diff --git a/autoresearch/prefill/prepare.py b/autoresearch/prefill/prepare.py
index b273d60..90427d4 100644
--- a/autoresearch/prefill/prepare.py
+++ b/autoresearch/prefill/prepare.py
@@ -46,7 +46,7 @@ def evaluate(report: dict, candidate) -> dict:
         ),
         "recursive_protocol": (
             critic.get("critic_protocol")
-            == "recursive_proof_decomposition_v2"
+            == "goal_anchored_recursive_gan_v3"
         ),
         "no_fallback": int(delta.get("fallbacks", 0)) == 0,
         "no_job_failure": int(delta.get("remote_job_failures", 0)) == 0,
diff --git a/autoresearch/prefill/program.md b/autoresearch/prefill/program.md
index 82e10bd..55c2a6f 100644
--- a/autoresearch/prefill/program.md
+++ b/autoresearch/prefill/program.md
@@ -17,7 +17,7 @@ Minimize `metric_cold_critic_prefill_s`. Lower is better.
 - Every compute segment must remain at or below 300 seconds.
 - Critic must receive the complete Generator response.
 - `critic_omitted_tokens` must equal zero.
-- Protocol must be `recursive_proof_decomposition_v2`.
+- Protocol must be `goal_anchored_recursive_gan_v3`.
 - Snapshot mode must remain `final_only`.
 - No fallback, local Primary Prefill, failed remote job, sampling, summary, or
   semantic simplification is allowed.
diff --git a/docs/ops/distributed-prefill-kv-network.md b/docs/ops/distributed-prefill-kv-network.md
index 61b4287..87b499d 100644
--- a/docs/ops/distributed-prefill-kv-network.md
+++ b/docs/ops/distributed-prefill-kv-network.md
@@ -349,13 +349,20 @@ and semantic fallback are forbidden. A global Critic score is valid only when
 `critic_omitted_tokens=0`. Long Prefill operations emit a heartbeat every 30
 seconds; on the 16GB allens worker, full-context Critic Prefill may take 15–25
 minutes.
-The Critic uses `recursive_proof_decomposition_v2`: numeric scores and blanket
+The interactive Critic uses `goal_anchored_recursive_gan_v3`: numeric scores and blanket
 approval are forbidden. It ignores prizes, money, prestige, style, and other
 proof-irrelevant facts. It attacks the central stopping claim, recursively
 decomposes it into proof obligations, records arguments/counterarguments and
 dependencies for every node, and loops until every leaf is either explicitly
 derived or a precisely stated open lemma. It then reports the smallest
 unresolved frontier and next adversarial step.
+The first substantive REPL input becomes an immutable research goal. `/continue`
+feeds the complete previous Generator and Critic outputs into the next
+Generator turn; `/new <goal>` is the only way to switch topics. Inputs beginning
+with `generator>`, `critic>`, `prompt>`, `[metrics]`, `[allens]`, errors, or
+tracebacks are rejected so Terminal output cannot contaminate the research
+objective. Every Critic turn emits an explicit `Goal Alignment` decision and
+restores the anchored proof frontier when drift is detected.
 The worker reserves estimated final-snapshot capacity before model compute,
 prevents adaptive shrink from consuming active reservations, then atomically
 publishes and leases the final snapshot before adding optional intermediate
diff --git a/scripts/agent_gan_inference_demo.py b/scripts/agent_gan_inference_demo.py
index b4f19df..d370223 100644
--- a/scripts/agent_gan_inference_demo.py
+++ b/scripts/agent_gan_inference_demo.py
@@ -40,14 +40,19 @@ def _output_metadata(text: str) -> dict:
     }
 
