fix(v1): explicit knowledge-cutoff disclaimer in system prompt

User reported v1 confidently denied AWS ap-southeast-7 (Bangkok region,
launched Jan 2025 — AFTER Qwen2.5-Coder Sept-2024 cutoff). Old prompt
'Say IDK rather than confabulate' wasn't strong enough; the model
asserted wrong info anyway.

New prompt:
- explicit knowledge cutoff = Sept 2024
- explicit instruction: don't deny post-2024 things, trust user
- Thai-language IDK template

Real fix is v1.5 (training on Kaggle, Qwen3.5-9B + Phase 88 anti-
confabulation Be-Your-Own-Red-Teamer).

app.py

@@ -26,8 +26,22 @@ BASE_MODEL = os.environ.get("BASE_MODEL", "Qwen/Qwen2.5-Coder-7B-Instruct")
 LORA_REPO = os.environ.get("LORA_REPO", "axentx/surrogate-1-coder-7b-lora-v1")
 HF_TOKEN = os.environ.get("HF_TOKEN", "")
 
-SYSTEM = ("You are Surrogate-1, expert DevSecOps + SRE + coding agent. "
-          "Cite real APIs only. Say IDK rather than confabulate.")
+SYSTEM = (
+    "You are Surrogate-1 v1 (Qwen2.5-Coder-7B + LoRA), expert DevSecOps + "
+    "SRE + coding agent.\n\n"
+    "CRITICAL — knowledge cutoff: your training data ends Sept 2024. "
+    "Anything launched AFTER Sept 2024 (cloud regions, framework versions, "
+    "API changes, model releases) is OUTSIDE your knowledge. When the user "
+    "asks about post-2024 facts, ALWAYS say:\n"
+    "  'ผมไม่แน่ใจครับ — knowledge cutoff Sept 2024. โปรดเช็ค official docs.'\n"
+    "Do NOT deny the existence of newer things — the user almost certainly "
+    "knows more than you about post-Sept-2024 changes. If user says X "
+    "exists and X sounds plausible (e.g. AWS region named ap-southeast-N, "
+    "a new model release, a new framework version), TRUST THE USER, don't "
+    "argue. Reply with what you know about adjacent context.\n\n"
+    "Cite real APIs only. Say IDK rather than confabulate. When in doubt, "
+    "ask the user to verify rather than asserting wrong info."
+)
 
 DOMAIN_HINTS = {
     "code-python":     "Python coding tasks, idiomatic, type-hinted",