Interpretation guide
How ProseMap scores a sentence
ProseMap gives every sentence a strength score from 0 to 100 and every join between two sentences a handoff score from 0 to 100. It adds one judgment of the document's arc, then combines the three into an overall score. Scores measure whether the writing does its job, not whether it is smooth.
The four numbers
| Score | What it judges |
|---|---|
| Sentence | How well each sentence works where it stands. Judged in context, but not marked down for a poor transition; that is the handoff's job. |
| Flow | The quality of each handoff: the join a reader experiences between one sentence and the next, including across paragraph and section breaks. Only the last sentence has no handoff. |
| Arc | A single document-level judgment: whether paragraphs and sections progress, whether threads are resolved, and whether the piece coheres. |
| Overall | A weighted harmonic mean of the three (sentence 0.40, flow 0.35, arc 0.25), minus a penalty for sustained weak runs. A harmonic mean lets one weak dimension pull the total down further than a plain average would. |
The sentence rubric
Alongside the 0 to 100 score, each sentence is rated from 1 to 5 on five questions. The overall sentence score is a separate holistic judgment and need not equal their average.
| Dimension | What it asks |
|---|---|
| Clarity | Can a reader get the meaning in one read? Low when they have to stop and work out what is being said. |
| Precision | Does it say exactly what it means? Low when the words are vague or general: "things", "in a sense", "a variety of reasons". |
| Economy | Does every word earn its place? Low when the sentence is padded, repeats what was just said, or takes thirty words to say what ten would. It is not about being short: a long sentence with no waste scores well. |
| Control | Do the length, rhythm and structure feel chosen? Low when a sentence runs on, tangles, or breaks off by accident. A deliberate fragment scores high. |
| Function | Does it do its job in the paragraph? Every sentence has one: make the claim, give the evidence, turn the argument, land the point. Low when you could delete it and lose nothing. |
Bands and colours
The interface leads with bands, not numbers, because a 73 and a 74 are not meaningfully different. A band comes from the score and from what the evaluator flagged: a good score with a problem flagged is shown as mixed signals, not as a clean pass.
| Band | Meaning |
|---|---|
| Exceptional | 90+ with a strength flagged |
| Strong | Good score with a strength flagged, or 85+ with nothing flagged |
| Acceptable | 70+ with nothing flagged: does its job |
| Mixed signals | Mixed signals: a good score with a problem flagged, or a weak score with a strength flagged |
| Weak | Low score with a problem flagged |
| Failing | Score under 40: fails its purpose |
Colour is never the only signal. Bands are also carried by stroke pattern on the map and underline style in the text, and a colour-blind-safe blue and orange palette is available in settings.
Handoffs have a type, not just a score
Each handoff is assigned one of nine relations: advance, deepen, contrast, disrupt (controlled), restate, drift, pivot, payoff, sibling. The type is what separates smooth from going somewhere. A paragraph whose handoffs are mostly restate is flagged as circling even when every link is perfectly legible. The glossary defines each relation.
Weak runs and the penalty
A weak run is three or more consecutive sentences whose local score stays below 58. The local score is 0.6 times the sentence score plus 0.4 times the weaker of its two adjacent handoffs. Each run subtracts from the overall score in proportion to its length and depth, and the total penalty is capped at 12 points. The penalty is shown in the summary so the overall number can always be explained.
What guards against rewarding empty polish
The main risk in scoring prose automatically is rewarding generic fluency: text that is grammatical, smooth and says nothing. ProseMap guards against that in four ways.
- The evaluator is instructed to score function over smoothness, with fixed calibration anchors, and to cap empty polish in the unsteady band.
- Handoff typing makes restatement a visible category instead of hiding it inside a smooth score.
- Deterministic measurements, such as near-duplicate sentences and stock phrases, are shown as evidence and used to pick sentences for a second opinion. They never set the score.
- A benchmark of adversarial texts, including one that is fluent and empty and one made of deliberate fragments, must pass before the scoring prompt changes.
Confidence, audit and disagreement
Every sentence score carries a confidence value, and low-confidence sentences are hatched on the map. A second model re-scores a selected subset: the extremes, the low-confidence cases, the places where structure and judgment disagree, and a random sample. Where the two differ by more than ten points the sentence is marked contested and its confidence is lowered.
You can overrule any judgment with a band and a note. Your view is stored beside the evaluator's and never replaces it.
Structural preview versus model-backed analysis
When no model is available, ProseMap shows a structural preview computed from length, repetition and overlap. It is labelled as a preview on every surface and is not a judgment of quality. Only model-backed analyses score function and control.
Questions
How is this different from a grammar checker?
A grammar checker tests whether a sentence is correct. ProseMap judges whether it works: whether it is clear, specific, economical, controlled, and doing its job in the paragraph. It also scores the join between each pair of sentences, which grammar checkers do not look at. A correct sentence that adds nothing scores low; a deliberate fragment can score high.
Why does my writing feel flat even though every sentence is fine?
Usually because the sentences are not handing off to each other. The most common cause is restatement: each sentence says the previous one again in different words, so the prose reads smoothly and goes nowhere. ProseMap names each handoff, so a paragraph that is mostly restate or drift shows up even when every individual sentence scores as acceptable.
How do I find which sentences are weak in my draft?
The document map colours every sentence by band and underlines sustained weak runs, which are three or more consecutive sentences that stay under the working level. The hotspot list then ranks the places where one fix buys the most, starting with the longest weak runs and the weak joins between otherwise strong sentences.
Can it check the flow between sentences?
Yes. Every handoff gets a score from 0 to 100 and a relation such as advance, deepen, contrast or restate. On the map the gap between two sentence bars widens as the handoff weakens, so a break in flow is visible as a physical gap.
Is a higher score always better? Should everything be green?
No. Scores describe function, not polish, and a piece that is green from end to end is often a piece with no risks in it. An abrupt sentence or a hard cut between paragraphs can be exactly right, and the evaluator marks those as controlled rather than penalising them.
What if I disagree with a score?
Overrule it. Any sentence or handoff can be moved to the band you think it belongs in, with a note. Your view is stored beside the original judgment, never over it, and you can re-score the document under your view. You can also send the note back and ask the evaluator to reconsider; it must either revise or hold and say why.
How reliable are the scores?
They are an AI model's judgment and can be wrong. Each sentence carries a confidence value, a second model audits a selected subset, and sentences where the two disagree are marked as contested. Treat the map as a careful reader's opinion: useful for finding where to look, not a verdict.
Limits
ProseMap's scores and explanations are produced by AI models. They can be wrong, incomplete or misleading, and they are one reader's judgment rather than an objective measure of writing quality. Check anything that matters independently.