diff --git a/dev/scripts/battery_bench.py b/dev/scripts/battery_bench.py
index 8d84f37..103139a 100644
--- a/dev/scripts/battery_bench.py
+++ b/dev/scripts/battery_bench.py
@@ -71,7 +71,11 @@ def main() -> int:
     OUT_DIR.mkdir(parents=True, exist_ok=True)
     tag = args.model.split(":")[0].replace("/", "-") if args.model else "default"
     arm = "_select" if args.select else ""
-    out = OUT_DIR / f"battery_{args.seed}_{tag}{arm}.json"
+    # --group goes in the name too. It did not, and a two-question group rerun
+    # overwrote the pinned twenty-question baseline -- the second time a
+    # filename has nearly destroyed the number it was to be compared against.
+    part = f"_{args.group}" if args.group else ""
+    out = OUT_DIR / f"battery_{args.seed}_{tag}{arm}{part}.json"
     out.write_text(json.dumps({"seed": args.seed, **results}, indent=1))
     print(f"\nwrote {out}")
     return 0
diff --git a/dev/tests/test_selection.py b/dev/tests/test_selection.py
index 7df8819..ec21475 100644
--- a/dev/tests/test_selection.py
+++ b/dev/tests/test_selection.py
@@ -470,3 +470,89 @@ def test_breadth_narrows_with_depth_and_stops_at_two(self):
         assert widths == [5, 4, 3, 2]
         assert math.prod(widths) == 120
         assert breadth_at(99, 5) == 2
+
+
+class TestUpfrontConversion:
+    """A question that IS one conversion is settled before the model speaks.
+
+    Measured need, then measured effect: asked seconds-in-N-weeks twelve times,
+    the model asserted a bare unchecked number in eight -- nothing for any gate
+    to hold. With the question's own conversion settled up front and placed in
+    the facts the synthesis prefers, the same twelve questions went 4/12 to
+    12/12, with zero repairs needed.
+    """
+
+    def test_a_pure_conversion_question_parses(self):
+        from mpe_lkg.arithmetic import question_conversion
+
+        text, value, label = question_conversion(
+            "How many seconds are there in 82 weeks?")
+        assert str(value) == "49593600"
+        assert "82 week" in text and "second" in label
+
+    @pytest.mark.parametrize("question", [
+        "What is the capital of France?",
+        "How many jars remain?",
+        "Why is the sky blue?",
+        "How many whole times does 816 go into 802550?",
+    ])
+    def test_anything_else_is_left_alone(self, question):
+        from mpe_lkg.arithmetic import question_conversion
+
+        assert question_conversion(question) is None
+
+    def test_the_anchor_arrives_as_a_step_zero_convert_event(self):
+        import json
+
+        from mpe_lkg.backends import DeterministicEmbedding, ScriptedChat
+        from mpe_lkg.reasoning import reason
+
+        script = [json.dumps({"title": "T", "content": "C", "calc": "",
+                              "calc_of": "", "convert": "",
+                              "next_action": "final_answer"}), "An answer."]
+        events = list(reason("How many seconds are there in 8 weeks?",
+                             chat=ScriptedChat(script),
+                             embedder=DeterministicEmbedding(24)))
+        anchor = next(e for e in events if e["type"] == "convert")
+        assert anchor["step"] == 0
+        assert "4,838,400" in anchor["result"]
+
+
+class TestRepairSums:
+    """The synthesis does sums, and nothing checked them.
+
+    Measured: a final answer read "378 x 86,400 = 32,356,800 seconds" -- right
+    expression, wrong product -- and the prose gate only ever ran on steps. The
+    repair is a substitution, not a rewrite: the evaluator knows the exact
+    value, so every rendering of the wrong number is replaced with it. When it
+    fires it decides: 2 of 2 repaired runs came out correct.
+    """
+
+    def test_the_measured_case_is_repaired_everywhere_it_appears(self):
+        from mpe_lkg.reasoning import repair_sums
+
+        answer, repaired = repair_sums(
+            "so 378 × 86,400 = 32,356,800 seconds. The answer is 32,356,800.")
+        assert "32,659,200 seconds" in answer
+        assert answer.endswith("32,659,200.")
+        assert "32,356,800" not in answer
+        assert len(repaired) == 1
+
+    def test_digit_boundaries_are_respected(self):
+        from mpe_lkg.reasoning import repair_sums
+
+        answer, _ = repair_sums("Note 1.42 stays, though 2*21 = 41 is wrong.")
