Constituency-level intelligence for the 2026 US midterms: a framework

From 300 seats in Dhaka to 435 districts in Washington #

Strategos was first commissioned as a constituency intelligence layer for Bangladesh's 2026 parliamentary elections, where 300 directly elected seats demanded a uniform analytical treatment despite huge variance in district size, ethnic composition, and party history. Each constituency was modeled as an autonomous agent with three internal stores: a static demographic profile, a rolling political baseline drawn from the previous three election cycles, and an event log fed by field reporters, news scrapers, and party sources. Agents communicated upward to regional supervisors and laterally to neighboring constituencies, allowing the system to flag contagion effects when, for example, a defection in one district shifted candidate calculus in adjacent ones.

The same skeleton transfers to the United States with minimal redesign. The 435 House districts replace the 300 Bangladeshi constituencies, the Senate adds a smaller layer of 33 statewide races per cycle, and gubernatorial contests slot in as a parallel state-level track. The agent contract stays identical: each district owns its data, exposes a standard query interface, and reports drift signals to a national coordinator. What changes is the data plumbing. Bangladesh relied on Election Commission gazette files and field census data; the US version pulls from FEC bulk filings, Census American Community Survey tables, state-level voter file aggregates, and the Cook Political Report partisan voter index.

Historical baseline construction #

Every district agent boots with a historical baseline that combines five inputs: the most recent Cook PVI score, presidential two-party vote shares for 2008, 2012, 2016, 2020, and 2024, House results from the last four cycles, demographic composition from the latest five-year ACS estimates, and a turnout vector indexed to presidential and midterm cycles separately. The baseline is not a single number but a distribution. For each district we compute a posterior over expected Democratic two-party share given midterm conditions, with uncertainty driven by within-district volatility across cycles and by the structural gap between presidential year and midterm electorates.

The PVI anchors the prior because it already encodes a national-environment-adjusted lean, but Strategos treats PVI as a slow-moving variable rather than a fixed truth. When a district's presidential margin moves more than four points between two consecutive cycles, the agent flags a structural realignment candidate and downweights older observations. This matters in 2026 because roughly forty House districts have shown sustained drift since 2020, particularly in Sun Belt suburbs and Rio Grande Valley counties, and stale PVI assumptions would mislead resource allocation.

Cycle-over-cycle drift detection #

The most operationally useful output of Strategos is not the level estimate but the drift estimate. For each district the system computes the change in expected Democratic share between the current cycle's posterior and the previous cycle's posterior at the equivalent point in the calendar, holding the national environment constant via a generic ballot adjustment. Drift exceeding 2.5 points in either direction triggers a review, and drift exceeding 5 points triggers an alert routed to the regional supervisor and the national desk.

Drift detection is more sensitive than simple polling comparison because it integrates registration changes from the voter file, candidate quality scores derived from FEC fundraising velocity and endorsement counts, and incident-level field reports. A district can show a flat polling average yet a clear underlying drift if, for example, the Republican incumbent has lost two consecutive quarters of small-dollar donor count while the Democratic challenger has built an organic field operation. Strategos surfaces these compositional changes before they show up in topline numbers, which is precisely the window in which campaign resources can still be reallocated.

Field intelligence integration #

Quantitative baselines are necessary but insufficient for live campaign work. Strategos accepts structured field intelligence through a lightweight reporter app: each submission carries a constituency identifier, an event type tag, a free-text observation, optional geocoded location, and a confidence score chosen by the reporter. Event types include candidate appearances, opposition activity, local press coverage, ground game indicators such as door-knock saturation or sign density, and incident reports such as polling location changes or disinformation flare-ups.

The agent for each district treats field reports as soft evidence updates. A single anecdote from a low-confidence reporter shifts the posterior trivially; a converging pattern of medium-confidence reports across ten districts in a media market produces a regional narrative flag. The system explicitly tracks who reported what and when, which is important both for accountability inside the campaign and for retrospective calibration, scoring which reporters proved predictive over the cycle and adjusting their future weight automatically.

Photocard generation for state operations #

State-level operations need a one-page summary per district that a regional director can scan in under two minutes. Strategos generates a photocard for each constituency on a daily refresh, rendered to PDF and pushed to a shared drive. The photocard fits on a single side of letter paper and contains the district identifier, current rating, drift indicator, top three field intelligence items from the last seven days, the candidate fundraising delta, and a small block of demographic context relevant to the current narrative.

Photocards are deliberately constrained in design. They use a fixed layout so that a director comparing twelve districts in a swing state can locate the same field at the same coordinate on every card. Color is reserved for drift direction and rating change; everything else is monochrome. The format borrows from the briefing cards used by congressional intelligence committees, where decision-makers learned long ago that consistency beats prettiness when the goal is accurate cross-comparison under time pressure.

Three deployment scenarios #

The architecture supports three concrete deployments at different scales. The first is a single congressional campaign that licenses the agent for its own district plus the four nearest comparable districts as benchmarks. The second is a state party committee covering all House districts in the state plus the Senate race if applicable. The third is a national committee or party-aligned super PAC running the full 435 House districts and the Senate map, with the regional supervisor layer activated for media market clustering. Pricing, refresh cadence, and field reporter capacity scale with deployment tier, but the underlying agent code is identical across all three.

Scenario selection drives the operational rhythm more than the analytical content. A single campaign typically wants daily photocards and weekly drift reviews; a state committee wants twice-weekly aggregated dashboards plus exception reports; a national committee wants a real-time alert stream with a Monday morning national rollup. Strategos exposes the same underlying state through different presentation layers, which keeps the analytical core honest because every customer is reading off the same numbers.

Why Strategos, and what it does not promise #

Strategos is not a forecasting model in the FiveThirtyEight or Decision Desk sense. It does not produce a single national seat count or a probabilistic call on the House majority. Its job is to maintain a coherent, auditable, district-level state that campaign and committee staff can act on, with explicit provenance for every claim and explicit uncertainty on every number. The Bangladesh deployment proved that a 300 agent architecture could run for an entire pre-election period without analytical drift, and the same operating discipline applies to the US.

Three boundary conditions are worth stating plainly. First, Strategos depends on the quality of upstream public data; FEC filing lags and ACS estimate vintages set the floor on how fresh the baseline can be. Second, field intelligence is only as good as the reporter network, and standing up a 300 plus reporter network for a national deployment takes roughly six weeks of recruitment and training. Third, the system is built to inform human decision-makers, not to replace them; every alert routes to a named supervisor with a recommended action, but the action itself is a campaign judgment call. Used inside those constraints, the platform gives state and national operations a shared, current, district-grounded picture that no single analyst could maintain by hand.

Sources #

• Cook Political Report PVI
• MIT Election Data and Science Lab
• FEC bulk filings
• Census ACS demographic estimates
• Catalist voter file research (cited)