METHODOLOGY

In short: rewards members whose campaigns are funded by lots of small individual donors rather than PACs or a handful of big donors and industries. The more spread-out and grassroots the money, the higher this score — regardless of party or chamber.

Measures five dimensions: (1) PAC dependency — a blend of the share of funding from PACs and how close contributing PACs are to their legal per-election cap, chamber-specific since Senate and House candidates rely on PAC money at structurally different rates; (2) the share of funding from small (<$200, unitemized) donors — the broadest possible funding base; (3) relative top-donor concentration — what fraction of the itemized external donor pool comes from the top 10 donors, with the member's own money and transfers from their own committees excluded; (4) source breadth — small-donor money counts fully, industry-classified money counts moderately, and opaque money counts least; and (5) industry concentration — the inverse Herfindahl-Hirschman Index (HHI) of industry donations, where funding concentrated in a single industry suggests potential regulatory capture. Components (4) and (5) were folded in from a separate Funding Diversity dimension in 2026-07 after finding the two dimensions correlated at r=0.72 across the Senate — the same underlying funding-profile signal under two labels, not two genuinely distinct ones. PAC dependency follows Stratmann (2005), [5] Stratmann 2005who found that PAC contributions are more strongly correlated with roll-call alignment than individual contributions. The concentration components apply the same intuition as HHI at the donor level, following Bonica (2014) [1] Bonica 2014 and the industrial-organization logic Rhoades (1993) [6] Rhoades 1993 built the HHI metric on.

"UNCLASSIFIED" money (committee transfers, joint-fundraising splits, donations lacking employer data — a real 32% median share of total funding across the Senate) is scored neutrally rather than penalized. It is a residual we cannot attribute at all, not evidence of concentration in one source — the same "missing data defaults to neutral" principle applied everywhere else on this page.

In short: checks whether a member's voting actually matches what their state or district elected them to do. Voting with your party is notpenalized on its own — for a safe-seat member, that often IS representing your constituents. The score moves below neutral only for a clear, readable sign of a mismatch — a voting position toward the party's flank for a seat that isn't safe for that flank — and moves above neutral for the mirror case: a voting position genuinely in step with the seat.

Measures how a member's voting compares to what their state elected them to do — not raw defection from party. Each member's contested-vote break rate is scored against a seat-specific expectation derived from state partisan lean (Cook PVI [4] Carson et al. 2010): an aligned safe seat expects near-base-rate dissent (~3%), a swing seat ~8%, and a seat whose electorate leans toward the opposing party up to ~20%. Matching the expectation scores ~50 — a typical partisan for that seat. The score is deliberately asymmetric around that expectation (v6.6), under one governing principle: it moves off neutral only for readableevidence a member is representing their constituents, and treats behavior whose meaning can't be read as neutral. Below-expected loyalty is notpenalized: it floors at neutral, never below. A low defection rate is unreadable — it may be faithful representation of the coalition that elected the member (not the geographic median voter), it is the structural norm for both parties in the modern Senate, and being "out of step" is a matter of ideological position, not a loyalty rate — so we decline to score the rate itself rather than penalize it. This loyalty floor was the only behavioral change in v6.6, and it is deterministic: a below-expected loyalist with no other signal available scores exactly 50 on this component. v6.7 adds one legible exception: if a member's cosponsorship-derived ideology score places them in their own party's most extreme third, and their seat isn't safely aligned for that extremity, they are discounted below neutral (scaled by how unsafe the seat is — not penalized at all in a genuinely safe seat, where that extremity is the structural norm for both parties). This targets ideological POSITION, not the loyalty rate itself, so it doesn't reopen the rate-is-unreadable problem above — a member can be maximally loyal and still be flagged if their position is a clear outlier for their seat. This discount's maximum strength was reduced in v6.8 after a fairness audit found it overlapped with coalition breadth (below) more than intended — see the scoring changelog and the Known Limitations note below. Above-expected crossing is the readable side: it earns credit only where it plausibly moves toward the seat's political center, discounted by seat lean (full credit in opposed and swing seats, near-neutral in deep aligned seats, since there the center sits with the party). A further discount for members positioned on their party's ideological flank — whose crossings more likely point awayfrom the center (Kirkland & Slapin 2017) — was designed but is notshipped, and checking it against live data found a deeper problem than a missing calibration number: the members who actually cross party lines most often all read as ideologically centrist on this platform's own cosponsorship-based ideology measure, not flank-extreme — the opposite of what the discount assumes. Crossing behavior and this measure of ideology turn out to be linked rather than independent, so this specific fix is shelved, not just uncalibrated. This is a deliberately humble use of the delegate model, with partisan lean standing in for issue-level constituent opinion — a measurable, disclosed simplification (see Known Limitations below, including why this measures the rate and direction of a member's deviation rather than the distance between their position and their constituency's, so it cannot yet positively credit representation achieved through congruent loyalty). Note on composition: confirmation votes on nominations make up a large share of recent Senate roll calls and count at full weight — they are genuine, whipped party-line tests.