 
-def build_critic_context(tokenizer, text: str) -> tuple[str, dict]:
+def build_critic_context(
+    tokenizer,
+    text: str,
+    *,
+    protocol: str = "recursive_proof_decomposition_v2",
+) -> tuple[str, dict]:
     full_ids = tokenizer.encode(text, add_special_tokens=False)
     return text, {
         "generator_full_tokens": len(full_ids),
         "critic_context_tokens": len(full_ids),
         "critic_omitted_tokens": 0,
         "review_scope": "full",
-        "critic_protocol": "recursive_proof_decomposition_v2",
+        "critic_protocol": protocol,
     }
 
 
diff --git a/scripts/agent_gan_repl.py b/scripts/agent_gan_repl.py
index cb03f65..82a1ffd 100644
--- a/scripts/agent_gan_repl.py
+++ b/scripts/agent_gan_repl.py
@@ -5,6 +5,7 @@
 import argparse
 import hashlib
 import json
+import re
 import signal
 import threading
 import time
@@ -46,7 +47,37 @@ def _telemetry_request(url: str, **kwargs):
         return None
 
 
-def build_generator_messages(prompt: str) -> list[dict[str, str]]:
+_RUNTIME_ARTIFACT = re.compile(
+    r"^\s*(?:generator>|critic>|prompt>|\[(?:metrics|allens|error|"
+    r"telemetry-warning|protected|supervisor)\]|Traceback\b)",
+    re.IGNORECASE,
+)
+
+
+def is_runtime_artifact_prompt(text: str) -> bool:
+    lines = [line for line in text.splitlines() if line.strip()]
+    return bool(lines) and bool(_RUNTIME_ARTIFACT.match(lines[0]))
+
+
+def build_generator_messages(
+    goal: str,
+    *,
+    steering: str = "",
+    previous_generator: str = "",
+    previous_critic: str = "",
+) -> list[dict[str, str]]:
+    feedback = ""
+    if previous_generator or previous_critic:
+        feedback = (
+            "\n\nComplete previous Generator response:\n"
+            f"{previous_generator}\n\nComplete previous Critic correction:\n"
+            f"{previous_critic}\n\nApply the Critic's Next Adversarial Step "
+            "while remaining anchored to the immutable goal."
+        )
+    steering_text = (
+        f"\n\nCurrent human steering (subordinate to the goal):\n{steering}"
+        if steering else ""
+    )
     return [
         {
             "role": "system",
@@ -59,14 +90,21 @@ def build_generator_messages(prompt: str) -> list[dict[str, str]]:
                 "Distinguish unknown from impossible."
             ),
         },
-        {"role": "user", "content": prompt},
+        {
+            "role": "user",
+            "content": (
+                f"IMMUTABLE RESEARCH GOAL:\n{goal}"
+                f"{feedback}{steering_text}"
+            ),
+        },
     ]
 
 
 def build_critic_messages(
-    prompt: str,
+    goal: str,
     generator_response: str,
     *,
+    steering: str = "",
     stop_reason: str,
     complete: bool,
 ) -> list[dict[str, str]]:
@@ -94,12 +132,17 @@ def build_critic_messages(
                 "Leaf Obligation Ledger; Smallest Unresolved Frontier; Next "
                 "Adversarial Step. Do not sample, summarize, simplify, or use a "
                 "fallback review."
+                " Begin with `Goal Alignment: ALIGNED` or `Goal Alignment: "
+                "DRIFTED`. If drifted, discard the off-topic branch and restore "
+                "the proof-obligation frontier for the immutable goal."
             ),
         },
         {
             "role": "user",
             "content": (
-                f"Task:\n{prompt}\n\nComplete response:\n{generator_response}\n\n"
+                f"IMMUTABLE RESEARCH GOAL:\n{goal}\n\n"
+                f"Current steering:\n{steering or '(none)'}\n\n"
+                f"Complete response:\n{generator_response}\n\n"
                 f"Completion: {stop_reason}; complete={complete}"
             ),
         },
@@ -257,11 +300,16 @@ def get_stats():
         return stats
 
