exceeds anything 2024 evidence directly supports and would require invoking candidate-specific dynamics not observable in the data (e.g., Saira's name ID running materially weaker than Parent's, a Thomas-specific GOTV operation outperforming Parent's 2024 ground game, or a Clayton-skewed E-Day surge).

Tuesday signals

This is a uniform-swing projection, not a candidate forecast. The scenarios are arithmetic implications of swings calibrated to the largest within-cycle shift the 2024 data showed; they are not assigned probabilities. The most honest output is the sensitivity chart — read off the projected margin at whatever swing Tuesday's early returns suggest.

3.8 · Methodology and caveats

Snapshot framing. All numbers in this report come from the 2026-05-16 snapshot, which reflects ballots returned through end-of-day 2026-05-15 (the final EV day). Snapshot date and data-through date are not the same thing — the standing daily-brief convention.

Cube source. Postgres analytics.ev2026_primary_precinct_daily_crosstab joined to analytics.ev2026_primary_precinct_geographies on entity_id = geog_id, filtered to state_senate_district = '044' and party = 'DEM'. Single-dimension breakouts hold all other dimensions at 'ALL' — the cube stores marginal totals at that sentinel value, so summing rows where multiple dimensions are non-'ALL' would double-count.

2024 comparison. Identical schema in the analytics.ev2024_primary_precinct_* tables. Precincts join on the 2026 SD-44 boundary using the precinct crosswalk. SD-44 is overwhelmingly in DeKalb and Clayton, so the precinct-vintage mismatches noted elsewhere (Dodge / Oconee / Worth / Fulton) do not apply.

Demographic categories. Race is BISG-style assignment from the BQ voter file; age is from DOB at snapshot date; party is from primary-ballot choice (the SOS file). Gender is self-reported; the small "NA / U / X" cells are excluded from share denominators in the charts but included in the totals.

2024 candidate results. Parent vs Thomas precinct-level results from analytics.unified_election_results filtered to election_date='2024-05-21' AND office_type='state_senate' AND district_number='44' AND party='Dem' AND source='sovc'. Other 2024 candidates (Tyriq Jackson, Andre W. Herman) had a combined <1% share and are excluded from the two-way Parent-vs-Thomas baseline.

Source-level redaction (DeKalb SOVC). 16 of 49 SD-44 precincts (all DeKalb — including Cedar Grove South, East Lake, Fernbank Elem, Lin-Mary Lin, Metropolitan, Glennwood, and 10 others) are intentionally redacted in the DeKalb County Statement of Votes Cast (verified against the official 2024 SOVC PDF — those precincts' per-method per-candidate cells are printed as **** with the annotation "Insufficient Turnout to Protect Voter Privacy"). The per-method ballot counts are published for those precincts (Cedar Grove South: 498 ED + 279 Advance + 14 Mail + 1 Provisional = 792 cast), and so is each candidate's total; only the per-candidate × per-method intersection is suppressed.

These 16 precincts hold 31% of the 2026 EV volume. The EV-matched scenarios in §3.7 handle the redaction by falling back to the 2024 blended Thomas share for those precincts (the only honest treatment given the source). The "2024 EV vs E-Day" per-HD amplification chart in §3.7 and the EV-vs-E-Day scatter in §8 are computed only on the 33-precinct subset where DeKalb did publish the breakdown — that subset is large enough to characterize the district-wide pattern but should not be read as a complete enumeration.

VCU sizing. Registered-voter counts from analytics.ev2026_primary_precinct_registered, propensity tiers as encoded by the BQ voter-file pipeline. Voted-EV counts from the same daily crosstab as the rest of the report. The local sanitized DB has no per-voter history, so "lapsed-EV" estimates are precinct-aggregate gaps rather than voter-level joins.

Explicitly out of scope: Election Day, post-election results, voter contact universe (VCU / E-VCU / X-VCU) performance against campaign-specific lists, candidate-level analysis, statewide-2026 comparisons, and 2022-special comparisons. Those are separate analyses.

8 · Precinct-share model review (detailed)

What the model assumes, where it has leverage, and where it doesn't

The standing uniform-swing model takes each SD-44 precinct's 2024 two-way Parent-vs-Thomas vote share and applies a district-wide swing to project the 2026 Saira-vs-Thomas margin. The 2024 baseline is real, recent, and at the precinct level — but it carries assumptions worth examining before reading the projection too literally.

The 2024 baseline. Parent won the 2024 SD-44 Dem primary 73.4% to Thomas's 26.6% on a two-way basis (17,855 to 6,468 votes across the 49 SD-44 precincts). That margin is district-wide; the precinct-level distribution is what matters for the swing model. Map 2 below shows it.

