Research · Four-arm workflow · Diagram C2 · Appendix A5 · RMC GRID 48x3=144

Arm C2: two more ideas in the same box — and it got worse

Same machine as C1. Only CHOOSE WHAT TO ASK changes: C1's rule stays, then two extra ideas stack underneath. On RMC GRID 48x3=144 the score sits below Arm B. Kept as a measured negative result; not put into use.

Arm C2: two more ideas in the same box — and it got worse Arm C2 stacks an expected-split rule and saved first-turn work onto C1's CHOOSE WHAT TO ASK; RMC GRID 48x3=144 correctness is 68/144 beside B 74/144, and 68 beside C1 77 on the same grid. while an answer could still change the advice and the question limit allows another question. 1 HEAR The model only turns Thai into facts. It decides nothing. 2 BUILD THE LIST supporting / contradicting / not yet asked. Likely, Possible, Cannot exclude. No number. 3 CHOOSE WHAT TO ASK Rank by whether the answer would change what the farm does. Pick one. also estimate expected split; also reuse the saved first-turn work. 4 ASK One question per message. Unasked questions get a written reason. 5 DECIDE WHETHER TO STOP Stop when no answer could change advice. Question limit 4 + 1 hard stop 10. 6 DECIDE WHAT MAY BE SAID Name one disease, an honest tie, or send to the veterinarian. Never a percentage. 7 SAY The model writes Thai from approved wording. Second and last time. 8 RECORD Evidence Trace (the full written record): the facts, each disease's tally and band, the question chosen, every question skipped and why, and the reason for stopping. 9 REFUSAL / VET DOOR BY COVERAGE METAPHOR LANES GEAR T1 TWO ADD-ONS JAM THE C1 LANES RMC GRID 48x3=144 · C2 acc 68/144 beside B 74/144 · RMC GRID 48x3=144 · C2 acc 68/144 beside C1 77/144 · answer key follows Arm B LEGEND Stage Alarm / no source given Record Main flow Loop / write

Full sentences · matched to the numbered chips

  1. 1

    HEAR

    The model only turns the Thai into facts. It decides nothing.

  2. 2

    BUILD THE LIST

    For each disease in the approved catalogue, sort its listed signs into supporting / contradicting / not yet asked. Then give it a plain word from a fixed 8-row table: Likely, Possible, or Cannot exclude. Never a number. Order the diseases still on the list. Reportables stay in the protected set and are always evaluated.

  3. 3

    CHOOSE WHAT TO ASK

    Rank every possible question by whether its answer would change what the Farm Operator does. If two diseases lead to the same Farm Operator actions, do not spend a question splitting them. Pick one. Also estimate the expected split from possible answers. Also reuse the saved first-turn work instead of doing the work again.

  4. 4

    ASK

    One question per message. Every question not asked is written down with its reason.

  5. 5

    DECIDE WHETHER TO STOP

    Stop when no remaining answer could change what the code advises the Farm Operator to do. Also stop when the question limit is reached. The limit is 4 answers, plus 1 extra safety question. That is about five questions at most in an ordinary case. There is a hard stop at 10 asking turns.

  6. 6

    DECIDE WHAT MAY BE SAID

    Name one disease, state an honest tie, or send to the veterinarian. Never show a percentage.

  7. 7

    SAY

    The model writes the Thai from approved wording. This is the second and last time.

  8. 8

    RECORD

    Evidence Trace: the facts, each disease's tally and band, the question chosen, every question skipped and why, and the reason for stopping.

  9. 9

    REFUSAL / VET DOOR

    REFUSE BY COVERAGE. Does the approved catalogue cover this Health Event at all? If not → send to the veterinarian. No confidence number is needed.

Caption

A measured negative result

Adding more cleverness to the same box made the machine worse. RMC GRID 48x3=144: C2 acc 68/144 beside B 74/144 — C2 sits below the Arm B baseline. The same grid records C2 68 beside C1 77. C2 is also below C1. The expected-split rule and the saved first-turn work hurt. Keep C1 as the only RMC one-step change worth testing again. Do not put C2 into use. The answer key follows Arm B, so the test tilts toward B-like answers.

Changed from

C1 — one box, and it hurt

Changed from C1: keeps all of C1, including the same starting machine. Adds two ideas in the same CHOOSE WHAT TO ASK box — the expected-split rule and the saved first-turn work. The addition hurt. Kept as a measured negative result, not put into use.

How to read

Marks on this drawing

  • Top strip: C1's rule on what the farm does, unchanged
  • Bottom strip (red): expected-split rule + saved first-turn work
  • Down arrow on the box: measured worse on RMC GRID 48x3=144
  • Same grid: wrong reportable flag 37/144 beside B 46/144; mean questions 3.87 beside B 3.58

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