How SI drafts, how drafts are checked before publication, and how scores are computed.

01

“Constitutional Solutions”

A VFA SI Solution is a drafted alternative to one official action. There are three kinds:

• Bill. A bipartisan working alternative that stays within Congress’s enumerated powers.
• Executive action. An assessment of whether the action is within the President’s Article II authority, followed by either the statute Congress would need to pass or a corrected order.
• Supreme Court opinion. A textualist re-decision: how the case would come out reasoning from the constitutional text.

The model is given the official record and nothing else: for a bill, the Congress.gov summary, explicitly flagged as a summary and not the full bill text, plus its latest action; for an executive action, the Federal Register text; for an opinion, the opinion text. Bills that have already become law or been sent to the President are not drafted against, because a summary is not enough to analyze final text.

Most drafts are written by a Muse model running on our own worker. If that worker does not finish, a backup model (DeepSeek, via OpenRouter) drafts instead, within a small daily cap. Each draft records which model produced it. A draft stays private until it passes the publication checks below.

02

“The Constitution citation gate”

The Constitution is the fixed backbone of every SI lane: Solutions, Founders’ Verdicts, Think Tank proposals and constitutional grades. Before any of them can be published, software checks that the output:

• cites at least one specific, real provision: an Article with a Section (Articles I–IV), Article V, VI or VII, or an Amendment from the First through the Twenty-seventh, with valid section and clause numbers;
• cites no provision that does not exist, such as “Article VIII”, a “28th Amendment” or “Article I, Section 11”;
• uses our house term “SI” for machine-generated work in published text.

Solutions must also meet a minimum length and structure, include the operative provisions required for their type, carry a note on what source they relied on and a note listing any claims they could not verify, and describe a path to passage without forecasting vote counts. These are checks of shape and disclosure. They do not prove the legal analysis is right, which is why every item is labelled as SI-generated and links to its source.

03

“Source handling and verification”

Records on officials come only from public, official sources, ingested on a schedule: Congress.gov, House and Senate roll calls and the Congressional Record daily; the Federal Register and CourtListener daily; Voteview, House periodic transaction reports and FEC receipts weekly. The full list, with what each is used for, is on the sources page. Each record links back to its source.

Before a Solution is published, the server confirms that its official source is available: the Congress.gov summary for a bill, the Federal Register text for an executive action, or the opinion text for a Supreme Court case. Think Tank proposals pin the exact text they quote and are re-checked against the source daily. Changes to constitutional grades, Founders’ Verdicts and official profiles are written to a hash-chained audit log that is verified every night.

Campaign-finance data on member profiles shows committees and organizations only. Contributions from private individuals are never aggregated or displayed.

04b

“The SI Editorial Standard”

Every Solution, Founders’ Verdict and SI Senate measure must meet one written standard before it is published. The Constitution comes first: the argument starts from the text and original meaning (enumerated powers, federalism and the Tenth Amendment, the Bill of Rights, separation of powers, limited government, fiscal restraint). It must also be genuinely bipartisan: something a principled conservative and a good-faith progressive could each vote for, with a concrete gain named for both.

• Checked by software: real constitutional citations; required cost, sunset, oversight and enforcement provisions; no partisan labels, insults or election advocacy; a gain named for both sides; plain English (reading level checked); no figures that are not in the official source; not a copy of the source.
• Checked by an independent SI critic: a separate SI pass scores each item from 0 to 5 on source grounding, constitutional support, balance, stage accuracy, honesty, originality, clarity, and constitution-first persuasive clarity. Grounding, constitutional support, balance, stage accuracy and honesty must score at least 4; the rest at least 3; the average at least 3.8. Any unsupported claim, misapplied citation, or partisan phrase fails the item.
• One repair, then reject: a failing item gets one SI rewrite and is checked again from the start. If it fails again it is not published, and the reasons are recorded.
• Checked again live: the official source is re-read just before publication. A bill that has become law is never presented as pending.
• After publication: sources are re-checked daily. If a source changes, the page says so; if it is removed, the item is withdrawn. Daily caps and an emergency stop limit what can go out.

