Guides / Cookbooks

Cookbooks · Techniques · Intermediate · about 8 min

Self-consistency: three wordings

Ask the same question in three wordings in one request. If the answers agree, the wording is sound; if one disagrees, find out why before you act on it.

1. When to use it

   Before you put a new question into production, and whenever a question's wording is changed. Three wordings in one request cost little more than one (Ask many questions in one request).

2. The request

   {
     "model": "ora-fin1",
     "state": {
       "symbol": "TSLA"
     },
     "questions": {
       "drop_a": {
         "type": "yes_no",
         "instructions": "Will it drop at least 5% at some point in the next month?"
       },
       "drop_b": {
         "type": "yes_no",
         "instructions": "Over the coming month, will it trade at least 5% below the latest close at any point?"
       },
       "drop_c": {
         "type": "yes_no",
         "instructions": "Within roughly 21 trading sessions, will the low reach 5% or more under the current price?"
       },
       "up_a": {
         "type": "yes_no",
         "instructions": "Will the stock be higher in a week?"
       },
       "up_b": {
         "type": "yes_no",
         "instructions": "Five trading days from now, will it close above the latest close?"
       },
       "up_c": {
         "type": "yes_no",
         "instructions": "Will next week end with the price up from where it is now?"
       }
     }
   }

   Open this in the Playground

3. The real response

   {
     "model": "ora-fin1",
     "answers": {
       "drop_a": {
         "type": "yes_no",
         "yes_no": 0.638
       },
       "drop_b": {
         "type": "yes_no",
         "yes_no": 0.593
       },
       "drop_c": {
         "type": "yes_no",
         "yes_no": 0.184
       },
       "up_a": {
         "type": "yes_no",
         "yes_no": 0.487
       },
       "up_b": {
         "type": "yes_no",
         "yes_no": 0.534
       },
       "up_c": {
         "type": "yes_no",
         "yes_no": 0.496
       }
     },
     "usage": {
       "input_tokens": 483,
       "output_tokens": 24
     },
     "context": {
       "source": "server",
       "symbol": "TSLA",
       "as_of": "2026-10-30"
     },
     "version": "ora-fin-0.1.1"
   }

   The week question is consistent: up_a, up_b, up_c span 0.047. The drawdown question is not: drop_a 0.638 and drop_b 0.593 agree, but drop_c is 0.184, a spread of 0.454. The third wording ("the low reach 5% or more under the current price") is the odd one out, so it is the one to drop, and the plain wording stays.

4. The code

   cURLPythonJavaScriptPHPGoJavaC#RubyCopy

   Requires curl 7.76+. Save as run.sh, then: bash run.sh

   curl -sS https://oramodel.com/v1/systemone \
     -H "Authorization: Bearer $JEV_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
     "model": "ora-fin1",
     "state": {
       "symbol": "TSLA"
     },
     "questions": {
       "drop_a": {
         "type": "yes_no",
         "instructions": "Will it drop at least 5% at some point in the next month?"
       },
       "drop_b": {
         "type": "yes_no",
         "instructions": "Over the coming month, will it trade at least 5% below the latest close at any point?"
       },
       "drop_c": {
         "type": "yes_no",
         "instructions": "Within roughly 21 trading sessions, will the low reach 5% or more under the current price?"
       },
       "up_a": {
         "type": "yes_no",
         "instructions": "Will the stock be higher in a week?"
       },
       "up_b": {
         "type": "yes_no",
         "instructions": "Five trading days from now, will it close above the latest close?"
       },
       "up_c": {
         "type": "yes_no",
         "instructions": "Will next week end with the price up from where it is now?"
       }
     }
   }'

   from statistics import median
   
   def consistent(answers, names, max_spread=0.10):
       v = [answers[n]["yes_no"] for n in names]
       return max(v) - min(v) <= max_spread, median(v)
   
   ok, p = consistent(data["answers"], ["drop_a", "drop_b", "drop_c"])
   if not ok:
       flag_for_review("drop wordings disagree")   # don't act on p

5. Act on it

   A wide spread is a problem with the question, not a signal about the stock. Rewrite or remove the outlier, re-run, and keep the clearest wording as the one you use from then on. A narrow spread lets you use the median and move on.

6. Variations

   • Use it once when designing questions, not on every call.
   • The Paraphrase check examples pair wordings with the real outcome.
   • Text questions: compare the winning choice across wordings, and treat disagreement as low confidence.

Information, not investment advice. Every number on this page is a real response for the session shown; answers change with each session.

What's next

• Confidence-gated actions