Getting better

The practice loop

Rhetoric is built around four steps: prepare, perform, measure, improve. The Getting Started checklist, a small button in the corner of the workspace that opens a compact list, names them as five milestones: write or paste a speech, rehearse with the teleprompter, read your Full Audit, try a debate, and review your History. Completing all five once is orientation. Getting better means running the loop again, deliberately, with the numbers from each pass guiding the next one.

This chapter explains what each number means, which numbers to trust for what, and how to build a rhythm that moves them.

Reading a score as a baseline

Your first Full Audit score is a baseline, not a verdict. It places your speech on a 0 to 100 scale so you can see whether later versions move up, down or sideways. The score ring on the report page shows the Overall Persuasion Score: a fixed weighted blend of six AI-scored dimensions, computed in code by Rhetoric, not by the AI model.

Because the formula is fixed, scores from different runs and different speeches are directly comparable. A speech that scores 62 today and 71 after a rewrite has genuinely improved on the same rubric.

In your Library, every speech card shows a score pill once it has been audited. Green means the score is above 75. Amber means it is above 45. Red means it is below 46. A card without a score shows Draft, meaning no audit has been run yet.

The Overall Persuasion Score is computed by Rhetoric from the six dimension scores. The AI judges the dimensions. Rhetoric computes the number.

Which numbers to trust for what

Rhetoric produces two kinds of numbers. Measured numbers come from code running on your device: word count, pace in words per minute, duration, script coverage, departures and recovery time on the teleprompter, and the freestyle sub-scores for pace, coverage and timing. These are deterministic. The same input always gives the same output.

AI judgments come from the analysis engine: the six Full Audit dimension scores (Ethos, Pathos, Logos, Structure, Delivery, Caution), findings and severities, coaching notes from the Coach pane, the Originality freshness scan, the freestyle integration read, the debate verdict, and the Case Doctor critique. These are model outputs, anchored to rubrics and stabilised by a consensus mechanism. A borderline dimension may shift by a point on a re-run.

• Use measured numbers to track delivery habits: pace, timing, filler count, recovery speed, script coverage.
• Use AI judgments to find structural and rhetorical problems: weak evidence, missing emotional appeal, logical gaps, predictable phrasing.
• Compare measured numbers across sessions directly. Compare AI scores across audits of the same speech, or across speeches that serve a similar purpose.

Where the report shows a Measured pill, the number is computed on your device. Where it shows an AI judgment pill, the number is a model output.

The six audit dimensions

Each dimension is scored 0 to 10 by the analysis engine. The overall score blends them with fixed weights: Logos 20%, Structure 18%, Pathos 18%, Ethos 15%, Caution 15%, Delivery 14%. A low dimension pulls the overall score down in proportion to its weight.

• Ethos (Credibility and character): whether the audience will trust the speaker. Look at sourcing, credentials, and tone.
• Pathos (Emotional resonance): whether the speech connects with the audience's concerns. Look at stories, imagery, and stakes.
• Logos (Logical rigor and evidence): whether the argument holds together. Look at claims, data, and reasoning chains.
• Structure (Architecture and arc): whether the speech moves the audience through a progression. Look at openings, transitions, and closings.
• Delivery (Diction and speakability): whether the text reads well aloud. Look at sentence length, rhythm, and cue density.
• Caution (Argument integrity): whether the argument survives scrutiny. A low Caution score means the engine found fallacies or manipulative technique.

To move a dimension, read its findings in the report. Each finding quotes your text, names the severity (strength, minor, major or critical), and offers a rewrite. Apply the rewrites that make sense, then re-audit to see the effect.

Teleprompter numbers

The session report after a teleprompter rehearsal shows four measured values: duration, pace (words per minute), coverage (percentage of the script you reached) and departures (places where you went off script). If the session used voice tracking, all four are measured from your speech on this device. If it used the timed engine, pace and departures show No data because no voice was captured.

Recovery time is the number that improves fastest with practice. Each departure row shows how many seconds it took you to find your place again. A shorter recovery means you are learning the material, not just reading it. The report also counts rhetorical pauses and lapses, so you can separate deliberate silence from lost-place silence.