     print(
-        "Kakeya Agent GAN REPL ready. Type a prompt; /quit exits.\n"
+        "Kakeya Agent GAN REPL ready. First prompt sets the immutable goal.\n"
+        "Use /continue to apply Critic feedback, /new <goal> to reset, "
+        "and /quit to exit.\n"
         "Each turn runs allens Prefill → Primary hot Generator → "
         "allens Prefill → Primary hot Critic.",
         flush=True,
     )
+    research_goal = ""
+    previous_generator = ""
+    previous_critic = ""
     with Client(args.address) as client:
         while True:
             try:
@@ -274,6 +322,33 @@ def get_stats():
             if prompt.lower() in {"/quit", "/exit"}:
                 print("[bye]")
                 break
+            if prompt.lower().startswith("/new"):
+                new_goal = prompt[4:].strip()
+                research_goal = new_goal
+                previous_generator = ""
+                previous_critic = ""
+                if not research_goal:
+                    print("[goal] cleared; enter a new research goal", flush=True)
+                    continue
+                prompt = research_goal
+                print(f"[goal] reset: {research_goal}", flush=True)
+            elif prompt.lower() == "/continue":
+                if not research_goal:
+                    print("[goal-error] no active research goal", flush=True)
+                    continue
+                prompt = ""
+            elif is_runtime_artifact_prompt(prompt):
+                print(
+                    "[input-rejected] runtime output cannot become a research "
+                    "prompt; use /continue or /new <goal>",
+                    flush=True,
+                )
+                continue
+            elif not research_goal:
+                research_goal = prompt
+                print(f"[goal] anchored: {research_goal}", flush=True)
+                prompt = ""
+            steering = prompt
             run_nonce = uuid.uuid4().hex
             telemetry_state["degraded"] = False
             run = _telemetry_request(
@@ -288,13 +363,22 @@ def get_stats():
                         "agents": ["generator", "critic"],
                         "rounds": 1,
                         "output_tokens": args.output_tokens,
+                        "goal_anchor": hashlib.sha256(
+                            research_goal.encode(),
+                        ).hexdigest(),
+                        "feedback_applied": bool(previous_critic),
                     },
                 },
             )
             remote_run = run is not None
             run_id = run["id"] if remote_run else f"local_{run_nonce[:16]}"
             try:
-                generator_messages = build_generator_messages(prompt)
+                generator_messages = build_generator_messages(
+                    research_goal,
+                    steering=steering,
+                    previous_generator=previous_generator,
+                    previous_critic=previous_critic,
+                )
                 generator_ids = tokenizer.apply_chat_template(
                     generator_messages,
                     add_generation_prompt=True,
@@ -348,6 +432,7 @@ def get_stats():
                 critic_context, context_metrics = build_critic_context(
                     tokenizer,
                     generator_text,
+                    protocol="goal_anchored_recursive_gan_v3",
                 )
                 if (
                     critic_context != generator_text
@@ -356,8 +441,9 @@ def get_stats():
                 ):
                     raise RuntimeError("Critic full-context invariant violated")
                 critic_messages = build_critic_messages(
-                    prompt,
+                    research_goal,
                     critic_context,
+                    steering=steering,
                     stop_reason=generator_actual["stop_reason"],
                     complete=generator_actual["complete"],
                 )
@@ -400,6 +486,8 @@ def get_stats():
                 )
                 if not critic_stage["ok"] and not telemetry_state["degraded"]:
                     raise _gate_failure("Critic", critic_warm, critic_actual)
+                previous_generator = generator_text
+                previous_critic = critic_text
                 completed = None
                 if remote_run:
                     completed = _telemetry_request(
diff --git a/tests/inference_engine/bench/test_prefill_autoresearch.py b/tests/inference_engine/bench/test_prefill_autoresearch.py
index 3bece3d..4bef89c 100644
--- a/tests/inference_engine/bench/test_prefill_autoresearch.py
+++ b/tests/inference_engine/bench/test_prefill_autoresearch.py
@@ -18,7 +18,7 @@ def _report(**overrides):
         "generator_full_tokens": 900,
         "critic_context_tokens": 900,
         "critic_omitted_tokens": 0,
-        "critic_protocol": "recursive_proof_decomposition_v2",
+        "critic_protocol": "goal_anchored_recursive_gan_v3",
         "delta": {"fallbacks": 0, "remote_job_failures": 0},
     }
     stage.update(overrides)
diff --git a/tests/inference_engine/bridge/test_agent_gan_repl.py b/tests/inference_engine/bridge/test_agent_gan_repl.py
index 896429c..2034eb5 100644
--- a/tests/inference_engine/bridge/test_agent_gan_repl.py
+++ b/tests/inference_engine/bridge/test_agent_gan_repl.py
@@ -11,6 +11,7 @@
     build_critic_messages,
     build_generator_messages,
     install_signal_protection,
+    is_runtime_artifact_prompt,
 )
 