2026-05-16T13:24:34.387662 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

The distribution is bimodal-ish: most precincts are lopsided in one direction or another, with a narrow band of competitive precincts in the middle. Below is the distribution weighted by 2024 ballot volume, so each precinct's bar is proportional to its 2024 turnout (not its land area).

2026-05-16T13:24:34.592427 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

16 precincts gave Parent 80%+; 5 precincts went Thomas-majority. Only 17 of the 49 sat in the genuinely competitive 40–60% Parent band. The uniform-swing model has the most leverage in those middle precincts: a 5pp district-wide swing toward Thomas moves a 55% Parent precinct to 50%, but a 95% Parent precinct stays solidly Parent.

2024 EV vs Election-Day Parent shares

A key assumption of the EV-only projection is that EV voters and E-Day voters vote alike at the precinct level. If they don't — if EV draws systematically more pro-Parent or pro-Thomas voters than E-Day — the EV-only projection misses the eventual final margin.

The 2024 data lets us check this directly. The scatter below plots each precinct's 2024 Parent share in advance/EV against its Parent share on Election Day, weighted by E-Day volume:

2026-05-16T13:24:34.623967 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

Across the 49 precincts, Parent's 2024 EV share was 65.2% and her Election-Day share was 70.3% — essentially identical. The precinct-level scatter sits tightly on the 45-degree line. That's the empirical justification for the uniform-swing approach: at least in 2024, the precinct-level vote share didn't depend much on which ballot type a voter cast. If 2026 follows the same pattern, the EV-only projection should be a reasonable approximation of the final margin.

Caveat: 2024 was a Parent re-election with high name recognition. 2026 is a special with two candidates who don't share that incumbent advantage, so EV-vs-E-Day differential could be larger. Treat the 2024 EV/E-Day equivalence as evidence that the swing approach is reasonable, not as proof.

Where the model has the most leverage

Leverage = volume × competitiveness. The precincts that will move the model output most under any given district-wide swing are the ones that are both high-volume in 2026 and were close in 2024. The table below ranks the 12 highest-leverage precincts (competitiveness = 50 − |Parent% − 50|, so a 50% precinct gets full weight and a 90% precinct gets only 10%):

PrecinctHDParent 2024 share2026 EV ballotsLeverage score
Cedar Grove ElemHD-11647.8%75936,313
Cedar Grove SouthHD-11653.4%77336,012
Ellenwood 2HD-07663.4%92233,785
MeadowviewHD-09049.0%62230,463
Mcnair HighHD-09064.8%75026,437
Morrow 5HD-07657.9%59425,011
Ellenwood 1HD-07662.9%67424,976
Gresham RoadHD-09055.9%56424,854
Morrow 8HD-07664.3%61822,062
Terry MillHD-09059.1%52621,508
Bouldercrest RoadHD-09046.7%42019,614
Morrow 10HD-07656.7%44219,149

The bulk of this list is HD-076 (Clayton — Morrow / Ellenwood) and the Cedar Grove pair. These precincts were close in 2024 AND have substantial 2026 EV volume. If the district-wide swing is non-trivial, these precincts will swing the final margin most.

Where the model is least informative

By contrast, the precincts below were so lopsided in 2024 that the uniform-swing model essentially passes them through unchanged. Mobilization shifts won't show up as candidate-preference changes; only as turnout-volume changes — which the swing model handles correctly but which campaign GOTV should account for separately:

PrecinctHDParent 2024 shareLean2026 EV ballots
Fernbank ElemHD-08995.9%Parent landslide440
Druid Hills HighHD-08995.7%Parent landslide440
Johnson Estates (Atl)HD-09095.2%Parent landslide294
Lin-Mary Lin Elem (Atl)HD-09094.3%Parent landslide620
Clairemont East (Dec)HD-08493.6%Parent landslide627
Candler Park (Atl)HD-09093.6%Parent landslide1,107
Winnona Park (Dec)HD-08493.3%Parent landslide453
Clairemont West (Dec)HD-08493.1%Parent landslide561

Candler Park, Clairemont West, Clairemont East, and Winnona Park are basically locked in for the non-Thomas candidate; the Morrow cluster and a couple of Cedar Grove precincts lean Thomas. For these precincts the model is mechanically correct but operationally near-irrelevant — they will swing neither candidate's margin meaningfully on a per-precinct basis. Pay attention to the middle of the distribution instead.

Limitation: this section evaluates the model's structural assumptions but does NOT backtest model predictions against actual 2024 results (which would require a 2022 baseline of the same candidates — not available). The strongest backtest available locally is the 2024 EV-vs-E-Day equivalence above, which the data passes cleanly.

9 · Voter contact universes (detailed)

Who's still out there to mobilize

SD-44 has 151,885 total registered voters across the 49 precincts. As of close of EV, 23,303 have voted in the 2026 DEM EV (15.3% of registered). That leaves 128,582 registered voters who have not yet cast a 2026 ballot — the universe available for Tuesday Election Day mobilization.