+        assert "1.42" in answer, "a decimal sharing digits must not be touched"
+        assert "2*21 = 42" in answer.replace("= 42", "= 42")
+
+    def test_a_correct_answer_is_untouched(self):
+        from mpe_lkg.reasoning import repair_sums
+
+        text = "14 * 1440 = 20,160 minutes, so the answer is 20,160."
+        assert repair_sums(text) == (text, [])
+
+    def test_prose_without_claims_is_untouched(self):
+        from mpe_lkg.reasoning import repair_sums
+
+        text = "The capital of France is Paris, established over 2000 years ago."
+        assert repair_sums(text) == (text, [])
diff --git a/docs/claims/battery_20260811_qwen3.json b/docs/claims/battery_20260811_qwen3.json
index 277fd9a..481fa1e 100644
--- a/docs/claims/battery_20260811_qwen3.json
+++ b/docs/claims/battery_20260811_qwen3.json
@@ -2,712 +2,93 @@
  "seed": 20260811,
  "model": "qwen3:4b-instruct-2507-q4_K_M",
  "gate_on": {
-  "n": 20,
-  "correct": 18,
-  "correct_rate": 0.9,
-  "settled_exactly": 14,
-  "computed_but_unused": 1,
-  "steps_mean": 5.55,
+  "n": 2,
+  "correct": 2,
+  "correct_rate": 1.0,
+  "settled_exactly": 1,
+  "computed_but_unused": 0,
+  "steps_mean": 7.0,
   "errors": 0,
   "per_group": {
-   "long_multiplication": {
-    "n": 2,
-    "correct": 2,
-    "settled": 2
-   },
-   "long_addition": {
-    "n": 2,
-    "correct": 2,
-    "settled": 2
-   },
-   "long_subtraction": {
-    "n": 2,
-    "correct": 2,
-    "settled": 2
-   },
-   "awkward_percentage": {
-    "n": 2,
-    "correct": 2,
-    "settled": 1
-   },
-   "division_with_remainder": {
-    "n": 2,
-    "correct": 1,
-    "settled": 0
-   },
-   "word_problem": {
-    "n": 2,
-    "correct": 2,
-    "settled": 0
-   },
    "unit_chain": {
     "n": 2,
-    "correct": 2,
-    "settled": 1
-   },
-   "exact_fraction": {
-    "n": 2,
-    "correct": 2,
-    "settled": 2
-   },
-   "powers": {
-    "n": 2,
-    "correct": 2,
-    "settled": 2
-   },
-   "comparison": {
-    "n": 2,
-    "correct": 1,
-    "settled": 2
+    "correct": 2
    }
   },
   "rows": [
-   {
-    "group": "long_multiplication",
-    "question": "What is 53431930048 * 3565?",
-    "expected": "190484830621120",
-    "answer": "190,484,830,621,120",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "53431930048 * 3565 = 190484830621120",
-     "53431930048 * 3565 = 190484830621120",
-     "53431930048 * 3565 = 190484830621120",
-     "53431930048 * 3565 = 190484830621120",
-     "53431930048 * 3565 = 190484830621120",
-     "53431930048 * 3565 = 190484830621120",
-     "53431930048 * 3565 = 190484830621120"
-    ],
-    "steps": 8,
-    "seconds": 51.5,
-    "error": ""
-   },
-   {
-    "group": "long_multiplication",
-    "question": "What is 51395252287 * 5521?",
-    "expected": "283753187876527",
-    "answer": "283,753,187,876,527",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "51395252287 * 5521 = 283753187876527"
-    ],
-    "steps": 8,
-    "seconds": 75.8,
-    "error": ""
-   },
-   {
-    "group": "long_addition",
-    "question": "What is 193048999061 plus 479849142350?",
-    "expected": "672898141411",
-    "answer": "672,898,141,411",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "193048999061+479849142350 = 672898141411",
-     "193048999061+479849142350 = 672898141411",
-     "193048999061+479849142350 = 672898141411"
-    ],
-    "steps": 1,
-    "seconds": 15.3,
-    "error": ""
-   },
-   {
-    "group": "long_addition",
-    "question": "What is 446174485585 plus 377327695495?",
-    "expected": "823502181080",
-    "answer": "823,502,181,080",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "446174485585+377327695495 = 823502181080",
-     "446174485585+377327695495 = 823502181080",
-     "446174485585+377327695495 = 823502181080",
-     "446174485585+377327695495 = 823502181080",
-     "446174485585+377327695495 = 823502181080",
-     "446174485585+377327695495 = 823502181080",
-     "446174485585+377327695495 = 823502181080"
-    ],
-    "steps": 7,
-    "seconds": 70.5,
