feat: starter questions are things to learn, not scenarios, and DeepSeek writes them without thinking

The pool was cases — "A 6-week-old with 3 days of projectile vomiting… which
fluid do you start?" — and the user, reading them on the assistant's front
page: "Vague nonsense clinical scenarios. I want concrete things people learn
from and pathophysiology if possible, all related to peds. No more clinical
scenarios. Use deepseek with its reasoning off."

So the brief is inverted. A question asks one definite thing with a definite
answer — a mechanism, a distinction between two look-alikes, the number that
decides, the reason behind a rule of practice — mechanism first, naming its
disease, drug or finding, and staying paediatric. The filter refuses a
vignette opener and a bare definition instead of demanding a number; the
prompt version moves to 3 so every deployment rebuilds the pool once; the
model defaults to DeepSeek flash with thinking disabled, which the assistant
already knows how to ask for.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016fZGJNyDvERbMgS2Uc2msP

src/utils/clinicalPromptPool.js

@@ -251,7 +251,7 @@ function parseJsonObject(text) {
 // Bumped whenever the generation prompt or the filter changes. A pool made by
 // an earlier version is served while a new one is built, so nobody sees an
 // empty screen and nobody has to remember to press Regenerate.
-var PROMPT_VERSION = 2;
+var PROMPT_VERSION = 3;
 
 function createClinicalPromptPool(opts) {
   opts = opts || {};
@@ -332,9 +332,10 @@ function createClinicalPromptPool(opts) {
 
   async function buildCorpusPromptPool() {
     var taxonomy = taxonomyWithQuotas(target);
-    // A small non-reasoning model writes the pool: a reasoning model spends its
-    // budget thinking about a twenty-line list and returns nothing.
-    var chatModel = await opts.getSetting('clinical_assistant.prompt_model', '') || process.env.CLINICAL_ASSISTANT_PROMPT_MODEL || 'openrouter-gpt-4.1-mini';
+    // DeepSeek with thinking off writes the pool (the user's choice): a list
+    // of questions needs no reasoning budget, and a reasoning model spent its
+    // budget thinking about a twenty-line list and returned nothing.
+    var chatModel = await opts.getSetting('clinical_assistant.prompt_model', '') || process.env.CLINICAL_ASSISTANT_PROMPT_MODEL || 'ds-deepseek-v4.1-flash';
     var generated = [];
     var startedAt = Date.now();
     // A build that produced nothing used to vanish without a word; now it says
@@ -357,11 +358,16 @@ function createClinicalPromptPool(opts) {
         var ai = await opts.callAI([
           {
             role: 'system',
-            content: 'Write starter questions for a pediatric clinical assistant, the kind a clinician actually types at 3 a.m. with a patient in front of them. Use only the provided titles/snippets as inspiration; do not answer the questions and do not mention source names. Every question is a case and a decision: it opens with a one-line vignette — an age with a number and unit (e.g. "a 6-week-old", "a 14-year-old"), the setting, and two or three specific findings of which at least one is a number (a vital sign, a lab value, a dose, a duration, a weight) — and then asks ONE thing the clinician must decide now: the next test, the threshold at which to act, a dose with units, admit or discharge, when to repeat, what changes the plan. Never ask a textbook question: no "What is…", "What are the causes/features/signs of…", "How is X managed?", "Which scores are useful…". 18 to 45 words. Return strict JSON only: {"questions":[{"label":"2-5 word label","prompt":"question ending with ?","category":"' + item.category + '","age_band":"' + item.ageBands.join('|') + '","intent":"diagnosis|management|red_flags|differential|counseling|dosing|admission|discharge|review"}]}.'
+            content: 'Write starter questions for a pediatric clinical assistant. They are what a resident or a paediatrician would ask to learn something concrete about children — not a case, not a scenario, not a vignette. Use only the provided titles/snippets as inspiration; do not answer the questions and do not mention source names. Each question asks one definite thing with a definite answer: a mechanism (why a disease produces a finding, how a drug or a test works, what goes wrong in the physiology), a distinction (what separates two conditions that look alike, and why), a threshold or a number that matters in children (a dose, a cut-off, an age at which something changes), or the reason behind a rule of practice. Prefer pathophysiology wherever the source supports it. Name the disease, drug, organism, test or finding in the question itself, and keep it paediatric: an age group, a childhood disease, or a paediatric dose or value. Never open with a patient ("A 6-week-old with…", "A child presents…"), never invent findings, never ask "What is X?" alone. 10 to 30 words. Return strict JSON only: {"questions":[{"label":"2-5 word label","prompt":"question ending with ?","category":"' + item.category + '","age_band":"' + item.ageBands.join('|') + '","intent":"diagnosis|management|red_flags|differential|counseling|dosing|admission|discharge|review"}]}.'
           },
-          { role: 'user', content: 'Category: ' + item.category + '\nAge bands: ' + item.ageBands.join(', ') + '\nIndexed pediatric source snippets:\n\n' + sourceText + '\n\nCreate 20 starter questions for this category, each a specific case with a specific decision. Vary the setting (emergency department, ward, clinic, neonatal unit, telephone advice), the decision (next test, threshold, dose, fluids, admit/discharge, timing, what if a finding changes), and the age. Example of the depth wanted: "A 6-week-old with 3 days of projectile non-bilious vomiting, weight down 8% from birth and a chloride of 88: which fluid do you start, and what corrects before theatre is safe?"' }
+          { role: 'user', content: 'Category: ' + item.category + '\nAge bands: ' + item.ageBands.join(', ') + '\nIndexed pediatric source snippets:\n\n' + sourceText + '\n\nCreate 20 starter questions for this category, each a concrete thing to learn, mechanism first. Vary what is asked (why a finding happens, how a treatment works, what distinguishes two conditions, which number decides, why a rule exists) and the age group. Examples of the kind wanted: "Why does pyloric stenosis produce a hypochloraemic metabolic alkalosis, and why is the urine paradoxically acidic?", "How does a ductal-dependent lesion stay hidden until the duct closes, and what does prostaglandin E1 do?", "Why are infants under 3 months given broader empiric cover for fever than older children?" No scenarios.' }
         ], {
           model: chatModel || undefined,
+          // The pool is a list, not a reasoning task: DeepSeek writes it
+          // without thinking, which is faster and — measured on the assistant
+          // itself — no less accurate for a grounded task.
+          reasoningEffort: 'none',
+          reasoningFormat: 'hidden',
           temperature: 0.78,
           // A case is three to four times the length of a heading; the old
           // ceiling cut the list off mid-JSON and a whole category was lost.
@@ -567,17 +573,20 @@ function createClinicalPromptPool(opts) {
     return taxonomyItem && taxonomyItem.ageBands && taxonomyItem.ageBands[0] || 'school_age';
   }
 