Before v4.2 this dimension was called Independent Voting and rewarded raw defection; it also exempted party-line votes on policy areas related to a member's top donor industries. That exemption is removed: donor industries are not a proxy for state interests, and it shielded exactly the votes most suspect for donor influence.

The score blends seat-relative vote alignment (70%, or 100% when roll-call ideal-point data is unavailable) with position congruence (30%, when available — v6.11): the member's DW-NOMINATE first-dimension position (Voteview; the standard roll-call-based position measure in political science) compared against a seat-conditional expectation — what a same-party member of a similarly-leaning seat typically holds, fit per chamber and per party from real data. Per-party fits deliberately avoid the swing-seat artifact a single pooled fit would create (Bafumi & Herron 2010). A position toward the party's flank relative to that norm scores below neutral (scaled by how unsafe the seat is — flank positions in genuinely safe seats are the structural norm and are not penalized); a position toward the seat's center scores above neutral (with the same seat-direction credit shape as surplus crossing). When active, this component supersedes the v6.7 cosponsorship-based discount above — same construct, better signal, and measuring it twice would repeat the exact double-count v6.8 fixed. Coalition breadth (20% here from v5 through v6.10) has moved to Legislative Effectiveness: bipartisan coalition-building is a legislative-effectiveness signal, not a constituent-alignment one — demand for bipartisanship varies with the seat's own makeup (Harbridge & Malhotra 2011), so a bipartisan member of a lopsided seat can be bipartisan and misaligned at once. Through v6.4 this dimension also included a Donor Independence component (25%, a heuristic based on the money associated with donor-vote topical overlaps) — removed in 2026-07 after finding it measured a close cousin of the Funding Independence signal (both keyed off total money raised and donor-industry concentration) while itself reducing to one of four fixed values for 85% of senators, since no data source discloses per-bill donor positions. Its freed weight now goes entirely to seat-relative vote alignment. We follow the methodological caution of Ansolabehere et al. (2003) [18] Ansolabehere et al. 2003 in interpreting donation-vote correlations generally: correlation does not prove causation. [5] Stratmann 2005

In short: measures whether a member is actually getting legislative work done — introducing bills that matter, moving them through Congress, and building a network of cosponsors other members trust. Introducing a substantive bill earns real credit even before it passes, matching how political scientists actually measure legislative productivity.

Measures whether a member is producing tangible legislative outcomes, following Volden & Wiseman's (2014) [34] Volden & Wiseman 2014 real published methodology: each sponsored bill is weighted by significance (5x for substantive bills — S./H.R./joint resolutions; 1x for commemorative simple/concurrent resolutions) and credited cumulatively across every stage it reaches — introducing a bill earns real credit on its own, not just bills that go on to pass a chamber or become law. Three components: bill significance & advancement (60%) — this cumulative stage-credit per congress served, compared against an expected credit for a sponsor of this chamber/majority-minority status; legislative leadership (25%) — cosponsorship-network PageRank, see below; and bipartisan coalition attraction (15%, v6.11 — moved here from Constituent Alignment): the share of cosponsors a member attracts to their own bills from the other party, chamber-median-normalized. Harbridge-Yong, Volden & Wiseman (2023) show that attracting cross-party cosponsors robustly predicts lawmaking success for majority and minority members alike — and that it is specifically the attraction of bipartisan cosponsors, not the offering of cosponsorships across the aisle, that carries the effect, so this component uses a receive-only rate rather than the blended give-and-receive bipartisanship figure shown on profiles. Their evidence is a strong association rather than a clean causal proof, and the measure is an input to effectiveness rather than realized output — both disclosed reasons its weight stays modest. When cosponsorship data is missing, the split reverts to exactly the prior 70/30.