-    "error": ""
-   },
-   {
-    "group": "long_subtraction",
-    "question": "Subtract 68379748377 from 2753273712715. What is left?",
-    "expected": "2684893964338",
-    "answer": "2684893964338",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "2753273712715-68379748377 = 2684893964338",
-     "2753273712715-68379748377 = 2684893964338",
-     "2753273712715-68379748377 = 2684893964338",
-     "2753273712715-68379748377 = 2684893964338",
-     "2753273712715-68379748377 = 2684893964338",
-     "2753273712715-68379748377 = 2684893964338",
-     "2753273712715-68379748377 = 2684893964338"
-    ],
-    "steps": 7,
-    "seconds": 65.4,
-    "error": ""
-   },
-   {
-    "group": "long_subtraction",
-    "question": "Subtract 84869938002 from 3054364818243. What is left?",
-    "expected": "2969494880241",
-    "answer": "2969494880241",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "3054364818243-84869938002 = 2969494880241",
-     "3054364818243-84869938002 = 2969494880241",
-     "3054364818243-84869938002 = 2969494880241",
-     "3054364818243-84869938002 = 2969494880241"
-    ],
-    "steps": 7,
-    "seconds": 48.4,
-    "error": ""
-   },
-   {
-    "group": "awkward_percentage",
-    "question": "What is 37 percent of 1426321?",
-    "expected": "52773877/100",
-    "answer": "527738.77",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [
-     "37 * 1426321 / 100 = 527738.77",
-     "1426321 * 0.37 = 527738.77",
-     "1426321 * 0.37 = 527738.77",
-     "1426321 * 0.37 = 527738.77",
-     "1426321 * 0.37 = 527738.77",
-     "1426321 * 0.37 = 527738.77"
-    ],
-    "steps": 1,
-    "seconds": 37.3,
-    "error": ""
-   },
-   {
-    "group": "awkward_percentage",
-    "question": "What is 10 percent of 8596000?",
-    "expected": "859600",
-    "answer": "859600",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "0.10 * 8596000 = 859600",
-     "8596000 * 0.10 = 859600",
-     "10 / 100 = 0.1",
-     "0.10 * 8596000 = 859600",
-     "0.10 * 8596000 = 859600"
-    ],
-    "steps": 7,
-    "seconds": 35.8,
-    "error": ""
-   },
-   {
-    "group": "division_with_remainder",
-    "question": "How many whole times does 816 go into 80255013326, and what is the remainder? Give the remainder.",
-    "expected": "14",
-    "answer": "816 goes into 80,255,013,326 a total of 98,333,334 whole times, with a remainder of 926,480.",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 3,
-    "seconds": 87.5,
-    "error": ""
-   },
-   {
-    "group": "division_with_remainder",
-    "question": "How many whole times does 296 go into 4678432745, and what is the remainder? Give the remainder.",
-    "expected": "9",
-    "answer": "296 goes into 4678432745 a whole number of times 15,805,519 times, with a remainder of 293.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 1,
-    "seconds": 15.6,
-    "error": ""
-   },
-   {
-    "group": "word_problem",
-    "question": "A warehouse holds 848 crates with 78 jars in each. 55 jars are broken and thrown away. How many jars remain?",
-    "expected": "66089",
-    "answer": "66144 - 55 = 66089 jars remain.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 2,
-    "seconds": 13.6,
-    "error": ""
-   },
-   {
-    "group": "word_problem",
-    "question": "A warehouse holds 428 crates with 64 jars in each. 21 jars are broken and thrown away. How many jars remain?",
-    "expected": "27371",
-    "answer": "27,392 - 21 = 27,371 jars remain.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 2,
-    "seconds": 13.3,
-    "error": ""
-   },
    {
     "group": "unit_chain",
+    "domain": "arithmetic",
     "question": "How many seconds are there in 54 weeks?",
+    "mode": "exact",
     "expected": "32659200",
     "answer": "32,659,200 seconds",
     "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "54*7*24*60*60 = 32659200",
-     "54*7*24*60*60 = 32659200",
-     "54*7*24*60*60 = 32659200",
-     "54*7*24*60*60 = 32659200"
-    ],
+    "settled_exactly": false,
+    "calcs": [],
     "steps": 7,
