diff --git a/.claude/skills/ci-gates/SKILL.md b/.claude/skills/ci-gates/SKILL.md
index e62382b5eb..1657aea3ab 100644
--- a/.claude/skills/ci-gates/SKILL.md
+++ b/.claude/skills/ci-gates/SKILL.md
@@ -7837,3 +7837,54 @@ regression, and expect the regression to be a defect you did not know
 about rather than a mistake in the change.
 
 Related: §241, a guard whose precondition had stopped holding.
+
+## 310. The repository already wrote the correct form, by hand, next door
+
+`gen-verilog` mapped `>>` to Verilog's `>>` unconditionally. Verilog's
+`>>` fills with zeros however the operand is declared; the arithmetic
+shift is `>>>`. On the CORDIC kernels this is a wrong gate, not a wrong
+number: simulated on the actual generated module,
+`cordic_x_next(100, -64, shift=2)` returned **-16268** where the spec, C
+and Zig all say **116**.
+
+The decisive evidence was not in the compiler. It was in the corpus:
+
+    generated Verilog, all 559 modules:   2 occurrences of `>>>`
+    ...and both were inside string literals
+    those string literals:  this project's own hand-written golden
+                            CORDIC RTL, which writes `y0 >>> 1`
+                            for the identical rotation
+
+**The project knew the right operator and the backend emitted it zero
+times.** A hand-written reference sitting in the corpus is an oracle the
+generator can be measured against, and it costs one grep.
+
+When a backend and a hand-written artefact in the same repository
+disagree about the same construct, that is not a matter of taste. Look
+for the artefact first: `grep` for the construct in the specs and in
+`docs/`, and see whether a human ever wrote it out.
+
+## 311. Acceptance could not have caught any of it
+
+Four defects were fixed in one pass. Not one moved an acceptance number:
+
+| defect | what it did | `cc` / `zig` / `iverilog` |
+|---|---|---|
+| `while (c) : (step)` | the step became the whole body | unchanged |
+| suffix dropped | `1u64 << n` shifted at u32 and panicked | unchanged |
+| `>>` on signed | filled with zeros, CORDIC did not converge | unchanged |
+| truncated test | reported OK in the backend that runs tests | unchanged |
+
+Every one produced output the target compiler was happy with. **A gate
+that asks "does it compile" cannot see any defect whose whole nature is
+that it compiles**, and the four above are the entire interesting class:
+the compiler agreed, and the program was wrong.
+
+What did see them: a second implementation to disagree with (the C
+backend against the Verilog one), a hand-written artefact to compare to
+(§310), and simulating the emitted module instead of reading it. Two of
+the four were found by fan-out audits told to prefer findings the target
+compiler ACCEPTS -- the instruction that made them look in the right
+place.
+
+Related: §262, what the comment was hiding.
diff --git a/docs/now/2026-08-30-the-oracle-was-in-the-corpus.md b/docs/now/2026-08-30-the-oracle-was-in-the-corpus.md
new file mode 100644
index 0000000000..aadf9b2596
--- /dev/null
+++ b/docs/now/2026-08-30-the-oracle-was-in-the-corpus.md
@@ -0,0 +1,11 @@
+# NOW -- The oracle was in the corpus (2026-08-30)
+
+## Two lessons from the signed-shift fix (Refs #2868)
+
+- `gen-verilog` mapped `>>` to Verilog's `>>`, which fills with zeros however the operand is declared; the arithmetic shift is `>>>`
+- simulated on the actual generated module: `cordic_x_next(100, -64, shift=2)` returned -16268 where the spec, C and Zig all say 116
+- the decisive evidence was in the CORPUS, not the compiler: all 559 generated modules held 2 occurrences of `>>>`, and both were inside string literals holding this project's own hand-written golden CORDIC RTL, which writes `y0 >>> 1` for the identical rotation
+- the project knew the right operator and the backend emitted it zero times; a hand-written reference in the repository is an oracle, and it costs one grep
+- four defects fixed this pass and NOT ONE moved an acceptance number: the step-as-body, the dropped width suffix, the logical shift, the truncated test
+- every one produced output the target compiler was happy with, so a gate asking "does it compile" could not see any of them
+- what did see them: a second backend to disagree with, a hand-written artefact to compare against, and simulating the module instead of reading it