 
@@ -149,13 +150,13 @@ def test_stage_includes_full_context_metrics():
             "critic_context_tokens": 100,
             "critic_omitted_tokens": 0,
             "review_scope": "full",
-            "critic_protocol": "recursive_proof_decomposition_v2",
+            "critic_protocol": "goal_anchored_recursive_gan_v3",
         },
     )
     assert stage["critic_context_tokens"] == 100
     assert stage["critic_omitted_tokens"] == 0
     assert stage["review_scope"] == "full"
-    assert stage["critic_protocol"] == "recursive_proof_decomposition_v2"
+    assert stage["critic_protocol"] == "goal_anchored_recursive_gan_v3"
 
 
 def test_telemetry_timeout_warns_without_stopping_inference(
@@ -185,17 +186,24 @@ def test_gate_failure_exposes_reuse_counters():
 
 
 def test_interactive_prompts_are_deterministic_for_kv_reuse():
-    generator_a = build_generator_messages("prove RH")
-    generator_b = build_generator_messages("prove RH")
+    kwargs = {
+        "steering": "continue the zero-free-region branch",
+        "previous_generator": "previous complete argument",
+        "previous_critic": "previous complete correction",
+    }
+    generator_a = build_generator_messages("prove RH", **kwargs)
+    generator_b = build_generator_messages("prove RH", **kwargs)
     critic_a = build_critic_messages(
         "prove RH",
         "complete generator response",
+        steering=kwargs["steering"],
         stop_reason="eos",
         complete=True,
     )
     critic_b = build_critic_messages(
         "prove RH",
         "complete generator response",
+        steering=kwargs["steering"],
         stop_reason="eos",
         complete=True,
     )
@@ -203,6 +211,10 @@ def test_interactive_prompts_are_deterministic_for_kv_reuse():
     assert critic_a == critic_b
     combined = repr(generator_a + critic_a)
     assert "Internal run" not in combined
+    assert "IMMUTABLE RESEARCH GOAL" in combined
+    assert "previous complete correction" in combined
+    assert "Goal Alignment: ALIGNED" in combined
+    assert "Goal Alignment: DRIFTED" in combined
     assert "open problem" in combined
     assert "recursive adversarial proof analyst" in combined
     assert "Never output a numeric score" in combined
@@ -212,3 +224,15 @@ def test_interactive_prompts_are_deterministic_for_kv_reuse():
     assert "Ignore prizes, money, prestige" in combined
     assert "smallest unresolved frontier" in combined
     assert "sample, summarize, simplify" in combined
+
+
+def test_runtime_output_cannot_replace_research_goal():
+    for text in (
+        "critic> ### Central Claim",
+        "[metrics] KV hit=100%",
+        "[allens] Critic Prefill: 30s",
+        "prompt> ",
+        "Traceback (most recent call last):",
+    ):
+        assert is_runtime_artifact_prompt(text)
+    assert not is_runtime_artifact_prompt("证明黎曼猜想")