Recognize wood log species in verified Naldo mining tasks

The initial Mine 1 wood log task rejected proven oak_log mining because wood_log was compared as a literal block ID. Resolve generic wood names to 1.19 log species and give the Creative critic the actual mining rules. Physical block evidence remains required; chat and inventory alone cannot prove mining. Three regression cases and 38 related checks passed.

tests/test_creative_mine_evidence.py

@@ -0,0 +1,101 @@
+"""A proven Creative mine without drops must pass; chat/give fakery must not.
+
+Live case: Naldo mined oak_log (verified block change, inventory oak_log 1, no
+drops in Creative) for "Mine 1 wood log", yet completed stayed 0: the generic
+name matched nothing, so evidence read "Verified 0/1 distinct wood_log" and
+both the model and the deny-only veto failed a real mine.
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+
+import voyager_verification as verification
+from voyager_creative_critic import CREATIVE_EVIDENCE_FACTS, VerifiedCreativeCritic
+
+
+def _live_mine_events(log_name="oak_log"):
+    """Final observation shaped like the live Naldo attempt."""
+    return [("observe", {
+        "inventory": {log_name: 1, "netherite_axe": 1},
+        "status": {
+            "position": {"x": 33.38, "y": 67.0, "z": -7.52},
+            "equipment": ["netherite_helmet", "netherite_chestplate",
+                          "netherite_leggings", "netherite_boots",
+                          "netherite_axe", None],
+            "health": 20.0,
+            "food": 20.0,
+        },
+        "verified_actions": [
+            {"kind": "mine", "name": log_name, "positions": [[33, 66, -7]]}
+        ],
+        "step_start_position": {"x": 26.5, "y": 64.0, "z": -6.5},
+    })]
+
+
+def test_proven_creative_mine_passes_without_drops():
+    # Every 1.19 log species satisfies the generic "wood log" objective.
+    for log_name in ("oak_log", "birch_log", "spruce_log", "jungle_log",
+                     "acacia_log", "dark_oak_log", "mangrove_log"):
+        supported, evidence = verification.verify_task(
+            _live_mine_events(log_name), "Mine 1 wood log")
+        assert supported is True, f"{log_name}: {evidence}"
+        assert verification.mechanical_verdict(
+            _live_mine_events(log_name), "Mine 1 wood log") is True
+    # Stone is not wood: the alias must not become a wildcard.
+    supported, _ = verification.verify_task(
+        _live_mine_events("stone"), "Mine 1 wood log")
+    assert supported is False
+
+
+def test_give_or_chat_alone_never_proves_a_mine():
+    # creativeGiveItem/chat fakery: inventory present, zero verified blocks.
+    events = [("observe", {
+        "inventory": {"oak_log": 1},
+        "status": {"position": {"x": 0.0, "y": 64.0, "z": 0.0}},
+        "verified_actions": [],
+    })]
+    supported, evidence = verification.verify_task(events, "Mine 1 wood log")
+    assert supported is False, evidence
+    assert verification.mechanical_verdict(events, "Mine 1 wood log") is False
+    # Obtaining is a different objective and still reads the inventory.
+    supported, _ = verification.verify_task(events, "Obtain 1 wood log")
+    assert supported is True
+
+
+def test_wrapper_passes_real_mine_and_vetoes_fakery():
+    class Upstream:
+        def __init__(self):
+            self.calls = []
+
+        def check_task_success(self, *, events, task, context, chest_observation, max_retries=5):
+            self.calls.append({"context": context})
+            return True, "model says success"
+
+    upstream = Upstream()
+    critic = VerifiedCreativeCritic(upstream)
+    success, _ = critic.check_task_success(
+        events=_live_mine_events(), task="Mine 1 wood log",
+        context="ctx", chest_observation={}, max_retries=5)
+    assert success is True
+    context = upstream.calls[0]["context"]
+    assert "Independent world observation" in context
+    assert "Verified 1/1" in context
+    assert "no item drops" in context
+    assert CREATIVE_EVIDENCE_FACTS in context
+
+    upstream = Upstream()
+    critic = VerifiedCreativeCritic(upstream)
+    fake = [("observe", {
+        "inventory": {"oak_log": 1},
+        "status": {"position": {"x": 0.0, "y": 64.0, "z": 0.0}},
+        "verified_actions": [],
+    })]
+    success, reason = critic.check_task_success(
+        events=fake, task="Mine 1 wood log",
+        context="ctx", chest_observation={}, max_retries=5)
+    assert success is False, "chat/give fakery must not complete a Mine task"
+    assert "contradicts" in reason

voyager_creative_critic.py

@@ -16,6 +16,19 @@ from __future__ import annotations
 from voyager_verification import mechanical_verdict, verify_task
 
