What it actually does with a case file

Scanned FIRs, photographed pages, digital PDFs, charge sheets, affidavits, depositions, correspondence, orders and office documents. Nothing needs to be retyped, renamed or reordered before it goes in.

• PDF, DOCX, RTF, TXT, CSV
• JPEG, PNG, WebP, TIFF, BMP — including phone photographs of pages
• Hindi, Punjabi, Urdu and English, transcribed in the original script

A digital PDF with a usable text layer is read locally and never leaves the server. A scanned bundle is split into batches of a few pages, each transcribed independently and several running concurrently.

That is the whole difference between a bundle that reads and one that times out. Sending forty pages in a single request means the model must read all forty before writing a word, and the reply runs out of room somewhere around page ten — leaving a document that looks complete and is two thirds missing.

The screen says page 12 of 48, because a forty-eight page scan is one document and forty-eight units of work. A progress bar that reads “0 of 1 documents” for four minutes looks frozen when the server is working perfectly well.

Close the tab if you like. A background worker continues the reading and the progress is there when you come back.

• Illegible handwriting is marked [illegible], never guessed at — and never a number, because a wrong FIR number is worse than a missing one.
• A failed batch prints [pages 5–8 could not be transcribed] in place, rather than joining page 4 to page 9.
• A transcription cut short by length is reported as cut short, with a button to re-read that document in smaller batches.
• A document where every page failed is marked failed. It is not recorded as read with a column of gap markers as its contents.
• A text layer that decodes into nonsense is discarded rather than stored, and the pages are transcribed instead.

• Strengths and weaknesses — each naming the document it came from and how heavily it weighs, with the way the other side will use it.
• Gaps — documents pleaded but not produced, dates asserted but not evidenced. Stated as gaps, never filled in by inference.
• Provisions — clearly engaged, arguably engaged, or asserted without any factual basis on the record.
• Procedure — limitation, sanction, jurisdiction, custody, investigation defects; each satisfied, defective, or undeterminable on this record.
• Authority — for and against, every item labelled unverified.
• Questions — what the papers cannot answer and you can.

Answer the questions it raises and run it again. Answers count as instructions — more reliable than inference, not themselves evidence on record. A critical question left blank holds confidence down however good the rest of the file is, and the same question is never asked twice.

Every run is kept, so you can see that the estimate moved after new papers came in, and by how much.

An advocate ambushed in court because the tool was being kind is worse off than one told plainly. It is instructed to say the uncomfortable thing.

• The figure — never 0 and never 100.
• A confidence level, worked out in code from how much of the file was readable, whether the documents this kind of matter turns on are present, and how many critical questions remain unanswered.
• The reasoning: why that number and not ten points either side.
• What would move it, and roughly by how much.

When confidence is low the figure is deliberately pulled toward the middle and the page says so in terms — including that it should not be repeated to a client as it stands.

Confidence is computed rather than asked of the model. A model asked to rate its own confidence answers “moderate” almost every time.

It is an estimate of prospects on the papers supplied. It has no knowledge of the bench, the listing, the other side's counsel, or anything not in your file.