A weekly-updated dashboard scoring every member of Congress on their independence from party leadership and bipartisan consensus.
Live site: https://congressionalindependence.deathbutt.org
Nothing here is pinned to one Congress. Which one is in scope is derived from the date — the 1st convened in 1789 and each runs two years — so the weekly job rolls itself over on January 3 of every odd year. A Congress convenes weeks before Voteview publishes its first files, so until that data lands the build falls back one Congress and keeps serving the outgoing one, then switches on its own. VOTES_CONGRESS pins a specific one for rebuilding an old term; party_code is normalised on the way in because older files ship it as a float (200.0), which would otherwise leave every member of an older Congress uncaucused and unscored.
Each member receives an Independence Score: how often they voted against their party majority on contested partisan votes, weighted by how united the party was. Breaking ranks on a vote where your party was unanimous is real independence; being on the losing side of a 50/50 party split is not. Each vote's weight is (party's share on the majority position - 0.5) x 2 — 1.0 when the party votes as a bloc, 0.0 when it's evenly divided. The plain unweighted number is still published as party_unity_pct if you prefer it.
Consensus deviation — how often they voted against bipartisan consensus (both parties agreed but they didn't) — is published beside the score, not inside it. It used to be averaged in, one half each. That was wrong: the two run on different scales, with a median of 3.5% against the weighted partisan median of 0.6%, so the mean tracked consensus deviation at r=0.96 and party discipline barely moved the ranking at all. Blending them made the average member look three times more independent than their voting record supports. They are different behaviors and they get separate columns.
The consensus column carries its denominator (62% of 67) because the chambers do not supply comparable ones. Most of the House's bipartisan rollcalls are suspension-of-the-rules bills — the classic noncontroversial passage vote — and it holds roughly three times as many as the Senate, whose bucket is mostly cloture motions and motions to proceed that happened to align. One Senate defection therefore moves the rate about three times as far, and senators post consensus deviation over 50%. (As of the 119th's 2026-09-06 build: House 191, of which 130 suspension bills; Senate 67. Those counts grow every week — the ratio is the durable part.) Filtering the bucket down to substantive passage votes was measured and rejected: it barely moves the House (191 -> ~148) and cuts the Senate to ~21, making the thin denominator thinner, and it would mean matching vote_question strings that Voteview can rename at any time. Instead the number is published with its denominator attached, and sorting that column groups the chambers rather than ranking them against each other.
| Tier | Score | Label |
|---|---|---|
| 0 | < 0.5% | Mindless Drone |
| 1 | 0.5–1.5% | Bobblehead |
| 2 | 1.5–4% | Squeaky Wheel |
| 3 | 4–8% | Free Agent |
| 4 | 8–20% | Heretic |
| 5 | 20%+ | Lone Wolf |
The cuts were measured against the 117th, 118th and 119th: every band is occupied in all three, the top stays rare (5–7 members a term), and the most loyal band is still the largest so the pile-up at zero remains visible. The previous cuts were calibrated for a score that averaged in consensus deviation, and against the current one they left ~90% of Congress in two bands. Every name reads the same direction — the scale disparages loyalty and credits deviation — which is why tier 3 is Free Agent and not Loose Cannon, the one rung that insulted the dissenter.
The bands live in exactly one place — TIERS in analyze_votes.py — and ship in data.json as tiers. The site builds its scale legend, filter options, histogram axis, ranges and every badge colour from that table, so renaming a band is a one-line edit and the browser test reads the names back out of the Python to prove it. Members carry independence_tier — the id only, since tiers is in the same file and a stored copy of a derived name is one more thing that can drift. Snapshot distributions are archived as label_dist, a list indexed by tier id rather than a dict keyed by name: the archive is the one thing that cannot be regenerated, and it has to survive a rename.
Attendance is tracked separately and does not feed into the score. Over every rollcall their chamber held between their first and last recorded vote, missed % counts the ones with no Yea or Nay from them — Present, Not Voting, and the ones Voteview has no row for at all. That last case matters: the Speaker votes at his own discretion and is simply absent from a large share of House rollcalls, which a row count would score as perfect attendance. Bounding by first and last vote keeps members who arrived or left mid-congress from being charged for votes held outside their service. Showing up is not the same thing as being independent, so the two numbers stay apart.
Yea and Nay each cover a range of Voteview cast codes (1–3 and 4–6); all of them are counted. Only decisive Yea/Nay votes go into the score denominators. Scores are cumulative over the whole term rather than a rolling window, which is why they barely move week to week. Members with fewer than MIN_VOTES (30) recorded votes are left unscored — too small a denominator to mean anything.
The table shows nine columns; party_unity_pct and the per-vote detail stay in data.json and on each member's page rather than widening it further. Fields nothing reads are not published: consensus_loyalty_pct was 100 - consensus_deviation_pct, weighted_partisan_deviation_pct became a duplicate of the score when the score stopped being a blend, and district was never rendered.
Floor leaders are flagged, not excluded. The Speaker plus each party's leader and whip in each chamber carry a leadership field; the table badges them and the filter can hide or isolate them. They schedule the votes they then vote on, so their loyalty is partly loyalty to an agenda they set themselves — in the 119th the Speaker came out the single most loyal member of his caucus, on the subset of votes he chose to cast. Excluding them outright would drop the first names anyone looks up, so the call is left to the reader.
docs/changes.html is a standalone page explaining a methodology change to people who saw the previous numbers, linked from a banner that retires itself on a date set in init(). It is not a running changelog — it gets rewritten the next time the numbers move for a reason a reader would notice.