Not all of those 128,582 are equally likely to vote on Tuesday given a contact. The voter file's propensity tiers (P0 through P78, where P78 is "voted in 7-8 of the last 10 relevant primaries") gives a strong prior on who's reachable. The breakdown:

2026-05-16T13:24:34.659655 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
Propensity binRegisteredVoted 2026 EVTurnout to dateNot yet voted (E-Day target)
P0 (no primary history)37,5276071.6%36,920
P12 (low)19,0074332.3%18,574
P34 (med-low)34,4792,3486.8%32,131
P56 (med-high)23,4274,69120.0%18,736
P78 (high — most reliable)37,44515,22440.7%22,221
SD-44 total 151,885 23,303 15.3% 128,582
22,221 P78 voters have not yet voted
This is the prime E-Day GOTV universe — high-propensity primary voters who are statistically very likely to vote on Tuesday given any contact. 41% of P78 voters in SD-44 already banked their EV ballot, which leaves 22,221 P78 voters as the highest-yield Election Day target. P56 adds another 18,736 medium-propensity voters where contact has substantially higher yield than the lower tiers.

Universe sizing by House District

The chart below stacks the not-yet-voted universe by propensity tier within each HD. Each bar's red+orange segment is the "yield-likely" target (P56 + P78); the gray portions are lower-yield but still on the voter file. HD-090 carries 42,647 not-yet-voted voters and HD-076 carries 32,511; HD-089 carries 30,275.

2026-05-16T13:24:34.703299 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/

Lapsed-EV proxy

The local sanitized DB does not contain per-voter history (the analytics.voter_history table is empty by design), so a true voter-level "voted DEM EV in 2024 but not in 2026" universe cannot be sized here. The closest precinct-level approximation: 2024 SD-44 DEM primary total turnout was 25,234 voters (across EV + E-Day + mail in 2024). 2026 EV alone is 23,303 voters — that's 92% of the entire 2024 Dem primary turnout already banked, with E-Day yet to come.

If Tuesday adds even half of what 2024 Election Day did, the 2026 SD-44 DEM primary will run materially larger than 2024 in total turnout. The GOTV question is therefore not "can we catch 2024" — that's already done — but "how much higher can E-Day push us, and which precincts have the most untapped P56+P78 voters."

Highest-yield precincts for E-Day GOTV

Map 4 colors each precinct by raw "registered − voted-EV-2026" (the total not-yet-voted universe per precinct). The hottest precincts — HD-090's Clayton-County-area precincts, Boulevard, the Cedar Grove pair, and the intown HD-090 corridor — are where E-Day contact will reach the most voters per door-knocked:

2026-05-16T13:24:34.549734 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
Caveat: a "registered" denominator includes inactive and uncontactable voters, voters who have moved within the district but not updated registration, and voters who are nominally registered DEM but won't vote DEM. The propensity-bin breakout in the table above is a more honest yield estimate. Per-voter deduplication against an external campaign list (VAN exports, etc.) is a separate analysis — see the standing civicfs-vcu-analysis skill, which takes campaign-defined VCU/E-VCU/X-VCU lists as input.

A · Fact-check appendix

Every numeric claim re-derived from postgres at render time

This appendix is auto-generated by re-running each headline calculation directly from postgres and comparing against the value printed elsewhere in the report. If any line shows FAIL the report is internally inconsistent.

PASS EV 2026 cycle total: report=23,303 vs recomputed=23,303 PASS EV 2024 cycle total: report=10,278 vs recomputed=10,278 PASS 2026/2024 ratio: report=2.27× vs recomputed=2.27× PASS Friday 5/15 total: report=5,883 vs recomputed=5,883 PASS In-person count: report=22,804 vs recomputed=22,804 PASS In-person % (×10): report=97.9% vs recomputed=97.9% PASS HD-076 cycle share %: report=23.9 vs recomputed=23.9 PASS 65+ Black 2026 cell count: report=5,282 vs recomputed=5,282 PASS 65+ Black 2024 cell count: report=3,249 vs recomputed=3,249 PASS Parent 2024 total votes: report=17,855 vs recomputed=17,855 PASS Thomas 2024 total votes: report=6,468 vs recomputed=6,468 PASS SD-44 registered total: report=151,885 vs recomputed=151,885 PASS Final-week (Mon-Fri) total: report=13,521 vs recomputed=13,521 PASS Week 1 total: report=4,354 vs recomputed=4,354
Generated by the SD-44 EV analysis pipeline from postgres analytics.ev2026_primary_precinct_*, analytics.ev2024_primary_precinct_*, analytics.unified_election_results, and analytics.turnout_summary. To re-run with a fresh snapshot, re-execute the same script — every figure above is derived live from postgres.