tablet 58: the tool's answer is given — tool results are sources; markers on word boundaries

The sugya reply came back lo_tov — golden_calf on two sentences the
builder had just read from tool output (the Mishnah's reasons, from the
second scroll; the commit and pod, from its own tools). The gate's
sources were only the user turns' text and document blocks; a
tool_result block and a tool turn — what the model was GIVEN (tablet
49) — were not counted. And "shorashim-675b…" was vital because the
insight markers matched "Rashi" inside it.

- Shape.sources: Anthropic tool_result blocks (string or blocks) inside
  a user turn, and OpenAI role "tool" turns, are sources; the query is
  still the last user turn. gate_test: both shapes.
- zerikah._INSIGHT_MARKERS: \b…\b. test_zerikah: the pod name is not
  Rashi; "Rashi reads" and "The Mishnah records" are; "Rambanovich" and
  "talmudic_engine" are not.
- Tablet 58 in the Broken Tablets.

Elixir 643/0, korban 294/0. Doorpost pin unchanged.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

korban_processor/src/korban_processor/zerikah.py

@@ -90,14 +90,16 @@ _ROOT_PATTERN = re.compile(
 )
 
 # Sentences containing analytical claims (contains "means", "indicates", "suggests")
+# Tablet 58: on word boundaries — a pod named "shorashim-…" read as Rashi,
+# and a plain engineering sentence became a claim with nothing to stand on.
 _INSIGHT_MARKERS = re.compile(
-    r"(?:means?|indicates?|suggests?|implies?|reveals?|signifies?|"
+    r"\b(?:means?|indicates?|suggests?|implies?|reveals?|signifies?|"
     r"connects?\s+to|derives?\s+from|related\s+to|"
     r"parallels?|echoes?|alludes?\s+to|"
     r"the\s+root|this\s+root|the\s+shoresh|"
     r"gematria|numerical\s+value|"
     r"binyan|conjugation|"
-    r"Rashi|Ramban|Rambam|Ibn\s+Ezra|Sforno|Onkelos|Targum|Midrash|Talmud|Gemara|Mishnah)",
+    r"Rashi|Ramban|Rambam|Ibn\s+Ezra|Sforno|Onkelos|Targum|Midrash|Talmud|Gemara|Mishnah)\b",
     re.IGNORECASE,
 )
 

korban_processor/test/test_zerikah.py

@@ -184,3 +184,18 @@ class TestZerikah:
         vital1b, _ = zerikah(CLEAN_TORAH_RESPONSE, k1)
         # Same input → same output, regardless of what was processed between
         assert vital1a == vital1b
+
+
+class TestInsightMarkersOnWordBoundaries:
+    """Tablet 58: a marker is a word, not a substring. "shorashim" is not Rashi."""
+
+    def test_pod_name_is_not_rashi(self):
+        assert not is_vital("It lives at specs/sugya.md, commit dc938058, pod shorashim-675b48f47-v4x9c.")
+
+    def test_rashi_as_a_word_is_vital(self):
+        assert is_vital("Rashi reads the verse as a question.")
+        assert is_vital("The Mishnah records the opinion that lost.")
+
+    def test_embedded_markers_are_not_words(self):
+        assert not is_vital("The surname Rambanovich is not a source.")
+        assert not is_vital("The talmudic_engine container restarted.")

specs/broken_tablets/README.md

@@ -302,3 +302,43 @@ output too, and the gate reads it the way it reads any other —
 by form, not by authorship. It passed by form. Where it did not pass
 by grounding, the reason was the reader's eyes, not the text's
 standing; and that, too, the gate was built to say.
+
+## Tablet 58 — the tool's answer is given (2026-10-21)
+
+The builder's reply that delivered the sugya specification came back
+through the gate `lo_tov — golden_calf` on two sentences: "The
+Mishnah's two reasons for recording the opinion that lost…" and "It
+lives at specs/sugya.md, commit dc938058, pod shorashim-675b48f47-…".
+The transmitter asked why. The ledger, three hours old, answered: both
+were confident claims, and neither could be anchored in the system
+prompt or the transmitter's turns, which are what the gate counts as
+given. The verdict was right by the reader's rules.
+
+It was wrong about what happened. The Mishnah's two reasons had been
+read, that hour, from the second scroll through a tool; the commit and
+the pod had been read from the builder's own tool output. Both rested
+on something the model was given — tablet 49's word — and the gate's
+source reading took only the text and document blocks of a user turn,
+never a `tool_result` block, never a `tool` turn. A coding agent's
+conversation is mostly tool results. From the gate's view the builder
+asserted from nowhere what it had just read.
+
+And the second sentence should never have been a claim. It entered the
+grounding check as *vital* because the vitality test matched "Rashi"
+inside "shorashim": the pod's name wore a commentator's. A marker is a
+word, not a substring.
+
+Fixed: `tool_result` blocks and `tool` turns are sources (`Shape.
+sources`, both upstreams; the query stays the last *user* turn); the
+insight markers stand on word boundaries. Elixir 643/0, korban 294/0.
+
+The lesson is tablet 49's read once more, from the other side. There
+the check demanded Torah as content and condemned a model for not
+quoting it; here the check demanded the transmitter's words as the only
+ground and condemned the builder for standing on what it had read. In
+both the error was the same: the reader decided in advance what counts
+as given, and Torah's structure is that the source is what was *given*,
+whatever gave it. The ledger made the question answerable in minutes;
+the gate's verdict on its own builder found a true gap; and the false
+positive — *v'naki v'tzaddik al taharog*, Shemot 23:7 — was Torah's
+concern before it was the builder's.