Three suites, no dependencies between them and none installed for them:
| What it covers | |
|---|---|
python3 tests/test_scoring.py |
The scoring pipeline against a synthetic Voteview fixture, plus every upstream-drift path |
node tests/test_site.mjs |
The page's functions against a stub DOM — sorting, filtering, the trend gate, the tier table |
node tests/test_browser.mjs |
The real page in a real browser: CSP violations, console errors, failed requests, what actually rendered, and the changes page |
The browser check exists because the other two cannot see what the browser decides. It builds the fixture site, serves it over HTTP so 'self' means something, and drives a headless Chromium over the remote-debugging protocol using node's built-in WebSocket — no browser-automation dependency and no downloaded browser binaries. It picks up Chrome, Chromium, Edge or Brave, in that order; CHROME_PATH overrides. It gates the weekly publish, which force-pushes the site with nobody watching.
Each weekly run archives a summary-only snapshot to docs/history/, tagged with its Congress. There is no per-member series: it was measured, and the median member moves 0.02 points week over week while the only large swings are new arrivals thrashing on a small denominator. The signal is at caucus level, so that is all the snapshots keep — about 110 lines a week instead of 4,500.
The Caucus drift chart plots average independence per caucus over time, House and Senate as separate panels. It appears on its own once the archive actually covers a term: at least three snapshots with scored members, the first within 90 days of the Congress convening, and no gap longer than 45 days. The gate is coverage rather than a Congress number because the problem it guards against is real: a cumulative average that had already converged before archiving began is a flat line by construction. Any term archived from its start gets a chart; one the archive only caught the tail of does not, and neither case needs a code change. (Archiving here began partway through the 119th, so that term shows no chart.)
The Speaker is derived, not configured. It is the only leadership post decided by a recorded vote, so Voteview already has it: vote_question is Election of the Speaker and the winner sits in vote_result as Surname (ST). The build takes the most recent such vote and matches it to a member, which means a mid-Congress replacement — 2023's McCarthy to Johnson — resolves with no edit at all. If the name doesn't match exactly one member, the Speaker goes unflagged and it is reported rather than guessed.
The other four posts per chamber are not derivable at any price: both parties elect their floor leaders and whips in closed conference and caucus meetings that never produce a rollcall. Election of the Speaker is the only leadership question Voteview records, and the Senate has no equivalent at all. Those four stay in LEADERSHIP_BY_CONGRESS, keyed by Congress because ICPSR numbers follow the person and not the post — one flat table carried forward would keep calling a former whip "Whip" for the rest of his career. An unlisted Congress leaves them unflagged and files an issue, which is how each new term gets the table filled in.
Voteview is a third party that can change its files whenever it likes, so the build checks what it gets. A missing or renamed column stops the run — there is no sensible way to score without it. A chamber, party_code, or cast_code the script doesn't recognise is the more dangerous case, because nothing would otherwise go wrong: an unknown cast code just falls into the "missed" bucket and the index is quietly incorrect. Those are recorded and the build still publishes.
Anything unexpected lands in schema-drift.txt, and the weekly workflow turns a non-empty report into a GitHub issue — commenting on the open one rather than filing a new issue every Monday. A clean run deletes the file, so a fixed problem stops nagging. Two maintenance conditions ride the same channel: an unmapped member who needs a caucus decision, and a Congress with no leadership table. That last one is what makes the January rollover a task in the tracker rather than a silent omission. Cases already reviewed and accepted go in ACCEPTED_UNSCORED so the weekly run stays quiet.
Data sourced from Voteview.com. Sanders and King are scored against the Democratic caucus they align with, but displayed as Independent. Anyone else Voteview codes as Independent has no caucus to measure deviation against and is left out of the index — the build prints their names so the omission is never silent. The President appears in Voteview's vote files (announced positions are recorded as cast codes) and is excluded entirely. So are the six non-voting delegates (DC, PR, VI, GU, AS, MP): barred from final-passage votes, their record is a Committee-of-the-Whole subset of a few dozen votes against a chamber norm of hundreds, and the resulting "missed" rate — well past 80% — is the franchise rather than attendance. Ranking them beside voting members would compare two different things, so they are dropped from the index and from the party-cohesion tallies — the build prints their names.
This one is already deployed; the steps below are for standing up a fork.
git init
git add .
git commit -m "Initial commit"
git remote add origin https://github.com/<your-username>/<repo-name>.git
git push -u origin main- Go to repo Settings → Pages
- Source: Deploy from a branch
- Branch:
site· Folder:/ (root) - The
sitebranch is created by the Action on its first run — run the Action before setting this. - Click Save
- Go to Actions → Update Voting Index
- Click Run workflow
After ~30 seconds the built docs/ tree is force-pushed to the site branch and the site is live. The bulk outputs (data.json, members/) are gitignored on main and only ever exist on site; the only thing committed back is docs/history/, which cannot be regenerated. It goes to the history branch, not main, because main requires the validate check on every push and an unattended bot commit can't have it.
The Action runs automatically every Monday at 8am UTC thereafter.
uv run --with requests python3 analyze_votes.py # generates docs/data.json, docs/members/, docs/history/
python3 tests/test_scoring.py # scoring self-check, no network needed
node tests/test_site.mjs # site self-check: empty-Congress render, trend chart
cd docs && python3 -m http.server 8000
# open http://localhost:8000The author has zero background in data science, python, or anything else. This may all be hallucinatory AI slop. Enter at your own risk. ooooh, scaaaaary...