05

“Founders’ Verdicts”

A Founders’ Verdict (also called the Framers’ verdict) is a structured debate about a bill. Four voices, modeled on the published reasoning of Alexander Hamilton, James Madison, Thomas Jefferson and John Marshall, each take a position (YES, NO or CONDITIONAL) with constitutional citations. A neutral clerk then writes a synthesis and a verdict: CONSTITUTIONAL, UNCONSTITUTIONAL or CONDITIONAL.

The Framers’ confidence (0 to 1) measures agreement among the four voices: 1.0 when all four agree, lower when they split. It is not a probability that the verdict is legally correct.

These are simulations of historical reasoning produced by SI. They are not statements by the Founders or by any living person. Every voice and the synthesis must pass the Constitution citation gate and the other publication checks above.

The SI Senate works the same way for new legislation. Three SI seats (conservative, progressive and independent; never personas of real people) take up the concern behind one pending bill and pass one measure of their own, which needs the votes of both the conservative and the progressive seat. Each measure is published only after it passes the standard above, with its own critic check that the seats’ votes and statements match the measure.

06

“SI Think Tank”

The SI Think Tank looks for principles shared across Solutions and turns them into broader proposals. Its inputs are pairs of published Solutions from the past year. For each candidate, the model either proposes or abstains. A proposal is published only if:

• every source it relies on is quoted word for word (at least 50 characters) and the quote matches the official text;
• a separate, skeptical critic pass approves it;
• it passes the Constitution citation gate and the “SI” wording check.

Like all SI output, Think Tank proposals are not individually reviewed by a person. Each proposal shows the state of its sources: current (checked in the last two days), source unverified, or source stale. Sources are re-checked daily; when a source changes, the proposal is marked stale and is never rewritten. Signed-in readers can upvote or downvote each proposal, one vote per account.

07

“What ‘Archived SI output’ means”

Some Solutions and Founders’ Verdicts were published under earlier versions of our process. Rather than delete them, we keep them on the record with the label “Archived SI output (machine-generated)”. Archived Solutions do not display a path to passage, and they are excluded from our public syndication feed. Check them against the original source.

08

“The constitutional grade”

Where the platform grades an action or a proposal, it applies five checks: enumerated powers (does the government have the authority?), the Bill of Rights, individual liberty (the burden is on the government to justify a restriction), decentralization (the four-question test on the principles page), and due process and takings. The result is a letter grade from A (constitutional) to F (a direct violation). Grades must cite the Constitution like every other SI output. F-graded content is published with a warning, never removed.

The Constitution Index

The Constitution Index applies this rubric (vfa-constitutional-grade/1.0) to every bill in the current Congress that has an official Congress.gov summary and to every executive order, using the Federal Register text. Each morning the newest measures are queued first, so a measure is graded within a day of its official summary or text appearing; then measures with recorded roll calls, then measures whose official text changed, then a steady backfill of the current Congress, most recently active first. Grades are drafted by our SI engine and must pass the SI Editorial Standard’s automated rules for grades before they are stored: a real, specific Constitution citation, no loaded partisan labels, insults or electioneering. Grades are not individually reviewed by a person.

Versioning. Every grade records the rubric version, the SHA-256 hash of the exact source text it graded, when that source was fetched, when it was graded, and whether the primary or the backup SI engine produced it. A grade is never overwritten. When the official summary or text changes, or the rubric version changes, the measure is graded again and the new grade supersedes the old one; both stay on the measure’s page with their permalinks.

Constitutional Voting Record v1. Each House and Senate roll call is matched to the measure it concerns and to that measure’s current grade. Only votes on passage or adoption count (passage, suspension and passage, agreeing to a resolution, concurrence, conference reports, veto overrides): a Yea is a vote for the measure and a Nay a vote against it. Cloture, motions to proceed, recommit or table, and amendment votes are listed as not counted, because a vote on them is not a vote for or against the measure. Nominations are not graded. A member’s record shows counts of Yea and Nay votes by grade band (A–B, C, D–F) with every vote linked to the official roll call and the grade. Shares appear only when a member has at least 10 counted votes across at least 5 graded measures. The record is a factual aggregate of public votes, not an endorsement. Since October 8, 2026 it is the basis of the Scoreboard (the Constitutional Voting Record ranking); it does not feed the profile accountability score.

