Methodology

Definition

Reader Tax is the percentage of reading effort the intended reader could skip without a material loss, given what the piece is for.

ReadMini first works out what the piece is for (to inform, explain, instruct, argue, evaluate, investigate or tell a true story) and who it is written for (a general reader, a beginner, a practitioner or an expert). What counts as material depends on both: a technical detail can be skippable for a general reader and essential for a practitioner. Fiction and poetry are not scored.

Formula

Reader Tax = 1 − (minimum sufficient word count / original meaningful word count)

• A model decides, passage by passage, what to cut, edit, move or keep. Every decision points at the exact passage it applies to.
• A second model checks those decisions both ways: for claims, numbers, evidence and caveats that were lost, changed or invented, and for passages that were kept but add nothing. It can only restore, correct or add decisions, before anything is counted.
• Application code builds the minimum sufficient version from the checked decisions, counts the words on both sides the same way, and calculates the percentage. No model outputs the number.
• Page furniture such as navigation, sign-up boxes, share prompts and comments, and decorative images, is not counted as meaningful words.
• The article keeps its own structure: headings, lists, links, emphasis, tables, code and informative images. Code, tables and images are never rewritten, and an image counts only its caption.
• The result is shown as a whole percentage.
• Reading time assumes 225 words a minute.

What affects the score

• Repetition, including restating a point or decision already made
• Unnecessary setup
• Over-explanation
• Redundant examples
• Tangents
• Empty phrasing
• Unnecessary complexity, where the same meaning fits in fewer or simpler words

Passages in the wrong place are pointed out as moves. A move never changes the percentage.

Reading lengths

Mini is the analysis itself and the only version with a Reader Tax. Longer and Shorter are derived from the same decisions, without another model call.

• Longer applies only the cuts and edits that remove filler, and keeps the examples and detail Mini leaves out.
• Shorter also leaves out prose that carries no core point and no caveat, instruction, decision or evidence.
• Shorter never leaves out code, tables, essential images or core points that Mini keeps.

What does not affect the score

• Whether the text was written by AI, or how likely that is
• Who the author is
• Whether ReadMini agrees with the author
• Political viewpoint
• Popularity
• Grammar or spelling on their own
• Em dashes
• AI-writing clichés on their own

Separately, ReadMini notes reading friction: content that is needed but harder to follow than it has to be. It is shown on its own and never changes the percentage.

Limitations

• ReadMini uses models to decide what is materially important. Those judgments may be wrong.
• Inspect the passage-level decisions in Writing mode or on “How this was scored” rather than trusting the percentage alone.
• Longer and Shorter follow fixed rules on top of the analysis. They are not separately checked.
• ReadMini does not check whether the article is factually correct.
• A low Reader Tax does not mean the writing is easy to read or good. It means little could be removed.

Reproducibility

The same text returns the same stored result. Models run with deterministic settings where the provider supports them. Changing the score definition, the prompts, the models or how text is extracted and split into passages starts a new methodology; earlier results are never recalculated.

Score version

readertax-v1

Analyst

analyst-v8

Verifier

verifier-v6