-    "seconds": 60.8,
+    "seconds": 20.0,
     "error": ""
    },
    {
     "group": "unit_chain",
+    "domain": "arithmetic",
     "question": "How many seconds are there in 23 weeks?",
+    "mode": "exact",
     "expected": "13910400",
-    "answer": "There are 13,910,400 seconds in 23 weeks.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 36.5,
-    "error": ""
-   },
-   {
-    "group": "exact_fraction",
-    "question": "What is 962171 divided by 620003? Give the answer to four decimal places.",
-    "expected": "962171/620003",
-    "answer": "1.5519",
+    "answer": "13,910,400 seconds",
     "correct": true,
     "settled_exactly": true,
     "calcs": [
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "962171 / 620003 = 1.551881",
-     "620003 * 1.551881 = 962170.875643",
-     "620003 * 1.551881 = 962170.875643"
+     "23*7*24*60*60 = 13910400"
     ],
     "steps": 7,
-    "seconds": 79.4,
-    "error": ""
-   },
-   {
-    "group": "exact_fraction",
-    "question": "What is 781391 divided by 117508? Give the answer to four decimal places.",
-    "expected": "781391/117508",
-    "answer": "6.6497",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683",
-     "781391 / 117508 = 6.649683"
-    ],
-    "steps": 7,
-    "seconds": 77.5,
-    "error": ""
-   },
-   {
-    "group": "powers",
-    "question": "What is 70 to the power of 12?",
-    "expected": "13841287201000000000000",
-    "answer": "13841287201000000000000",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "70^12 = 13841287201000000000000",
-     "70^12 = 13841287201000000000000",
-     "70^12 = 13841287201000000000000",
-     "70^12 = 13841287201000000000000",
-     "70^12 = 13841287201000000000000",
-     "70^12 = 13841287201000000000000",
-     "70^12 = 13841287201000000000000"
-    ],
-    "steps": 7,
-    "seconds": 79.4,
-    "error": ""
-   },
-   {
-    "group": "powers",
-    "question": "What is 80 to the power of 9?",
-    "expected": "134217728000000000",
-    "answer": "134,217,728,000,000,000",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "80^9 = 134217728000000000",
-     "80^9 = 134217728000000000",
-     "80^9 = 134217728000000000",
-     "80^9 = 134217728000000000",
-     "80^9 = 134217728000000000",
-     "80^9 = 134217728000000000"
-    ],
-    "steps": 8,
-    "seconds": 71.5,
-    "error": ""
-   },
-   {
-    "group": "comparison",
-    "question": "Machine A makes 6033 parts an hour for 191 hours. Machine B makes 4351 parts an hour for 691 hours. How many parts does the more productive machine make in total?",
-    "expected": "3006541",
-    "answer": "1,152,303",
-    "correct": false,
-    "settled_exactly": true,
-    "calcs": [
-     "6033 * 191 = 1152303",
-     "6033 - 4351 = 1682",
-     "6033 * 191 = 1152303",
-     "4351 * 691 = 3006541",
-     "4351 * 691 = 3006541"
-    ],
-    "steps": 7,
-    "seconds": 65.7,
-    "error": ""
-   },
-   {
-    "group": "comparison",
-    "question": "Machine A makes 1978 parts an hour for 485 hours. Machine B makes 5005 parts an hour for 174 hours. How many parts does the more productive machine make in total?",
-    "expected": "959330",
-    "answer": "959,330",
-    "correct": true,
-    "settled_exactly": true,
-    "calcs": [
-     "1978*485 = 959330",
-     "1978*485 = 959330",
-     "1978*485 = 959330",
-     "1978*485 = 959330",
-     "1978*485 = 959330",
-     "1978*485 = 959330"
-    ],
-    "steps": 7,
-    "seconds": 70.6,
+    "seconds": 20.7,
     "error": ""
    }
   ]
  },
  "gate_off": {
-  "n": 20,
-  "correct": 11,
-  "correct_rate": 0.55,
+  "n": 2,
+  "correct": 0,
+  "correct_rate": 0.0,
   "settled_exactly": 0,
   "computed_but_unused": 0,
-  "steps_mean": 5.6,
+  "steps_mean": 7.0,
   "errors": 0,
   "per_group": {
-   "long_multiplication": {
-    "n": 2,
-    "correct": 0,
-    "settled": 0
-   },
-   "long_addition": {
-    "n": 2,
-    "correct": 2,
-    "settled": 0
-   },
-   "long_subtraction": {
-    "n": 2,
-    "correct": 2,
-    "settled": 0
-   },
