Methodology

Every statement enters the system the same way. One standardized input. Every model receives it identically.

• Statement text captured with source link, timestamp, speaker identity, and context
• Input standardized to a single format before processing
• No preprocessing varies between models. Every model gets exactly the same input.

All AI systems process the same input at the same time. No system can see another’s output.

• Outputs are never combined, averaged, or editorially rewritten before display
• Separation is structural, not optional. You always see the variance.

Every output is displayed exactly as returned. Sources and context travel with every surface.

• Context presented as supporting information and links. Not editorial judgment.
• No output is summarized, merged, or paraphrased before the user sees it.

Consensus is computed after all independent outputs exist. It never overrides them.

• Summarizes shared patterns across models and highlights where they diverge.
• Consensus is additive. It sits alongside individual outputs, never replaces them.

Every input, output, and consensus result is stored as an immutable record.

• Historical data preserved. Never overwritten, never retroactively edited.
• Any statement can be re-evaluated against its stored outputs.
• Continuous audit trail. Every record has a timestamp and source chain.

Every measurement links back to its origin. The system does not replace source reading.

• Sources are URLs to public records, official transcripts, or verified platforms.
• Path back to the original statement always preserved for independent verification.

Surfaces past statements by the same figure on the same topic, ranked by semantic similarity.

• Each statement is compared against everything the figure has previously said on that topic.
• Match strength scored 0.0-1.0 (semantic similarity). Tiers: Very High (≥0.90), High (≥0.75), Moderate (≥0.60), Low (<0.60).
• Match limits vary by tier. Core: 3, Pro: 5, Pro+: unlimited.
• Patterns measured, not interpreted. Recurrence surfaced. Meaning left to the user.

Maps rhetorical structure across five framing dimensions. Pentagon chart rendered per model.

• Five axes: Adversarial/Oppositional, Problem Identification, Commitment/Forward-Looking, Justification/Reactive, Imperative/Directive.
• Each axis computed independently per AI model.
• Describes rhetorical structure, not moral character. A high Adversarial score means oppositional language, not that the speaker is wrong.
• Variance between models on the same axis is shown. If models disagree on “Justification” vs “Commitment,” you see each one.

Every model’s output displayed side-by-side. Consensus computed after. Disagreement displayed, not resolved.

• Outputs include: primary framing classification, signal metrics, and contextual notes.
• Consensus layer identifies shared patterns. Variance layer identifies divergence.
• Agreement and disagreement are both treated as signal.

Measures how far each provision drifts from a bill’s stated purpose.

Scored 0-100.

• Each provision measured against the bill’s title and purpose clause by semantic distance.
• Drift tiers: Low (0-25), Moderate (26-50), High (51-75), Very High (76-100).
• High-drift provisions surface riders, earmarks, and thematically distant amendments.
• Source links to original bill text always provided.

Tracks how bill provisions change across legislative versions: introduced, committee, floor, enrolled.

• Provision-level diffs between version pairs: additions, removals, and modifications.
• Magnitude scoring (0–1) for each provision change based on text delta size.
• Aggregate mutation percentage across the full bill.
• Mechanical text diff computation. No AI opinion involved.
• Source links to original bill versions always provided.

Surfaces spending data tied to bills and provisions. Grounds legislative analysis in dollars.

• Two data sources: CBO scores (official estimates) and extracted dollar figures (from bill text).
• Per-provision spending with section identifiers and spending categories.
• Version-over-version spending deltas when multiple versions exist.
• Factual financial data only. No AI opinion involved.

Two figures on the same topic, same surface. Direct framing comparison without editorial selection.

• Shared-topic matching via semantic overlap on the same legislative or policy subject.
• Framing differences presented side-by-side. No “winner” declared.

Pulsing avatar ring on feed cards and figure profiles indicating recent signal activity level.

• Driven by statement volume and signal variance in the trailing window.
• Visual-only. No analysis required to read it.

Aggregate framing tendencies rendered as a unique visual signature per figure.

• Computed from historical Framing Radar™ axis averages.
• Each figure’s fingerprint is distinct and evolves over time.

Dot-based navigation between figures, topics, and framing patterns. Each node sized and colored by activity.

• Connections mapped from shared topics and framing similarity.
• Navigation, not analysis. Explore, don’t interpret.

Detailed variance breakdown inside Lens Lab™ when independent systems disagree.

• Highlights the specific words, phrases, and axis scores where models diverge.
• Disagreement is data. Shown in detail.

Shows shared framing and topic overlaps across figures and time on Statement Detail.

• Links statements by topic, framing signature, and timing.
• Tappable chips navigate to related statements.

Complete analytical profile for a single figure. Every measurement surface consolidated into one view.

• Aggregates The Receipt™ history, Framing Radar™ averages, signal trends, and vote record.
• Exhibit plate format: museum-grade presentation of longitudinal data.

Detects when independent figures begin using similar framing simultaneously. B2B exclusive.

• Measures framing similarity across figures within defined time windows.
• Convergence presented as a signal. Causation is not implied.

Four independent scores computed per statement by each AI model.

Each scored 0-100.

• Repetition: How closely language mirrors the figure’s prior statements on the same topic.
• Novelty: How much new language or framing the statement introduces versus established patterns.
• Affect: Rate of emotionally charged language. Intensity markers, urgency signals, sentiment-loaded phrasing.
• Entropy: Topical spread. Higher = multiple subjects. Lower = tight focus.

How far a statement’s signal metrics deviate from the figure’s historical average.

• Each metric measured against the figure’s own rolling average.
• Positive means above-typical signal. Negative means below typical. Zero means on baseline.
• Measures shift, not position. A high Affect score is context. A high Affect delta is signal.

How many models converge on similar measurements. Shown as a ratio.

• Convergence ring shows proportion of agreement (e.g., 2/3 models aligned).
• Computed only after all models return independently.

When models produce significantly different results, a variance indicator appears.

• Triggered when models diverge on primary framing classification or signal metric values.
• Variance banner displayed prominently. Not an error. Genuine measurement divergence.
• Disagreement displayed as data. Never suppressed.

Per-member, per-bill detail across the full congressional record.

• Every vote. Every member. Every bill. Displayed as recorded (teal) or not recorded (gray).
• Never color-coded by position. Neutrality is structural.
• Tracked separately from speech metrics. Voting behavior and rhetorical behavior are independent surfaces.

Tracks how a figure’s language patterns shift over time.

• Signal metrics, framing axes, and similarity scores plotted across statements over weeks, months, and sessions.
• Trends observed, not predicted. No forecasting.