 
+# Facts the upstream critic prompt cannot know about this deployment. Appended
+# to the context (never to critic.txt, never to an official method): Creative
+# mining breaks blocks without drops, so only the verified-blocks line proves
+# a mine; 'wood log' is any 1.19 log; obtaining is not mining.
+CREATIVE_EVIDENCE_FACTS = (
+    "Creative mode facts: mining changes blocks but produces no item drops, so "
+    "a Mine task is proven only by the verified-blocks line above, never by "
+    "inventory deltas, chat messages or /give commands. 'wood log' means any "
+    "1.19 log (oak, spruce, birch, jungle, acacia, dark_oak, mangrove). "
+    "Obtaining an item and mining a block are different objectives."
+)
+
+
 class VerifiedCreativeCritic:
     def __init__(self, upstream_critic):
         self.upstream = upstream_critic
@@ -32,7 +45,7 @@ class VerifiedCreativeCritic:
         model_success, model_critique = self.upstream.check_task_success(
             events=events,
             task=task,
-            context=f"{context}\nIndependent world observation: {evidence}",
+            context=f"{context}\nIndependent world observation: {evidence}\n{CREATIVE_EVIDENCE_FACTS}",
             chest_observation=chest_observation,
             max_retries=max_retries,
         )

voyager_verification.py

@@ -237,6 +237,24 @@ def task_history(tasks, limit=48):
     return ", ".join(entries) or "None"
 
 
+# "Mine 1 wood log" names no species. In 1.19 a log is one of these overworld
+# logs (stems are not logs). A generic name must match any of them, in mined
+# blocks and in inventory alike; otherwise a real oak_log mine reads as
+# "Verified 0/1 distinct wood_log" and both the model and the veto fail it.
+WOOD_LOG_IDS = frozenset({
+    "oak_log", "spruce_log", "birch_log", "jungle_log",
+    "acacia_log", "dark_oak_log", "mangrove_log",
+})
+GENERIC_WOOD_NAMES = frozenset({"wood_log", "log", "logs", "wood", "woods"})
+
+
+def _resolve_names(name):
+    """Expand a generic item name to the concrete 1.19 IDs it can mean."""
+    if name in GENERIC_WOOD_NAMES:
+        return sorted(WOOD_LOG_IDS)
+    return [name]
+
+
 # The objective kinds verify_task() actually has a mechanical postcondition for.
 # Anything else (kill, use, breed, ...) returns None from mechanical_verdict, so
 # the Critic keeps full authority over it.
@@ -289,20 +307,29 @@ def verify_task(events, task):
         return False, "Objective is not a simple quantified verb form; judge from the observation."
     kind, amount, name = spec
     if kind in {"mine", "place"}:
+        # Mining is proven by changed blocks, never by inventory: Creative
+        # produces no drops, and a /give or chat message is not a mine.
+        wanted = _resolve_names(name)
         positions = set()
+        matched = set()
         for evidence in final.get("verified_actions", []):
-            if evidence.get("kind") != kind or evidence.get("name") != name:
+            if evidence.get("kind") != kind or evidence.get("name") not in wanted:
                 continue
+            matched.add(evidence.get("name"))
             for position in evidence.get("positions", []):
                 if isinstance(position, list) and len(position) == 3 and all(
                     isinstance(v, (int, float)) and math.isfinite(v) for v in position
                 ):
                     positions.add(tuple(position))
+        detail = name if len(wanted) == 1 else (
+            f"{name} (as {', '.join(sorted(matched)) if matched else 'no matching log mined'})"
+        )
         return len(positions) >= amount, (
-            f"Verified {len(positions)}/{amount} distinct {name} blocks for {kind}."
+            f"Verified {len(positions)}/{amount} distinct {detail} blocks for {kind}."
         )
     if kind in {"obtain", "craft", "smelt", "cook"}:
-        count = (final.get("inventory") or {}).get(name, 0)
+        inventory = final.get("inventory") or {}
+        count = sum(inventory.get(item, 0) for item in _resolve_names(name))
         return isinstance(count, (int, float)) and count >= amount, (
             f"Live Inventory has {count}/{amount} {name}."
         )