Because introduction itself earns credit, a member who sponsors many substantive bills can score well even before any of them advance further — this is Volden & Wiseman's real design, not a bug: their published methodology counts a bill's contribution at every stage it reaches, and most sponsored bills never advance at all (our own corpus measures Senate majority sponsors advancing bills at 3.6% vs. 2.4% for minority sponsors; House 6.4% vs. 2.4%). The expected-credit baseline accounts for that majority/minority gap, so scoring everyone against one absolute threshold doesn't silently penalize whichever party is out of power. The score explanation on each profile breaks the substantive-bill count into introduced-only / advanced-further / became-law so the volume-vs-advancement split is visible as real numbers.

Both the expected-credit baseline and the majority/minority-status adjustment above are calibrated separately for the House and Senate (v6.9) — the two chambers' real bill volumes and advancement rates genuinely differ, and comparing every member against one shared, pooled-across-chambers figure previously understated House members' effectiveness and overstated the Senate's. See the scoring changelog for the live-population numbers behind that fix.

Measures legislative influence using the PageRank algorithm [32] Brin & Page 1998applied to cosponsorship networks. When Senator A cosponsors Senator B's bill, that creates a directed link in the network. PageRank computes centrality: a senator whose bills attract many cosponsors — especially from other influential senators — receives a higher score. This mirrors GovTrack's leadership methodology. [33] Tauberer 2012

The algorithm uses power iteration with a damping factor of 0.85 and converges in ~50 iterations. Raw PageRank values are rescaled to [0, 1] using a logarithmic transformation to compress the heavy-tailed distribution, then displayed as 0-100.

Network centrality structurally takes years to build — a freshman senator's raw score is near-zero not because they lead poorly but because they haven't had time to accumulate cosponsorship connections yet. Both the score component and the displayed number shrink the raw value toward neutral 50 for senators with under 6 years in office, confidence-scaled to a full term, so a brand-new senator reads as "not enough track record yet" rather than "bad at leadership."

Raw cosponsorship-network centrality can't on its own tell a substantive bill from a message bill introduced with no real chance of passing — a senator who signs onto ten symbolic resolutions accrued the same network weight as one who cosponsors ten bills that actually became law. Since v6.2, each cosponsorship is weighted by what happened to the underlying bill: full weight if it became law, reduced weight if it passed a chamber or cleared committee, and further reduced (not zeroed — a stalled bill is still real evidence of a working relationship) if it never advanced.

Computes a behavioral ideological position using Singular Value Decomposition (SVD) on the cosponsorship matrix, following Tauberer (2012). [33] Tauberer 2012The second singular vector (first is trivially related to overall activity) captures the primary ideological dimension — the axis along which senators most differ in who they cosponsor. This is analogous to DW-NOMINATE [20] Poole & Rosenthal 1985but derived from cosponsorship patterns rather than roll-call votes.

The ideology score is oriented so that lower values correspond to progressive positions and higher values to conservative positions, calibrated by checking the mean score of each party. It serves as a Bayesian prior for the partisan depth calculation: when a senator has few recorded votes, the ideology score regularizes the estimate; as vote data accumulates, the prior weight drops to zero. [19] Efron & Morris 1975

Combines the leadership and ideology scores into a human-readable label (e.g., "progressive Democratic leader" or "conservative Republican backbencher"). The label encodes three dimensions: ideological position (progressive/moderate/conservative), party affiliation, and influence tier (leader/rank-and-file/backbencher).