Coverage of 100% means you reached the last word, not that you read every word aloud. The tracker follows your position in the script.

Freestyle numbers

The speaking assessment after a freestyle drill shows a score from 0 to 100. It is a local blend of three measured sub-scores, each 0 to 10: pace (distance from the 120 to 160 words-per-minute persuasive band), coverage (how many of the three anchored ideas you wove in), and timing (how close you landed to the planned length). The blend weights are pace 34%, coverage 33%, timing 33%.

The freestyle score is computed entirely in code on your device. No AI is involved.

Below the score, the coaching notes section may appear in two forms. When the AI coaching service is available, the notes are phrased by AI from your measured numbers. When it is not, the notes are rule-based, computed locally with no AI involved. Both forms are labelled so you can see which you received.

The integration read, when it appears, is a separate AI judgment. It evaluates how the anchored ideas entered your unscripted flow: data framing, credibility and emotion balance, and the bridges between your own thoughts and the provided material. It carries no score. The only number on the assessment is the local blend above.

A session with fewer than 30 spoken words is not scored. Allow the microphone and speak for the full planned length to receive a score.

Debate numbers

The debate verdict screen shows a result (win, loss or draw) computed in code from the round-by-round scorecard. The judge, an AI, scores each round 0 to 10 for both you and the opponent. Rhetoric sums the round scores and compares the totals. The result is the comparison, not a model output.

Below the scorecard, two panels list what you did well and what to work on. These are AI judgments. They name specific moments in the transcript. On spoken debates, a section called How you spoke shows measured numbers: pace, filler words, response time, pauses, hedging and vocabulary. These are computed from the transcript and timings on your device. A Measured pill marks the section.

A weekly rhythm

One useful pattern is to rotate the three practice modes across the week. Each mode trains a different skill. The teleprompter trains delivery of prepared text: pacing, recovery, and coverage. Freestyle trains improvisation: weaving material into unscripted speech under a clock. Debate trains argumentation: responding to opposition, structuring rebuttals, and defending a position.

1. 1Early in the week, write or revise a speech in the editor. Use the Coach pane to get a quick analysis of the six dimensions.
2. 2Mid-week, rehearse the speech on the teleprompter. Note your pace, coverage and recovery times in the session report.
3. 3Run a freestyle drill on the same topic to practise speaking about it without the script.
4. 4Late in the week, set up a debate on the speech's motion. The verdict and the delivery section show how you argue under pressure.
5. 5At the end of the week, open History to compare session results across the week.

This is a suggestion, not a requirement. Adjust the order and frequency to suit your schedule. The point is that each mode surfaces a different weakness, and reviewing History shows whether those weaknesses are shrinking.

When to re-audit

Re-run the Full Audit after you have made substantive changes to the text: restructured sections, replaced evidence, rewritten a weak opening or closing. If a stale notice appears on the audit page, it means your draft has changed since the last audit and the report reflects the earlier version.

Do not re-audit after small edits like fixing a typo or rephrasing a sentence. The audit reads at full depth and uses one of your monthly allowance. A re-run on unchanged text is free because the server returns the cached result, so there is no cost if you want to reopen a report you already have.

What can go wrong

If your teleprompter report shows No data for pace and departures, the session used timed scroll, which does not capture voice. Open Stage settings and switch Tracking to Voice to get measured delivery numbers. The article "Troubleshoot teleprompter tracking and pace" covers engine selection and common problems.

If your freestyle assessment shows no score, either the microphone was blocked or fewer than 30 words were captured. Allow microphone access and speak for the full planned length. The article "Use freestyle drills" explains the setup.

If a Full Audit dimension changes by a point on a re-run of the same text, that is normal for a borderline score. The consensus mechanism stabilises the dimensions, but small fluctuations remain possible. Compare the overall trend across several audits rather than focusing on one dimension in one run. The article "The Rhetorical Audit, originality, and usage limits" explains allowances and stale reports.

If a debate verdict feels wrong, read the round-by-round breakdown. The per-round scores and notes explain what the judge weighted. The result is a sum of those scores, not a separate opinion. The article "Understand the debate modes" covers formats and scoring.