-  // A starter question is a case and a decision. "What red flags in a child's
-  // headache history warrant investigation?" is a chapter heading; it survives
-  // nothing here. A number is what tells a case from a heading — an age, a
-  // vital sign, a lab, a dose, a duration — so one is required, and the
-  // openers that only ever introduce a heading are refused outright.
-  var TEXTBOOK_OPENER = /^(what (is|are)( the)? (definition|cause|causes|feature|features|sign|signs|symptom|symptoms|treatment|management|differential|complication|complications|indication|indications|role|difference)\b|what (is|are) (a |an |the )?[a-z\- ]{1,40}\?$|how (is|are|should) [a-z\- ]{1,60} (managed|treated|evaluated|diagnosed)\?$|which (clinical )?(scores?|tools?|scales?) (is|are) useful\b)/i;
+  // A starter question is one concrete thing to learn, mechanism first. The
+  // pool used to be cases — "A 6-week-old with 3 days of vomiting…: which
+  // fluid?" — and the user found them "vague nonsense clinical scenarios":
+  // "I want concrete things people learn from and pathophysiology if possible,
+  // all related to peds. No more clinical scenarios." So a vignette opener is
+  // refused, and so is the bare definition ("What is X?") that teaches
+  // nothing; a question must name its subject and ask one definite thing.
+  var VIGNETTE_OPENER = /^(a|an|the|this|my|our|in a|in an|for a|for an|you (are|have|see)|there is)\s+(\d+[\s-]*(day|week|month|year)s?[\s-]*old|\d+\s*(yo|y\/o|mo|wk)\b|(newborn|neonate|infant|baby|toddler|child|boy|girl|teen|adolescent|patient|term|preterm)\b[^?]{0,80}\b(presents?|presenting|brought|arrives?|is seen|comes? in|admitted|on the ward|in (the )?(ed|clinic|emergency|nicu|picu))\b)/i;
+  var BARE_DEFINITION = /^what (is|are) (a |an |the )?[a-z\- ]{1,40}\?$/i;
   function isUsefulQuestion(prompt, item) {
     if (!prompt || prompt.length < 40 || prompt.length > 320) return false;
     if (prompt.indexOf('?') === -1) return false;
-    if (!/\d/.test(prompt)) return false;
-    if (TEXTBOOK_OPENER.test(prompt)) return false;
+    if (VIGNETTE_OPENER.test(prompt)) return false;
+    if (BARE_DEFINITION.test(prompt)) return false;
     if (!hasPediatricSignal(prompt, item)) return false;
     return !/\b(source|snippet|textbook|chapter|document|database)\b/i.test(prompt);
   }

