Check every document, and every summary or note made from one, against four questions: Was it obtained through a court process (discovery, subpoena, notice to produce, inspection)? Is it legally privileged or a settlement communication? Is it under a suppression or non-publication order? Does it contain other people's personal information? A yes to any of them means it does not go in without advice. Privacy settings, redaction or a public source do not by themselves give permission. In the NSW Supreme Court, SC Gen 23 paragraphs 9A and 9B set conditions on AI use of restricted material; this site's rule, that compelled material never goes into an AI tool, is stricter than that and is our policy. No tool may generate, alter or rephrase affidavit or witness evidence. Where permission is unresolved, do the task by hand.
You want to examine how one letter communicates—not award a trophy for sounding confident.
Visible counts and examples you can question, repeat and compare with the underlying record.
A professional conduct and litigation-risk solicitor.
It measures posture, not legal merit
Who is controlling the evidentiary battlefield?
A strong communicator tends to set the facts, cite the record and state a position. A weaker posture tends to follow the opponent’s structure, rely on inference and ask the opponent to explain or act. The analyser turns those differences into a transparent score with visible counts, examples, ratios and working.
A confident letter can still be legally wrong. A defensive letter can still be legally correct. The score describes the communication posture of one letter; it does not decide the case.
The eight-metric model
Legal hedge density
Substantive uncertainty language per 100 words.
Declarative ratio
Direct statements compared with dependent or conditional sentences.
Paragraph economy
Average paragraph length after non-body material is removed.
Argument concentration
Main propositions per estimated page, with a compound-proposition test.
Assertion to request ratio
How often the writer states rather than asks the opponent to act.
Concession rate
Substantive ground given away or accusations softened.
Reactivity ratio
A five-level classification of who is setting the narrative agenda.
Evidentiary reliance
Specific record-based assertions compared with inference and concern.
The weighted result places the letter in one of five posture bands: Dominant, Strong, Contested, Defensive or Weak. Evidence and narrative posture carry the most weight; style carries the least.
Copy the complete v4.2 analyser
This is the upload-only version. It accepts properly uploaded PDF or DOCX legal correspondence and refuses pasted text or other document types. The exclusions, thresholds, formula and required output structure are part of the invention—copy the complete prompt rather than extracting only the scoring table.
Download the complete prompt as a text file
Use one clean correspondence document
- Copy the complete prompt into a fresh AI chat.
- Wait for the analyser’s exact ready message.
- Upload one PDF or DOCX letter. Do not paste its text.
- Confirm the file is legal correspondence—not an affidavit, pleading, order, submission or evidence bundle.
- Read the metric counts and examples before looking at the composite score.
- Run a second time if needed and compare the working. Variation is information about the model’s reliability.
The unit of analysis is the advocacy body of one letter. Letterhead, signatures, annexures, schedules, metadata and reproduced evidence are excluded. Combining different document types produces a misleading result.
A score is a lens, not a verdict
Scoring the other side's correspondence is a way of reading it. A score must never appear in a letter, affidavit or filed document, and a high score for "authority" says nothing about whether an authority is real or right. Do not upload without-prejudice or Calderbank offers, which are settlement communications; the analyser is for open correspondence. Read a truthful qualification, a proper concession or a legitimate request as what it is, not as weakness; the metric weights are the author's, not a legal standard.
- It does not assess whether the law is correct.
- It does not predict who will win or whether a settlement will occur.
- It does not prove confidence, motive or truth.
- Its thresholds and weights are authored heuristics, not a validated clinical or legal scale.
- Different AI models—or repeated runs—may classify sentences differently.
- A low score may reflect a legitimate need to request information.
- A high score may reflect overconfidence unsupported by the underlying case.
The useful question is not simply “What number did it receive?” It is: Which counted features produced the number, and do those classifications survive human checking?