Export. Grades and voting records are public as CSV and JSON under /api/constitution-index/ (grades.csv, grades.json, voting-records.csv, voting-records.json, coverage.json, and one grade or one member as JSON). Read-only, no key, 60 requests per minute. Attribution is required: cite “Values for America Constitution Index” and link the grade permalink. Each measure page has a “Cite this” block.

The Constitution Explorer

The Constitution Explorer gives every Article, Section, Clause and Amendment its own page. The text is the National Archives transcription. Each page lists every public VFA SI Solution, Constitution Index grade, SI Think Tank idea and SI Senate measure that cites the provision, found with the same parser the Constitution gate uses, plus the roll-call votes on measures whose current grade cites it. Counts cover published items only and update within minutes.

Explainers are SI-generated, a few each day, most-cited provisions first. Before one is published, code checks that every quotation is copied word for word from the provision text or from a founding-era primary source in our verified set (each checked against its public-domain source), that every Supreme Court case named comes from a verified case record, that the originalist and living-constitutionalist readings are both present at comparable length, and the language rules. An independent SI critic then scores sourcing, accuracy about the text, balance and honesty (each must reach 4 of 5), with one repair allowed. Explainers are not individually reviewed by a person. A provision without a published explainer still shows its text, sources and live citations.

Search and Ask VFA

Search covers every published SI Solution, SI Research Paper, Constitution Index grade, SI Think Tank idea, SI Senate measure and Constitution provision, plus the official bill records and members of Congress they rest on. Drafts, retracted items and anything hidden by the kill switch are never indexed. Filters cover type, provision, policy area and Congress; there is no party filter.

Ask VFA answers a question only from that published record. If our published work does not cover a topic, no answer is written: you see “We haven’t analyzed that yet” and the closest related work, and the question is counted as a topic suggestion for the SI Think Tank and Solutions queues. Otherwise a short SI-generated answer is written from the matching passages, and code checks it before it is shown: every [n] citation must point to a passage it was given, every quotation must appear word for word in those passages, it must name a specific provision of the Constitution that the cited work itself cites, it must follow our language rules and never describe its own process, and it must not give voting or candidate advice. Questions asking whom to vote for are not answered. Answers are cached for a week and are not individually reviewed by a person. Questions are logged without IP addresses or account ids.

“SI Research Papers”

Once a week the record chooses the question for an SI Research Paper: the provision of the Constitution that VFA’s public Solutions, grades, Think Tank ideas and SI Senate measures cite most often and that has no paper from the past year, together with the policy areas of the bills behind those citations. Code then assembles the evidence pack the paper may use: the provision’s text, VFA’s own citation counts, founding-era primary sources from our verified set, Supreme Court cases from our verified case record, the published Solutions and Think Tank ideas that cite the provision, and the official texts of the measures they answer.

The paper is written section by section over several days, within a fixed daily budget: background, constitutional analysis, policy options compared, a bipartisan path, implementation with cost and limits, and finally an executive summary of the approved sections. Before a section is accepted, code checks that every paragraph cites at least one item of the evidence pack, that every quotation is word for word from a cited source, that every case named is in the case record, that every figure appears in the evidence, that both schools of interpretation are present at comparable length, that one option is VFA’s proposal and is labelled as a proposal, the language rules, and grammar. An independent SI critic then scores sourcing, constitutional accuracy, balance, honesty, grammar and fluency, and standalone clarity (each must reach 4 of 5); one repair is allowed. When every section passes, a second critic reads the whole paper for coherence; it may send named sections back for one revision. Only then is the paper published, with a stable id (VFA-RP-YEAR-NUMBER), a citation block and a printable PDF. References are formatted by code from the evidence pack, not written by the model. Papers are machine-generated and not individually reviewed by a person.

09

“The Scoreboard: Constitutional Voting Record ranking”

The Scoreboard ranks members of Congress only on positive, sourced evidence: their recorded votes on the final passage or adoption of measures graded in the Constitution Index. Every vote it counts links to the official House Clerk or Senate roll call and to the published grade.