test/clinical-assistant-prompt-pool.test.js

@@ -24,7 +24,7 @@ test('clinical assistant starter prompts are Redis or indexed-source backed', ()
   assert.match(pool, /CLINICAL_ASSISTANT_PROMPT_POOL_TARGET, 1000/);
   // Built once, then only from the admin button: no timer unless the environment asks for one.
   assert.match(pool, /CLINICAL_ASSISTANT_PROMPT_POOL_REFRESH_MS, 0\)/);
-  assert.match(pool, /process\.env\.CLINICAL_ASSISTANT_PROMPT_MODEL \|\| 'openrouter-gpt-4\.1-mini'/);
+  assert.match(pool, /process\.env\.CLINICAL_ASSISTANT_PROMPT_MODEL \|\| 'ds-deepseek-v4\.1-flash'/);
   assert.match(route, /loadStoredPromptPool: loadLatestPromptPoolSnapshot/);
   assert.match(pool, /PEDIATRIC_TAXONOMY/);
   assert.match(pool, /taxonomyWithQuotas\(target\)/);
@@ -82,18 +82,19 @@ function fakePool(overrides) {
   }, overrides || {})) };
 }
 
-test('a starter question is a case and a decision, not a chapter heading', () => {
+test('a starter question is one concrete thing to learn, not a scenario', () => {
   const { pool } = fakePool();
   const ok = q => pool.isUsefulQuestion(q, { category: 'respiratory', intent: 'management', age_band: 'infant' });
-  // What the old pool was full of.
-  assert.equal(ok('What red flags in a child\'s headache history warrant further investigation?'), false, 'no number, no case');
-  assert.equal(ok('Which clinical scores are useful for assessing asthma severity in children?'), false, 'a heading');
-  assert.equal(ok('How is bronchiolitis managed in infants?'), false, 'a heading');
-  assert.equal(ok('What are the causes of neonatal jaundice presenting at 2 days?'), false, 'a heading with a number in it is still a heading');
-  // What it should hold.
-  assert.equal(ok('A 6-week-old with 3 days of projectile non-bilious vomiting, weight down 8% and a chloride of 88: which fluid do you start, and what corrects before theatre is safe?'), true);
-  assert.equal(ok('A 14-year-old on the ward with asthma, SpO2 91% after two salbutamol nebulisers: at what point do you add magnesium, and what dose?'), true);
-  assert.equal(ok('In a neonate with bilious emesis and a scaphoid abdomen at 2 days of life, which imaging comes first and what wait is acceptable?'), true);
+  // What the user refused: "vague nonsense clinical scenarios".
+  assert.equal(ok('A 6-week-old with 3 days of projectile non-bilious vomiting, weight down 8% and a chloride of 88: which fluid do you start, and what corrects before theatre is safe?'), false, 'a vignette');
+  assert.equal(ok('A 14-year-old on the ward with asthma, SpO2 91% after two salbutamol nebulisers: at what point do you add magnesium?'), false, 'a vignette');
+  assert.equal(ok('An infant presents with bilious emesis and a scaphoid abdomen: which imaging comes first?'), false, 'a vignette without an age');
+  assert.equal(ok('What is bronchiolitis?'), false, 'a bare definition');
+  // What it should hold: mechanisms, distinctions, the numbers that matter.
+  assert.equal(ok('Why does pyloric stenosis produce a hypochloraemic metabolic alkalosis in infants, and why is the urine paradoxically acidic?'), true);
+  assert.equal(ok('How does a ductal-dependent heart lesion stay silent in a newborn until the duct closes, and what does prostaglandin E1 do?'), true);
+  assert.equal(ok('Why are infants under 3 months given broader empiric antibiotic cover for fever than older children?'), true);
+  assert.equal(ok('What separates croup from epiglottitis in a child, and why does one respond to dexamethasone?'), true);
 });
 
 test('a pool built by an older prompt is served, then rebuilt in the background', async () => {