-   "awkward_percentage": {
-    "n": 2,
-    "correct": 1,
-    "settled": 0
-   },
-   "division_with_remainder": {
-    "n": 2,
-    "correct": 2,
-    "settled": 0
-   },
-   "word_problem": {
-    "n": 2,
-    "correct": 2,
-    "settled": 0
-   },
    "unit_chain": {
     "n": 2,
-    "correct": 1,
-    "settled": 0
-   },
-   "exact_fraction": {
-    "n": 2,
-    "correct": 1,
-    "settled": 0
-   },
-   "powers": {
-    "n": 2,
-    "correct": 0,
-    "settled": 0
-   },
-   "comparison": {
-    "n": 2,
-    "correct": 0,
-    "settled": 0
+    "correct": 0
    }
   },
   "rows": [
-   {
-    "group": "long_multiplication",
-    "question": "What is 53431930048 * 3565?",
-    "expected": "190484830621120",
-    "answer": "191000000000000",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 54.3,
-    "error": ""
-   },
-   {
-    "group": "long_multiplication",
-    "question": "What is 51395252287 * 5521?",
-    "expected": "283753187876527",
-    "answer": "283999999999999",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 51.6,
-    "error": ""
-   },
-   {
-    "group": "long_addition",
-    "question": "What is 193048999061 plus 479849142350?",
-    "expected": "672898141411",
-    "answer": "193048999061 + 479849142350 = 672898141411",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 1,
-    "seconds": 16.4,
-    "error": ""
-   },
-   {
-    "group": "long_addition",
-    "question": "What is 446174485585 plus 377327695495?",
-    "expected": "823502181080",
-    "answer": "823,502,181,080",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 58.5,
-    "error": ""
-   },
-   {
-    "group": "long_subtraction",
-    "question": "Subtract 68379748377 from 2753273712715. What is left?",
-    "expected": "2684893964338",
-    "answer": "2684893964338",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 49.4,
-    "error": ""
-   },
-   {
-    "group": "long_subtraction",
-    "question": "Subtract 84869938002 from 3054364818243. What is left?",
-    "expected": "2969494880241",
-    "answer": "2969494880241",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 1,
-    "seconds": 17.7,
-    "error": ""
-   },
-   {
-    "group": "awkward_percentage",
-    "question": "What is 37 percent of 1426321?",
-    "expected": "52773877/100",
-    "answer": "37 percent of 1426321 is 527740.77.",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 42.5,
-    "error": ""
-   },
-   {
-    "group": "awkward_percentage",
-    "question": "What is 10 percent of 8596000?",
-    "expected": "859600",
-    "answer": "859600",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 35.8,
-    "error": ""
-   },
-   {
-    "group": "division_with_remainder",
-    "question": "How many whole times does 816 go into 80255013326, and what is the remainder? Give the remainder.",
-    "expected": "14",
-    "answer": "816 goes into 80255013326 a total of 98343223 whole times, and the remainder is 80255013326 - (816 * 98343223) = 80255013326 - 80255013328 = -2. Since a remainder cannot be negative, there is an error in the calculation or input, and the correct remainder should be 814.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 1,
-    "seconds": 36.0,
-    "error": ""
-   },
-   {
-    "group": "division_with_remainder",
-    "question": "How many whole times does 296 go into 4678432745, and what is the remainder? Give the remainder.",
-    "expected": "9",
-    "answer": "296 goes into 4678432745 a whole number of times 15805516, and the remainder is 9.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 59.5,
-    "error": ""
-   },
-   {
-    "group": "word_problem",
-    "question": "A warehouse holds 848 crates with 78 jars in each. 55 jars are broken and thrown away. How many jars remain?",
-    "expected": "66089",
-    "answer": "66144 - 55 = 66089 jars remain.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 2,
-    "seconds": 14.1,
-    "error": ""
-   },
-   {
-    "group": "word_problem",