Reflects approval trajectory and coalition retention. Gallup, this platform's original approval source, ended presidential approval tracking entirely in February 2026 after 88 years; approval data now comes from the American Presidency Project (presidency.ucsb.edu), which is still updated for the sitting president, aggregating AP-NORC/CNN-SSRS/Marist/Pew/Verasight. This covers every president from Truman onward — 70% average approval over the term, 30% the trend from term-start to term-end, both scored against real population statistics computed from every president's actual polling history. Presidents before Truman have no polling era at all, so their Public Mandate uses UCSB's historical election-margin data instead — the average margin of victory across their own election win(s), the pre-polling-era proxy. The five presidents who never won a presidential election in their own right have neither and show N/A for this dimension, not a fabricated number.

Measures tangible economic outcomes: GDP growth (60%) and job creation (40%). GDP growth is computed for the full presidency — BEA/FRED for the modern era, MeasuringWorth's real-GDP series (1790-present) for earlier presidents, both producing the same “average annual growth over the term” figure, with the term's first calendar year excluded when per-year data allows it (that year mostly reflects the outgoing administration's policy). Job creation comes from BLS nonfarm payroll data, which only exists from 1939 onward — presidents before that are scored on GDP growth alone, renormalized to 100% of the formula's weight, not defaulted on the missing component.

Measures how well executive agency actions align with stated presidential priorities, via Federal Register rulemaking data — the count of final and proposed rules published during the term, and what fraction were finalized rather than left pending. This is a digitization wall, not a conceptual one: notice-and-comment rulemaking was a real, functioning practice well before the 1990s, but no machine-readable record of it exists that far back — checked directly (2026-07) rather than assumed: federalregister.gov's API returns zero results for any pre-1994 president, and govinfo.gov's own structured Federal Register data starts at year 2000. Earlier issues exist only as scanned page images with no structured document-type or agency tagging, and reconstructing rulemaking counts from those would mean OCR'ing and classifying decades of raw scanned text — the same kind of unreliable pipeline already rejected for Follow-Through and Competence's court-success-rate. Every president before Clinton shows N/A for this dimension, excluded from their overall score entirely rather than scored on a proxy.

Covers what none of the other three dimensions can: crisis leadership, moral authority, vision, and similar historical-consequence judgments that don't reduce to GDP growth, approval polling, or rulemaking volume. Sourced from C-SPAN's Presidential Historians Survey — ~142 professional historians in the 2021 cycle (the most recent; the 2025 cycle was explicitly postponed by C-SPAN, citing the risk of turning “historical analysis” into “punditry” with a former president returning to office), scored across ten categories and aggregated into one point total. This is categorically different from the hand-set Independence/Follow- Through values removed elsewhere: a real, external, periodically-run survey with a documented methodology, not a single number invented for this platform — the same “trust a well-documented external institution” category as citing BLS or Federal Register data, just survey-based rather than administrative-record-based. Only rates presidents whose terms were complete as of the 2021 cycle — every currently-serving or just-departed president shows N/A here, genuinely unrated by the survey's own cadence, not a fetch gap.

Being real, external, and methodologically documented does not make this survey unbiased, and we don't present it as neutral ground truth. Political scientists who study these historian-ranking surveys have documented real, specific patterns in them: professional historians as a field skew toward favoring presidents who expanded federal/executive power, which plausibly inflates FDR, Wilson, and LBJ relative to how a more ideologically mixed panel might rate them; and historians are reluctant to rank very recent presidents at all until enough distance has passed to assess their legacy, which is the direct reason Obama and George W. Bush's scores may still be unsettled and Biden and the current president have none. Weighting this survey at 35% means this platform's ranking inherits those biases at roughly that same strength, not zero.

NEWS ANALYSIS PIPELINE

Eight RSS feeds from seven newsrooms — AP News, NPR (Politics and World desks, counted as one source), PBS NewsHour, BBC World, The Hill, Politico, and Roll Call — are parsed hourly; opinion and editorial sections are filtered out of every feed. Under common media-bias ratings this mix spans center to lean-left, with no right-of-center outlet currently included — a disclosed limitation of the source diet, not a neutral sample of all coverage. Each article is filtered for U.S. policy relevance using embedding cosine similarity against policy area prototypes — the same sentence-transformer model used throughout the platform. Articles that pass the relevance threshold are clustered by semantic similarity to group coverage of the same story across sources.