• Aligned votes. A Yea on a measure graded A or B, or a Nay on a measure graded D or F, is aligned with the grade. A Nay on A or B, or a Yea on D or F, is not. Measures graded C (significant issues) are neutral: the votes are shown but not scored. Procedural and amendment votes are never counted.
• Grade confidence. Each scored vote is weighted by how firmly the grade’s five rubric checks support the letter: the share of checks passed for an A or B, or failed for a D or F, never below 0.4 (one grave defect can carry a D or F on its own).
• Result. 90% weighted alignment share plus 10% attendance (the share of those graded roll calls on which the member voted Yea or Nay rather than Present or not voting), on a 0–100 scale.
• Findings. A statement–vote or conflict-of-interest finding that meets the evidence standard below lowers the result by 2 points, up to 10. Having no findings is neutral: it never adds anything.
• Minimum evidence. A member is ranked only with at least 10 scored passage votes across at least 5 graded measures. Everyone else is listed as “Not enough graded votes yet” with their count. That is a coverage gap, not a judgment; it shrinks as the Index grades more measures (up to 150 grades a day, measures with recorded passage votes first).
• No false precision. Results are shown as whole numbers with a 95% interval (Wilson interval over the confidence-weighted number of votes) and a quintile tier (top, upper, middle, lower, bottom). Members whose whole-number results are equal share a rank and are listed by evidence volume (scored votes, then measures), then name.

The ranking measures how a member’s votes line up with the published grades under one rubric applied to every measure. It is not an endorsement of, or opposition to, any member, party or candidate, and the grades themselves are SI-generated and labelled as such on every grade page. Recomputed daily (rank version constitutional-voting-record-ranking/2.0).

Election pause. Ranks and tiers are not shown within 60 days before a federal general election (the Tuesday after the first Monday in November of even years), and return the day after it. During the pause the Scoreboard lists the members who meet the evidence threshold alphabetically with their vote counts, and every member’s votes stay public on their page and in the export. The pause is a setting (SCOREBOARD_ELECTION_PAUSE, on by default; SCOREBOARD_ELECTION_PAUSE_DAYS, 60) kept on while the organization’s legal structure is settled.

Change log. Until October 8, 2026 the Scoreboard ranked members by the profile accountability score below. Once unevidenced findings stopped counting, 117 members tied at 100.0 because that score reads 100 when nothing adverse is found; the ranking measured missing data. It was replaced by the Constitutional Voting Record ranking, logged at /corrections.

Profile accountability score

Each member of Congress gets a composite accountability score from 0 to 100, recalculated daily from public records only. It combines these sub-scores:

Absence of data is never scored as bad. A sub-score with no underlying records is left out, and the remaining weights are rescaled so they add up to 100%. A member with no scored inputs at all gets no composite and is listed as “not enough public data yet”. Each profile has a Data Coverage Panel showing which sources are connected for that member; when coverage is below 60%, the profile shows a caveat. Several planned sources are not connected yet, so coverage is currently partial for everyone. The profile score is not used to rank members.

Evidence standard for findings. A finding that a member said one thing and voted another, or voted on a matter tied to their own finances, counts toward a score, or adjusts a Scoreboard result, only when it carries both (a) a verbatim quote from the source record and (b) a link to an official source: Congress.gov, the House Clerk or a house.gov site, the Senate or a senate.gov site, GovInfo, or the FEC. A finding without both is kept on file, marked as not meeting the evidence standard, and is not published; it never lowers a score or a rank, and the member is scored exactly as if it did not exist. Since October 8, 2026 this excludes all legacy model-written findings, none of which had quotes or source links. The recalculation is logged at /corrections.

Trust

“Error reports, corrections and the weekly audit”

• Report an error. Every Solution, Founders’ Verdict, SI Senate measure, Think Tank idea and grade has a “Report an error” form. No account is needed.
• Automated re-check. Each report starts a re-check of the item against its official source: the live source check, the code checks and the independent SI critic run again, judging the item only against its official source and the SI Editorial Standard (the report text is never shown to the critic). An automatically published item that fails a code check or two independent critic runs is unpublished; an older item that fails is flagged. Target: a result within 72 hours.
• Corrections log. Every unpublish, source-change flag, correction and retraction is recorded with the time, what changed and why at /corrections. Affected items link to their entry.
• Weekly audit. A random sample of the week’s automated publications, with each rule’s critic score, plus gate pass and reject rates and the top rejection reasons, is at /methodology/audit.

