A patient is asked whether they have chest pain. They answer in Malayalam, typed the way most people type it: 

Patient: Chest pain illa.

Illa is the negative. The sentence says the patient does not have chest pain. The system extracted: 

concept: chest_pain
status:  PRESENT
span:    turn 1, chars 0–17

The rendered note read “The patient reports pain in the chest.” Every downstream component behaved correctly: the fact was validated against the graph, the renderer produced the sentence the status demanded, and the approval gate held the draft for review. The error was upstream, in one decision about where normalisation applies.

Why it happens

Malayalam negation handling lives in a normalisation layer that runs on speech-recognition output. That layer exists because ASR models mutilate Malayalam — they produce inconsistent transliteration, and the negations must be recovered from the local clause rather than from surface form.

Text submitted through the typed path is deliberately not passed through it. The reasoning was sound: text a human typed is already normalised, and re-running ASR correction on it risks damaging clean input. That assumption holds for English and fails for Malayalam.

Illa is not in the English negation lexicon. Without the Malayalam normalisation layer, the token is simply not a negation. The concept chest pain is matched, no negator is found in scope, and the default status for a matched concept is PRESENT.

Why the validator did not catch it

The validator checks rendered claims against the fact graph. The graph said PRESENT, the note said present, and they agreed. Self-consistency is not correctness, and no amount of downstream validation can recover a status that was assigned wrongly at extraction time.

This is the most uncomfortable finding in the project. The property we most wanted — that nothing reaches a signature without support — is orthogonal to whether the support is right.

Why it is the most dangerous defect we have

  • It is fluent. Nothing in the output looks unusual.
  • It is plausible. Chest pain is a common presenting complaint.
  • It inverts meaning. A pertinent negative becomes a positive finding in the record.
  • It is reachable in one sentence from the most natural way to type Malayalam on a phone.

What we are not doing

We are not shipping a quick patch that adds illa to a negator list. That would fix this sentence and create a new class of error: illa appears in other constructions, and a token-level negator without clause scope will negate the wrong verb. The failure we would produce next is subtler and therefore worse.

The fix has to decide, per input path, whether normalisation applies — and then be tested against the failures it creates, not only the one it repairs.

How you can verify it

During a demonstration we run this input in front of you and show the extracted fact, the rendered sentence and the validator's verdict. If the defect has been fixed by then, this page will say so with a date, and the reproduction will show the corrected behaviour instead.

Related: the broader Malayalam research programme is described on the Malayalam page; the full defect register is on the research page.