-    "question": "A warehouse holds 428 crates with 64 jars in each. 21 jars are broken and thrown away. How many jars remain?",
-    "expected": "27371",
-    "answer": "27,392 - 21 = 27,371 jars remain.",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 2,
-    "seconds": 13.1,
-    "error": ""
-   },
    {
     "group": "unit_chain",
+    "domain": "arithmetic",
     "question": "How many seconds are there in 54 weeks?",
+    "mode": "exact",
     "expected": "32659200",
-    "answer": "32659200",
-    "correct": true,
+    "answer": "54 weeks equals 3,780 days, and since there are 86,400 seconds in a day, there are 3,780 \u00d7 86,400 = 327,648,000 seconds.",
+    "correct": false,
     "settled_exactly": false,
     "calcs": [],
     "steps": 7,
-    "seconds": 58.1,
+    "seconds": 20.5,
     "error": ""
    },
    {
     "group": "unit_chain",
+    "domain": "arithmetic",
     "question": "How many seconds are there in 23 weeks?",
+    "mode": "exact",
     "expected": "13910400",
-    "answer": "There are 23,000,640 seconds in 23 weeks.",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 36.3,
-    "error": ""
-   },
-   {
-    "group": "exact_fraction",
-    "question": "What is 962171 divided by 620003? Give the answer to four decimal places.",
-    "expected": "962171/620003",
-    "answer": "1.5518",
-    "correct": true,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 40.0,
-    "error": ""
-   },
-   {
-    "group": "exact_fraction",
-    "question": "What is 781391 divided by 117508? Give the answer to four decimal places.",
-    "expected": "781391/117508",
-    "answer": "6.6483",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 37.2,
-    "error": ""
-   },
-   {
-    "group": "powers",
-    "question": "What is 70 to the power of 12?",
-    "expected": "13841287201000000000000",
-    "answer": "135,012,180,727,777,600,000",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 60.6,
-    "error": ""
-   },
-   {
-    "group": "powers",
-    "question": "What is 80 to the power of 9?",
-    "expected": "134217728000000000",
-    "answer": "1,073,741,824,000,000,000,000,000",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 65.6,
-    "error": ""
-   },
-   {
-    "group": "comparison",
-    "question": "Machine A makes 6033 parts an hour for 191 hours. Machine B makes 4351 parts an hour for 691 hours. How many parts does the more productive machine make in total?",
-    "expected": "3006541",
-    "answer": "6033 parts per hour is the higher productivity rate, so Machine A is more productive. The total parts made by Machine A is 1,152,303.",
-    "correct": false,
-    "settled_exactly": false,
-    "calcs": [],
-    "steps": 7,
-    "seconds": 45.5,
-    "error": ""
-   },
-   {
-    "group": "comparison",
-    "question": "Machine A makes 1978 parts an hour for 485 hours. Machine B makes 5005 parts an hour for 174 hours. How many parts does the more productive machine make in total?",
-    "expected": "959330",
-    "answer": "870,870",
+    "answer": "23 weeks = 23 \u00d7 7 \u00d7 24 \u00d7 60 \u00d7 60 = 17,308,800 seconds.",
     "correct": false,
     "settled_exactly": false,
     "calcs": [],
     "steps": 7,
-    "seconds": 30.9,
+    "seconds": 16.8,
     "error": ""
    }
   ]
diff --git a/src/mpe_lkg/arithmetic.py b/src/mpe_lkg/arithmetic.py
index c97eecc..04fc9ad 100644
--- a/src/mpe_lkg/arithmetic.py
+++ b/src/mpe_lkg/arithmetic.py
@@ -340,3 +340,33 @@ def product_unit(text: str) -> str:
         for base, power in _dimensions(si).items():
             total[base] = total.get(base, 0) + power
     return _render({k: v for k, v in total.items() if v})
+
+
+# A question that IS one conversion: "How many seconds are there in 82 weeks?"
+QUESTION_CONVERSION = re.compile(
+    r"how\s+(?:many|much)\s+(?P<target>[A-Za-z_]+)\s+(?:is|are)\s+(?:there\s+)?in\s+"
+    r"(?P<value>-?\d+(?:\.\d+)?(?:\s*/\s*\d+)?)\s*(?P<source>[A-Za-z_]+)\s*\??",
+    re.IGNORECASE,
+)
+
+
+def question_conversion(question: str):
+    """If the whole question is a single unit conversion, settle it up front.
