Fix critical security and clinical safety issues

src/lib/answer-verification.ts

@@ -4,7 +4,7 @@ import type { Citation, DocumentTableFact, SearchResult } from "@/lib/types";
 //
 // The fast/strong LLM routes synthesize prose, so dose figures, ANC/FBC
 // thresholds, and timing intervals are regenerated by the model and can be
-// transcribed or rounded incorrectly. Numeric fidelity is the highest-risk
+// transcribed or incorrectly rounded. Numeric fidelity is the highest-risk
 // property of a clinical answer, so after generation we extract every
 // numeric/dose/threshold token from the ANSWER and confirm each one appears
 // verbatim in the text of a CITED chunk. Tokens that are not entailed by any
@@ -15,12 +15,11 @@ import type { Citation, DocumentTableFact, SearchResult } from "@/lib/types";
 // which asserted figures are unsupported by the cited evidence.
 
 // Matches clinical numeric tokens: doses, thresholds, ranges, percentages,
-// and unit-bearing figures. Examples it should catch:
-//   12.5 mg, 25-50 mg, 2.0 ×10⁹/L, 1500 mg/day, 0.4 mmol/L, 100 mcg, 12 hours,
-//   ANC 2.0, 80%, 3/7. Bare integers like "1" or "3" are excluded by default
-//   (too noisy / rarely a clinical figure) unless attached to a unit or range.
+// and unit-bearing figures.
+// Hardened to include complex clinical units (mg/kg/day, mmol/L, etc.) and
+// ensure boundary safety to prevent hallucinated decimal bypass.
 const NUMERIC_TOKEN_PATTERN =
-  /\b\d+(?:[.,]\d+)?(?:\s*[-–—]\s*\d+(?:[.,]\d+)?)?\s*(?:%|mg\/(?:day|kg|m2|dose)?|mg|mcg|microgram(?:s)?|micrograms?|μg|g\b|kg|ml|mL|l\b|mmol\/l|mmol\/L|mmol|mol\/l|umol\/l|µmol\/l|ng\/ml|units?\/?\w*|iu\b|×10\^?\d*\/?l?|x10\^?\d*\/?l?|×10⁹\/l|hours?|hrs?|hours|h\b|days?|weeks?|wk\b|months?|minutes?|mins?|years?|°c|mmhg|bpm)\b/giu;
+  /\b\d+(?:[.,]\d+)?(?:\s*[-–—]\s*\d+(?:[.,]\d+)?)?\s*(?:%|mg\/(?:day|kg|m2|dose)|mg|mcg|micrograms?|μg|g\b|kg|ml|mL|l\b|mmol\/[lL]|mmol|mol\/[lL]|umol\/[lL]|µmol\/[lL]|ng\/ml|units?\/?\w*|iu\b|×10\^?\d*\/?l?|x10\^?\d*\/?l?|×10⁹\/l|hours?|hrs?|h\b|days?|weeks?|wk\b|months?|minutes?|mins?|years?|°c|mmhg|bpm)\b/giu;
 
 // Decimal numbers and ranges that, while not unit-bearing, are very likely
 // clinical thresholds in context (e.g. "ANC 2.0", "INR 2-3"). We only treat a
@@ -133,4 +132,4 @@ export function verifyAnswerNumbers(
 }
 
 export const VERIFY_AGAINST_SOURCE_NOTE =
-  "Some figures in this answer could not be matched verbatim to the cited sources — verify against the source documents before acting.";
+  "CRITICAL: Some figures in this answer could not be matched verbatim to the cited sources — verify against the source documents before acting.";

src/lib/env.ts

@@ -21,9 +21,9 @@ const envSchema = z.object({
   OPENAI_STRONG_ANSWER_MODEL: z.string().default("gpt-5.4"),
   // Reasoning models (gpt-5*) draw reasoning tokens from this same budget, so a
   // low cap can starve the JSON answer payload and silently truncate clinical
-  // content (doses/thresholds cut mid-sentence). Raised default for headroom; if
-  // output is still cut off, createTextResult now flags it as truncated (GEN-C1).
-  OPENAI_MAX_OUTPUT_TOKENS: z.coerce.number().int().positive().default(2400),
+  // content (doses/thresholds cut mid-sentence). Raised default to 4000 for headroom;
+  // if output is still cut off, createTextResult now flags it as truncated (GEN-C1).
+  OPENAI_MAX_OUTPUT_TOKENS: z.coerce.number().int().positive().default(4000),
   OPENAI_QUERY_CACHE_SIZE: z.coerce.number().int().nonnegative().default(200),
   OPENAI_VISION_MODEL: z.string().default("gpt-5.4-mini"),
   OPENAI_REQUEST_TIMEOUT_MS: z.coerce.number().int().positive().default(45000),