diff --git a/.env.example b/.env.example
index 095b67b..9b5b127 100644
--- a/.env.example
+++ b/.env.example
@@ -1,4 +1,4 @@
 # Copy to .env (never commit .env). Production uses Secret Manager.
 # GROQ_API_KEY=gsk_xxxx
-# GROQ_MODEL=openai/gpt-oss-20b
+# GROQ_MODEL=qwen/qwen3.8-27b
 # GROQ_BASE_URL=https://api.groq.com/openai/v1
diff --git a/README.md b/README.md
index ad06515..55b8e66 100644
--- a/README.md
+++ b/README.md
@@ -11,7 +11,7 @@ Deterministic systems first; LLM only where semantic reasoning is required.
    provider-neutral ModelClient). Python is spec artifacts + bounded future
    service per ADR-001.
  - Agent count: 1 Investigation Agent (bounded, provider-neutral). No APPROVE/DENY/PAY by AI, ever.
- - LLM provider: Groq (`llama-3.1-8b-instant`, code default) contracted per ADR-002, wiring
+ - LLM provider: Groq (`qwen/qwen3.8-27b`, code default) contracted per ADR-002, wiring
    deferred behind ModelClient (stub/FakeModelClient until key); no Ollama.
 
 ## Prerequisites
diff --git a/apps/api/cmd/agent/groq_provider_selection_test.go b/apps/api/cmd/agent/groq_provider_selection_test.go
index 05f0a51..43b8c26 100644
--- a/apps/api/cmd/agent/groq_provider_selection_test.go
+++ b/apps/api/cmd/agent/groq_provider_selection_test.go
@@ -51,10 +51,11 @@ import (
 
 const (
 	// canonicalGroqModel is the ADR-002 adjudicated canonical model string
-	// (Option A, 2026-09-26). It must equal the code default in
+	// (Option A, 2026-09-26; re-adjudicated to a servable production model
+	// in APA-47, 2026-09-28). It must equal the code default in
 	// groq_model.go and the README line; groq_env_contract_test.go pins
 	// that three-way agreement so the three cannot drift apart silently.
-	canonicalGroqModel = "llama-3.1-8b-instant"
+	canonicalGroqModel = "qwen/qwen3.8-27b"
 
 	// sentinelGroqKey is a fake key material. It exists so the no-leak
 	// assertions have something specific to hunt for. It is not a real
diff --git a/apps/api/internal/investigate/orchestrate/groq_env_contract_test.go b/apps/api/internal/investigate/orchestrate/groq_env_contract_test.go
index 0782c49..4636035 100644
--- a/apps/api/internal/investigate/orchestrate/groq_env_contract_test.go
+++ b/apps/api/internal/investigate/orchestrate/groq_env_contract_test.go
@@ -35,10 +35,11 @@ import (
 
 const (
 	// canonicalModelString is the ADR-002 adjudicated canonical model
-	// string (Option A, 2026-09-26). It is duplicated here as a literal on
+	// string (Option A, 2026-09-26; re-adjudicated to a servable production
+	// model in APA-47, 2026-09-28). It is duplicated here as a literal on
 	// purpose: if the code default ever changes, this test fails and forces
 	// an explicit decision rather than letting code and docs drift.
-	canonicalModelString = "llama-3.1-8b-instant"
+	canonicalModelString = "qwen/qwen3.8-27b"
 
 	// sentinelKey is fake key material used as a leak canary. It is not a
 	// real credential and must never be committed as one.
diff --git a/apps/api/internal/investigate/orchestrate/groq_model.go b/apps/api/internal/investigate/orchestrate/groq_model.go
index f6d1502..9249700 100644
--- a/apps/api/internal/investigate/orchestrate/groq_model.go
+++ b/apps/api/internal/investigate/orchestrate/groq_model.go
@@ -15,8 +15,16 @@ import (
 )
 
 // Groq defaults.
+//
+// defaultGroqModel is the ADR-002 adjudicated canonical model, re-adjudicated
+// to a currently servable production model in APA-47 (2026-09-28). The prior
+// value, llama-3.1-8b-instant, was retired from Groq and answers HTTP 404
+// model_not_found, so it cannot serve production traffic. Re-verified live the
+// same day: HTTP 200 with non-empty choices[0].message.content under this
+// client's existing wire format. Provider-neutral seam and client contract are
+// unchanged; only the model identity moved.
 const (
-	defaultGroqModel   = "llama-3.1-8b-instant"
+	defaultGroqModel   = "qwen/qwen3.8-27b"
 	defaultGroqBaseURL = "https://api.groq.com/openai/v1"
 	groqTimeout        = 30 * time.Second
 )
diff --git a/docs/adr/002-groq-llm-provider.md b/docs/adr/002-groq-llm-provider.md
index 6f7addd..5c86a60 100644
--- a/docs/adr/002-groq-llm-provider.md
+++ b/docs/adr/002-groq-llm-provider.md
@@ -19,12 +19,23 @@ outcome is recorded here.
 
 ## Decision
 
-- Provider: Groq. Model: `llama-3.1-8b-instant` (code default in
+- Provider: Groq. Model: `qwen/qwen3.8-27b` (code default in
   `groq_model.go`; canonical string for the Groq qualification run).
 - Adjudication 2026-09-26 (Option A): the ADR originally contracted
-  `openai/gpt-oss-20b`, but the shipped code default is
+  `openai/gpt-oss-20b`, but the shipped code default was
   `llama-3.1-8b-instant` and working code is not changed to satisfy a
   stale ADR. ADR and README amended to the code default instead.
+- Re-adjudication 2026-09-28 (APA-47): `llama-3.1-8b-instant` has since
+  been retired from Groq and answers HTTP 404 `model_not_found`, so the
+  config contracted above would have failed every inference call. Model
+  re-adjudicated to `qwen/qwen3.8-27b`, confirmed servable live against
+  the configured provider (HTTP 200, non-empty
+  `choices[0].message.content`) under the existing client wire format
+  (`max_tokens`, `temperature: 0`, `stream: false`). This is a config and
+  documentation change only: no `response_format`/JSON mode, no reasoning
+  parameter, no prompt change, and no model-specific response parsing were
+  added, and the model remains untrusted input behind the unchanged
+  deterministic validation/grounding boundary.
 - Client pattern: OpenAI-compatible (`base_url` + env key). Works with
   `openai.OpenAI`, LangChain `ChatOpenAI`, LangGraph nodes unchanged.
 - Ollama is out of scope. localgcp Vertex proxy is reserved for