Research · SEALED 8-CASE PACK · Diagram A · Pack baseline

Arm A: fill a checklist, then hold a scoring contest

Every arm answers the same three questions. What do I ask next? When do I stop? What may I say? Arm A fills every blank. Then it runs fourteen paid scores. It always names ranks 1 and 2.

Arm A: fill a checklist, then hold a scoring contest Arm A fills a fixed checklist, scores fourteen catalogue diseases plus None in one batch, and always names ranks one and two; refusal is one option on a scoring podium, not a door. while checklist items are still blank. 1 HEAR Thai arrives. Path picker: 4 paths. Only diagnosis is drawn here. 2 BUILD THE LIST Same 15 always. 14 + the word None. 3 CHOOSE WHAT TO ASK Fixed checklist. Blanks get asked whether or not they would change anything. 4 ASK All blanks are sent together, in one message. 5 DECIDE WHETHER TO STOP Stop when every item has an answer. Then read back and tidy the note. 6 DECIDE WHAT MAY BE SAID 14 paid calls, one table. always take ranks 1 and 2 300 VH · 250 H · 150 M no source in the files 380 rule when None is 1st 7 SAY Thai from those two ranks. None adds a not-enough information line. 8 RECORD Record what each model call cost. No reason is stored for a question that was not asked. 9 REFUSAL / VET DOOR BY SCORE Observed: one Thai word → 7 questions, 2,109 characters; the next turn repeated the same seven with no change. METAPHOR CHECKLIST · EVERY BLANK TICKED None whether or not the answer could change anything. None · warning outline SEALED 8-CASE PACK · A pack baseline acc 4/8 (0.50) beside C 8/8 (1.00) · SEALED 8-CASE PACK · A McNemar p 0.0625 beside C 8/8 · does not pass 0.05 LEGEND Stage Alarm / no source given Refusal door Record Main flow Loop / write 14 diseases 'None' candidate

Full sentences. Each note matches a numbered chip.

  1. 1

    HEAR

    The Farm Operator's Thai arrives. A path picker picks one of four paths. The four paths are general chat, diagnosis, search, or not-clear-yet. Only the diagnosis path is drawn below.

  2. 2

    BUILD THE LIST

    The code uses the same fifteen candidates every time. They are the 14 catalogue diseases plus one candidate literally called 'None'.

  3. 3

    CHOOSE WHAT TO ASK

    Go through a fixed checklist of facts the code wants. Whatever is still blank gets asked, whether or not the answer would change anything.

  4. 4

    ASK

    The code sends all the blank checklist items together, in one message.

  5. 5

    DECIDE WHETHER TO STOP

    Stop asking when every checklist item has an answer. Then read the facts back once for confirmation. Then tidy the note.

  6. 6

    DECIDE WHAT MAY BE SAID

    Run the 14 disease prompts together as one batch. That is 14 separate paid model calls. Each call returns a score. Those scores build one table covering 14 diseases plus 'None'. Sort it and always take ranks 1 and 2. Turn each score into a word. 300 or more is 'Very High'. 250–299 is 'High'. 150–249 is 'Medium'. Otherwise the word is 'Low'.

  7. 7

    SAY

    Write the Thai reply from those two ranks. If 'None' is first or second, add a line saying there is not enough information. When 'None' is first, the second-ranked disease is still named — but only when that disease's score is 380 or less.

  8. 8

    RECORD

    Record what each model call cost. The code stores no reason for a question that was not asked.

  9. 9

    REFUSAL / VET DOOR

    REFUSAL BY SCORE. Arm A can refuse, but refusal stands inside the contest. 'None' can only be said when it out-scores 14 diseases. It is one option on the list, not a door.

Caption

A reading of the 14-prompt code

The Arm A measured in the SEALED 8-CASE PACK (n=8 cases; pack baseline: this arm) was a simpler test version. That test used one live model call, no catalogue check, and no stop rule. It was not this 14-prompt process. On this pack: C as one group 8/8 (1.00) beside A 4/8 (0.50). Exact one-sided McNemar p (paired yes/no test) = 0.0625. It does not pass 0.05. Nothing is deployed.

Changed from

None — this is the baseline

Changed from the previous arm: none. This arm takes no working parts from an earlier arm. This is the baseline. It adds the fixed checklist and the scoring contest. Later arms reuse only this shared nine-stage drawing frame.

How to read

Marks on this drawing

  • Red: always take ranks 1 and 2, score limits with no source given, 380 rule
  • Cyan: the refusal door is real. Do not read Arm A as unable to refuse.
  • Loop: keep asking while checklist items are still blank

Research only - nothing is deployed - no percentage ever reaches the Farm Operator - the 14 Disease Profiles are vet-reviewed - reviewed, not signed.