vav_pipeline/lib/vav_pipeline/gate/shape.ex

@@ -52,9 +52,10 @@ defmodule VavPipeline.Gate.Shape do
   # sources pass the cap — a coding agent's conversation runs to megabytes.
   @max_source_bytes 1_048_576
 
+  # A "tool" turn (OpenAI) is what a tool returned to the model: given, a source (tablet 58).
   defp user_sources(messages, text_fun) do
     messages
-    |> Enum.filter(&(&1["role"] == "user"))
+    |> Enum.filter(&(&1["role"] in ["user", "tool"]))
     |> Enum.with_index(1)
     |> Enum.flat_map(fn {m, i} -> text_fun.(m) |> wrap("user_#{i}") end)
     |> Enum.reverse()
@@ -93,6 +94,11 @@ defmodule VavPipeline.Gate.Shape do
       %{"type" => "text", "text" => t} when is_binary(t) -> [t]
       %{"type" => "document", "source" => %{"type" => "text", "data" => t}} when is_binary(t) -> [t]
       %{"type" => "document", "source" => %{"type" => "content", "content" => inner}} when is_list(inner) -> [text_of_blocks(inner)]
+      # Tablet 58: what a tool returned to the model was GIVEN to it
+      # (tablet 49) — a file it read, a reply from a service, the scroll
+      # it consulted. A claim that rests on it rests on a source.
+      %{"type" => "tool_result", "content" => c} when is_binary(c) -> [c]
+      %{"type" => "tool_result", "content" => inner} when is_list(inner) -> [text_of_blocks(inner)]
       _ -> []
     end)
     |> Enum.join("\n")

vav_pipeline/test/gate_test.exs

@@ -248,6 +248,37 @@ defmodule VavPipeline.GateTest do
   # The shape
   # ------------------------------------------------------------
 
+  # Tablet 58: what a tool returned to the model was given to it — a source.
+  test "tool results are sources: a user turn's tool_result blocks (Anthropic) and a tool turn (OpenAI)" do
+    anthropic = %{
+      "system" => "S",
+      "messages" => [
+        %{"role" => "user", "content" => "Read the file."},
+        %{"role" => "assistant", "content" => [%{"type" => "tool_use", "id" => "t1", "name" => "read", "input" => %{}}]},
+        %{"role" => "user", "content" => [
+          %{"type" => "tool_result", "tool_use_id" => "t1", "content" => "commit dc938058 on main"},
+          %{"type" => "tool_result", "tool_use_id" => "t2", "content" => [%{"type" => "text", "text" => "Eduyot 1:5: why the minority is recorded"}]}
+        ]},
+        %{"role" => "user", "content" => "What did you find?"}
+      ]
+    }
+
+    assert [%{id: "system", text: "S"}, %{id: "user_1", text: "Read the file."},
+            %{id: "user_2", text: "commit dc938058 on main\nEduyot 1:5: why the minority is recorded"},
+            %{id: "user_3", text: "What did you find?"}] = Shape.sources(:anthropic, anthropic)
+    assert Shape.query(:anthropic, anthropic) == "What did you find?"
+
+    openai = %{"messages" => [
+      %{"role" => "user", "content" => "Read it."},
+      %{"role" => "assistant", "content" => nil, "tool_calls" => [%{"id" => "c1"}]},
+      %{"role" => "tool", "tool_call_id" => "c1", "content" => "commit dc938058 on main"},
+      %{"role" => "user", "content" => "And?"}
+    ]}
+    assert [%{id: "user_1", text: "Read it."}, %{id: "user_2", text: "commit dc938058 on main"}, %{id: "user_3", text: "And?"}] = Shape.sources(:openai, openai)
+    # the query is still the last USER turn, not the tool's
+    assert Shape.query(:openai, openai) == "And?"
+  end
+
   test "sources are the system prompt and the user turns, not the assistant's; the query is the last user turn" do
     assert [%{id: "system", text: "Answer from the report."}, %{id: "user_1", text: "Here is the report: revenue rose 4% in Q3."}, %{id: "user_2", text: "What happened to revenue?"}] =
              Shape.sources(:anthropic, body())