TRENDING TOPIC INTEGRATION

Clusters are ranked using a weighted combination of civic actionability (40%) — whether officials are named and how closely the story resembles the ingested corpus of civic documents — coverage breadth (35%), how many independent newsrooms cover the story, and trending relevance (25%), whether the topic aligns with what the public is actively discussing. Actionability leads because it is what makes an issue something a citizen can act on; trending is weighted least because it is the most volatile of the three signals. Trending signals are drawn from Google Trends and policy-relevant Reddit communities, cross-referenced with the news clusters via embedding similarity.

NON-PARTISAN SUMMARIZATION

The top-ranked issues are summarized by the LLM with explicit instructions to present objective facts, avoid opinion or editorial framing, and recommend actions that do not assume which side of an issue the reader supports. Recommended actions include contacting representatives, attending public hearings, and reviewing primary source documents — not advocating for or against any policy position.

CROSS-REFERENCING

When a ranked politician is involved in a trending issue, the Action Center links directly to their scorecard. Related government documents from the Explore database are matched using semantic search. Source articles include direct links to the original reporting.

GOVERNMENT ACTIVITY TABS

Dedicated tabs for all three branches of government — Legislative (Senate and House), Executive, and Judicial — display the most recent government documents: floor speeches, executive orders, proposed rules, court opinions, and notices, pulled directly from the Explore database.

NATIONAL MONITORS

When an issue persists in the news across multiple days, the system automatically creates a National Monitor — a dedicated tracking page for that ongoing concern. Monitors build a sourced timeline of developments, detect when separate news stories are facets of the same underlying event using embedding similarity, and merge duplicate monitors automatically. Monitors transition to "watching" status when coverage subsides and reactivate when new developments appear.

YEAR-IN-REVIEW TIMELINE

Each day's top issue is permanently recorded in a timeline that accumulates throughout the calendar year. The Timeline tab provides a month-by-month chronological view of what mattered most, with top policy themes calculated for each month and the year as a whole. At year's end, this becomes a complete "Year in Review" of the issues that shaped civic life.

ELECTIONS TAB

The Elections tab displays upcoming election dates, Senate races with incumbent scores linked to their scorecards, and an interactive U.S. map for selecting states. State-specific information helps users understand their local races in the context of national trends.

INTERACTIVE GLOBAL NEWS MAP

The World tab features a 3D interactive globe that visualizes U.S.-related international news coverage. Countries mentioned in current news feeds are highlighted with points scaled by article count. Clicking a country scrolls to recent headlines about U.S. relations with that nation, linking to the original source articles.

HOW IT WORKS

We implement a nearest-centroid classifier (Rocchio 1971) [22] Manning, Raghavan & Schütze 2008in sentence-embedding space. Each party's known platform positions on each policy area (taxes, healthcare, environment, etc.) are embedded as centroids using the same sentence-transformer model used throughout the pipeline. Bill text is then embedded and compared to both party centroids via cosine similarity.

The stance direction(pro/anti) disambiguates cases where both parties have positions on the same topic: a "pro" environment bill (strengthen EPA enforcement) aligns with the Democratic platform, while an "anti" environment bill (roll back regulations) aligns with the Republican platform. This encodes the saliency-plus-direction model from manifesto research. [23] Laver & Garry 2000

TWO-SIGNAL FUSION

Content analysis is the primary signal for party alignment. Vote tallies from roll-call data serve as a secondaryrefinement. When both agree, confidence is high. When they disagree, content wins unless the vote data shows a clear party-line split (which is itself informative — the bill was important enough to whip). This follows Snyder & Groseclose (2000) who demonstrated that vote outcomes reflect party discipline as much as ideology. [24] Snyder & Groseclose 2000

ADAPTIVE LEARNING

Platform position descriptions are seed prototypes bootstrapped from published party platforms. As the pipeline processes bills, sponsor party data from Congress.gov serves as supervised ground truth — bills sponsored by a single party are labeled examples that refine the classifier over time. This follows the self-training paradigm. [25] Yarowsky 1995