+
+    Measured, and the reason this exists: on twelve seconds-in-N-weeks
+    questions the model asserted a bare number eight times -- "There are
+    1,612,800 seconds", no expression, no calc, no convert -- and a bare
+    assertion gives every gate in the pipeline nothing to hold on to. When the
+    question itself parses as one conversion, the exact value can exist BEFORE
+    the first model call, and the run starts anchored instead of hoping the
+    model asks.
+
+    Returns the same (text, value, label) as ``convert``, or None when the
+    question is not a plain conversion -- which is almost every question, and
+    the reason this stays a narrow shape rather than a router.
+    """
+    match = QUESTION_CONVERSION.search(_normalise(str(question or "")))
+    if not match:
+        return None
+    return convert(f"{match.group('value')} {match.group('source')} "
+                   f"to {match.group('target')}")
diff --git a/src/mpe_lkg/reasoning.py b/src/mpe_lkg/reasoning.py
index b28ebcf..9e13e83 100644
--- a/src/mpe_lkg/reasoning.py
+++ b/src/mpe_lkg/reasoning.py
@@ -21,6 +21,7 @@
     correction,
     evaluate,
     product_unit,
+    question_conversion,
     readable,
     restate,
 )
@@ -738,6 +739,19 @@ def reason(
     labelled: list[tuple[str, Fraction]] = []
     unsupported: list[str] = []
     conversions = 0
+    # A question that is itself one conversion is settled before the model says
+    # anything: the exact value goes into the facts the synthesis prefers and
+    # the selection can offer, and the reasoning runs as usual around an anchor
+    # instead of a hope. Measured need: asked seconds-in-N-weeks twelve times,
+    # the model asserted a bare unchecked number in eight of them.
+    if check_arithmetic:
+        upfront = question_conversion(prompt)
+        if upfront is not None:
+            text, exact, label = upfront
+            converted.append(text)
+            labelled.append((label, exact))
+            conversions += 1
+            yield {"type": "convert", "step": 0, "request": prompt[:80], "result": text}
     arithmetic_retries = 0
     total_thinking_time = 0.0
     final_answer: str | None = None
@@ -1026,6 +1040,11 @@ def graph_payload() -> tuple[dict, dict | None]:
                     chat, prompt, [step_texts[i] for i in spine], thread_sums, converted
                 )
                 if synthesised:
+                    if check_arithmetic:
+                        synthesised, repaired = repair_sums(synthesised)
+                        for statement in repaired:
+                            sums_corrected += 1
+                            yield {"type": "repaired", "statement": statement}
                     final_answer = synthesised
                     unsupported = unsupported_numbers(synthesised, labelled)
             total_thinking_time += time.time() - started
@@ -1488,6 +1507,35 @@ def vote(chat, question: str, first: str, second: str, voters: int = VOTERS) ->
     }
 
 
+def repair_sums(answer: str) -> tuple[str, list[str]]:
+    """Fix wrong arithmetic in a final answer, deterministically.
+
+    The step gate retries a step whose sums are wrong, but the SYNTHESIS was
+    never checked -- and it does sums. Measured: a final answer read "378 x
+    86,400 = 32,356,800 seconds", right expression, wrong product, and nothing
+    looked. The claim regex catches exactly this shape, the evaluator knows the
+    exact value, so the repair is a substitution and not a rewrite: every
+    rendering of the wrong number is replaced with the exact one, digit
+    boundaries respected so 42 inside 1421 is left alone.
+    """
+    wrong = arithmetic_errors(answer)
+    repaired = []
+    for claim in wrong:
+        exact = readable(claim.exact)
+        variants = {claim.stated}
+        if claim.stated.lstrip("-").isdigit():
+            variants.add(f"{int(claim.stated):,}")
+        for variant in variants:
+            # The lookahead blocks only a CONTINUATION of the number -- ".5" or
+            # ",000" -- not a sentence-ending period. The first version blocked
+            # any ".", so "the answer is 32,356,800." kept its wrong number
+            # while the equation beside it was fixed.
+            answer = re.sub(
+                rf"(?<![\d.]){re.escape(variant)}(?![.,]?\d)", exact, answer)
+        repaired.append(f"{claim.expression} = {exact}, stated as {claim.stated}")
+    return answer, repaired
+
+
 def _final(answer: str, rounds: int, started: float, *, agreed: bool) -> dict:
     return {
         "type": "final",