Participation

“Citizen Consensus”

Every public VFA SI Solution and SI Think Tank idea carries a one-tap vote: support, oppose or unsure. Results are at /consensus.

• One vote per account per item. You can change or withdraw it at any time; only your current vote counts.
• Verification tiers. Unverified: the vote is stored but not counted anywhere. Email verified (you clicked the link we emailed): counts in national results. District attested (you also told us your home ZIP code and confirmed you live there): counts in your state and congressional district too. This is self-attestation, not identity verification; we do not check addresses.
• Districts. ZIP codes are matched to 119th-Congress districts with the U.S. Census Bureau’s ZIP Code Tabulation Area to Congressional District relationship file (2020 ZCTAs, public domain). When a ZIP spans more than one district, you pick yours from the Census candidates. ZIP+4 is optional and does not change the match.
• Minimum sample. National, state and district results publish only at 30 or more verified votes; below that the page says “not enough verified votes yet” and shows no percentages.
• Margins. Each share shows a 95% sampling margin (1.96 × √(p(1−p)/n)). Participants choose to take part, so results describe verified VFA participants, not all voters or residents; the margin does not correct for self-selection.
• Never representative claims from unverified totals. Votes from unverified accounts are never published, in any total.
• Anti-brigading. Per-account and per-connection rate limits. A second account voting on the same item from the same browser, or a sixth account voting on it from the same connection within 24 hours, is held out of results. A sudden spike (at least 30 voters in an hour and five times the item’s usual hourly rate) is flagged for review; while it is open, votes from accounts younger than seven days are held. Held votes are stored and can be restored after review. Connection and browser identifiers are stored only as salted one-way hashes.
• Reasons. An optional reason (up to 280 characters) is checked automatically for links, contact details and profanity, and is never shown unless an editor approves it. Approved reasons appear without names.

Participation

“Follows and alerts”

Signed-in readers can follow bills, presidential actions, members of Congress, Constitution provisions, SI Think Tank topics, Congress.gov policy areas and their own district. Your alerts are at /account/alerts.

• Only public records. Every 15 minutes we look for new public events: a Constitution Index grade (or a change of letter on a re-grade), a published SI Solution, a new latest action on a followed bill (Congress.gov) or a new Federal Register status on a followed presidential action, a House or Senate roll call, a published adoption-tracker match, a new SI Think Tank idea or SI Research Paper, and an entry in the corrections log. Nothing private and nothing unpublished becomes an alert, and the automated-publication kill switch withholds alerts for automated items exactly as it hides them on the site.
• Matching. A grade, Solution, Think Tank idea or Research Paper matches a provision when it cites that provision or any part of it (following Article I catches a citation of Article I, Section 8, Clause 3), using the same citation parser as the Constitution gate. Policy areas and Think Tank topics are Congress.gov policy areas of the bills involved. “My district” means the U.S. Representative for the district on your account and both U.S. Senators for its state.
• Votes. A member’s alert shows how they voted, as recorded by the House Clerk or the Senate, and the measure’s current VFA Constitution grade when it has one. Quorum calls and roll calls older than 14 days are not sent.
• Adoption matches say only that new official text contains language matching a VFA SI Solution; they never claim VFA caused it.
• One alert per event. An event you follow through several items reaches you once. Emails batch every pending alert into one message on your schedule (as they happen, daily at 7 a.m. or weekly on Monday at 7 a.m. your time), never during your quiet hours, only to verified addresses, and every email has a one-click unsubscribe link. Grades, Solutions and Think Tank items in alerts are SI-generated and carry the same labels as on the site.

10

“Known limits”

• Summaries are not full text. Bill Solutions are drafted from the official Congress.gov summary, not the full bill.
• SI output can be wrong. Automated checks reduce errors; they do not eliminate them. Nothing here is legal advice.
• No per-item human review. Items are published after automated checks, not after a person reads them.
• Coverage is partial. Some sources, such as lobbying filings and state legislatures, are not yet public on the site, and scores reflect only the records we hold.
• Older output predates the current checks. Items labelled “Archived SI output” were published under earlier rules.

Found an error? Use “Report an error” on the item, or email hello@valuesforamerica.com with a link to the primary source. See the corrections log.