BILL POLICY AREA CLASSIFICATION

Bills and votes are classified into 18 policy areas (healthcare, defense, energy, etc., plus a procedural catch-all for non-substantive motions) using a tiered adaptive strategy:

• 1.Learning store exact match — bills classified in prior pipeline runs are recalled instantly by ID. This is the experience replay pattern. [10] Lin 1992
• 2.kNN against reference corpus — the k=7 most similar previously-classified bills in the vector store are retrieved and the policy area is assigned by similarity-weighted majority vote. [9] Cover & Hart 1967This is retrieval-augmented classification: the reference corpus grows with each pipeline run, improving accuracy over time. [26] Lewis et al. 2020
• 3.Embedding similarity against policy descriptions — cosine similarity between bill text embeddings and pre-computed policy area description embeddings. This is the cold-start fallback using nearest-centroid classification. [7] Reimers & Gurevych 2019

The policy taxonomy is based on the Congressional Research Service (CRS) policy area scheme used by Congress.gov. The approach follows the text-as-data paradigm reviewed in Grimmer & Stewart (2013). [27] Grimmer & Stewart 2013Stance derivation (pro/anti/neutral) uses embedding cosine similarity against direction prototypes — the bill text is compared to semantic signatures of supportive, restrictive, and reform-oriented legislative language, following the Comparative Agendas Project coding tradition. [28] Baumgartner & Jones 1993Zero LLM calls are used for bill classification.

DONOR AND INDUSTRY CLASSIFICATION

Donor classification uses a five-tier strategy:

• 1.FEC metadata — structured fields from the Federal Election Commission API encode committee type and designation codes, providing ground-truth classification for PACs vs. individual donors.
• 2.Semantic detection — embedding cosine similarity against category prototypes replaces ~200 lines of hardcoded string patterns. This generalizes to unseen entities because distributed representations capture semantic meaning. [29] Bengio et al. 2003
• 3.Learning store lookup — previously classified entities are recalled instantly by name.
• 4.Embedding cosine similarity — donor names are compared against pre-computed industry description embeddings. Industry descriptions include exemplar company names as anchoring tokens, following the zero-shot classification setup. [30] Yin, Hay & Roth 2019
• 5.k-Nearest Neighbor (kNN) — remaining unclassified donors are classified by the k=7 most similar already-labeled entities using distance-weighted majority voting. [9] Cover & Hart 1967This mirrors prototypical networks for few-shot learning [14] Snell et al. 2017where classification is performed by comparing query embeddings to accumulated real examples.

The kNN approach was chosen over LLM-based classification after empirical testing showed the LLM hallucinated invalid categories (producing labels like "SPORTS" or "RESTAURANT" outside the valid taxonomy) and was orders of magnitude slower. The kNN classifier processes ~5,000 donors in under 5 seconds versus 40+ minutes for the LLM, with more consistent results.

LEARNING STORE AND ADAPTIVE CLASSIFICATION

All classifications are persisted in a learning store (SQLite table) that functions as an evolving knowledge base. On subsequent pipeline runs, previously classified entities are retrieved instantly without recomputation. This is analogous to experience replay in reinforcement learning [10] Lin 1992 — past decisions inform future ones, improving both speed and accuracy over time.

The learning store also feeds into the self-training loop [25] Yarowsky 1995 — high-confidence classifications from prior runs become labeled examples for kNN and reference corpus retrieval in future runs. The system literally gets better each time the pipeline runs.

To prevent stale data from persisting when analysis algorithms are updated, the pipeline implements version-aware artifact management. At the start of each run, a SHA-256 fingerprint of all analysis source files is compared to the stored hash from the previous run. If the code is unchanged, all learning data is preserved to promote self-training. If the code has changed, stale artifacts (LLM results, learned classifications, kNN reference corpus) are automatically cleared so updated algorithms start fresh. The API cache (raw data from government APIs) is never cleared.

SEMANTIC SEARCH (EXPLORE)

The Explore feature uses dense passage retrieval [11] Karpukhin et al. 2020 to enable free-text search over government documents — Senate and House floor speeches, presidential actions (executive orders, proclamations, memoranda), Supreme Court opinions, and Federal Register rulemaking documents. Bill text is notindexed here; it is used separately, title-only, for the kNN bill-classification step in the scoring pipeline. Each document gets a single embedding (no chunking) over its title, summary, and first 800 characters of body, encoded with Snowflake all-MiniLM-L6-v2 and stored in sqlite-vec for nearest-neighbour retrieval. This outperforms keyword search (BM25) for conceptual queries like "climate policy" where exact term overlap is low.

EMBEDDING MODELS (CLASSIFICATION + SEARCH)

Two sentence-transformers, both 384-dimensional and both around 22M parameters. Snowflake Arctic-XS handles classification: bill policy areas, donor industries, party alignment, motion types, and the k-nearest-neighbour reference corpus. all-MiniLM-L6-v2 [8] Wang et al. 2020handles the semantic-search index and the Action Center's similarity gates. Sentence-transformers produce dense vector representations where cosine similarity correlates with semantic similarity [7] Reimers & Gurevych 2019 — making them ideal for classification-by-comparison tasks where category definitions exist.

The split is measured, not accidental. Arctic is retrieval-asymmetric: it packs same-register text into a narrow raw-cosine band (~0.55-0.87), which left several similarity thresholds unable to separate genuine matches from noise. Against this platform's own live failure cases, all-MiniLM-L6-v2 measured roughly 4x the separation margin on document anchoring and 3x on policy relevance. Classification stays on Arctic because its thresholds were calibrated against that model's geometry — moving a classification gate to a different embedding space without re-measuring the threshold is how thresholds quietly stop meaning anything.

LLM (NARRATIVE SYNTHESIS)

LFM2.5-1.2B-Instruct via llama.cpp [16] Gerganov 2023 handles tasks requiring natural language understanding and multi-step reasoning:

WHAT AI DOES NOT DO

WHY THESE TECHNIQUES WERE CHOSEN

We follow a strict hierarchy: structured metadata first, then embedding similarity, then kNN, then LLM — reserving each more expensive technique only for tasks the cheaper ones cannot handle. The pipeline contains zero hardcoded keyword lists, regex patterns, or string-matching heuristics for classification decisions. Every classification is made mathematically via embedding cosine similarity against natural-language prototypes. [12] Jurafsky & Martin 2023

Embeddings (not LLM) for classification: sentence embeddings excel at text classification tasks when labeled examples or category descriptions exist. They are deterministic, fast, and avoid the hallucination risks inherent in generative models. [13] Minaee et al. 2021The kNN classifier further leverages accumulated labeled data as a growing reference set — a well-established approach in few-shot and semi-supervised learning settings. [14] Snell et al. 2017

Content analysis (not votes) for party alignment: roll-call votes confound ideology with legislative strategy. Analyzing what a bill does relative to published party platforms recovers ideological alignment more accurately, following the manifesto analysis tradition. [31] Laver, Benoit & Garry 2003

LLM for narrative synthesis: tasks like promise-vote cross-referencing and PAC identification require world knowledge and multi-step reasoning that embeddings alone cannot provide. These are inherently generative tasks suited to language models. [15] Wei et al. 2022

MODEL AND ARCHITECTURE

The inference model is LFM2.5-1.2B-Instruct, a compact open-weight language model running natively via llama.cpp [16] Gerganov 2023 compiled with ARM-specific optimizations (cortex-a76, dot-product, fp16). This provides faster inference compared to containerized runtimes, generating ~14 tokens/second on the Raspberry Pi 5 CPU. Results are cached in a local database so each unique analysis is computed at most once.

The embedding models are Snowflake Arctic-XS for classification (bills, donors, industries, party alignment) and all-MiniLM-L6-v2 [8] Wang et al. 2020 for the search index and similarity gates — both 384-dimensional, both in the 22M-parameter class, so carrying two costs little. Vectors are stored in sqlite-vec, a single-file SQLite extension that runs on the Pi with no separate vector server. Every model runs entirely on-device with no external API calls.