diff --git a/docs/MATCHING_COOKBOOK.md b/docs/MATCHING_COOKBOOK.md
index 8d29562..0923abf 100644
--- a/docs/MATCHING_COOKBOOK.md
+++ b/docs/MATCHING_COOKBOOK.md
@@ -2158,3 +2158,67 @@ is a rule, not a trick:
 > **When the frame is a clean multiple of 8 bytes short and everything else is identical, add an
 > unreferenced ARRAY local of that size.** cc1 homes an unreferenced *array* but not an unreferenced
 > *scalar*.
+
+### 132. When the original's SHORT path is the fall-through, INVERT the condition — 2 independent instances (worker B)
+
+Worker B's `0x800A8984` (188 B, a table-slot release helper) was **30 bytes off** on layout alone:
+
+    if (a0 < 8)  s0 = table + a0*24; else s0 = 0;     ->  30 bytes of wrong layout
+    if (a0 >= 8) s0 = 0; else s0 = <table>;           ->  residual drops to 3 bytes
+
+**The original BRANCHES INTO the table path and falls through to the zero path.** Worker B notes this
+is the *same* shape as its `0x800FCA90` (where the original branched into the body). **Two independent
+instances, both fixed the same way:**
+
+> **When the original's short path is the fall-through and the long path is the branch target, invert
+> the condition in the source.**
+
+This is the mirrored-layout family (findings 28/43/100) with a **stated direction** rather than a
+"try inverting it" heuristic. Note it is the *opposite* of the usual instinct, which is to write the
+guard as the early-exit.
+
+**And worker B's second residual was exactly one register** — the loaded pointer in **a1** in the
+original versus **a0** in the candidate. Introducing an explicit `int *p` local to change the pseudo
+order had **no effect**, so it classified after two spellings. That is the allocator class, correctly
+identified and correctly abandoned.
+
+### 133. A FAMILY GIVES YOU THE SHAPE, NOT THE DECLARATIONS (worker A)
+
+Worker A's `0x800506E4` was **one byte wrong**: the sibling's source shape matched, `int i = 0` was
+kept, and cc1 still emitted `sltiu` — but **for this row the signed form emits `slt`.** Declaring
+`unsigned int i` reproduces the original's `sltiu`, and the next row (`0x80050548`) then matched
+**first try with that declaration inherited.**
+
+> **When a family row is one byte off on a comparison opcode, try `unsigned` on the induction
+> variable before anything else.**
+
+**The family transfers the SHAPE and the LEVERS; the DECLARATIONS must still be re-derived.** That is
+finding 55's limit stated precisely: the family gives you a 90% draft and you must still check the
+types.
+
+### 134. A family hit is also a hint about the CALLEE, and the search does not model it (worker A)
+
+Worker A observed that the family relation **crosses the call graph in both directions**:
+
+- `0x8006B2D4` (a family row) **calls** `0x800C022C` — a row worker A had already classified, sharing
+  the sub-object layout `mid+352` / `mid+396`.
+- `0x8003022C` (a family row) **calls** `0x8002FE3C`, which is itself on the unclaimed worklist.
+
+**So a family hit is also a hint about its callees:** when a row's sibling calls a function you have
+already seen, check whether the callee is an unclaimed row with the same object layout. **That is a
+second row from one reading**, and `tools/sf3_family` does not model it — it compares bodies, not
+call edges.
+
+### 135. A predicted family member that does NOT exist is a result (coordinator, negative)
+
+Worker A mapped a **four-member family** completely and predicted the fourth by pattern — bases
+stepping by −0xD8 and −0x300, so `src = 0x80116DD0`, `dst = 0x8012FADC`, with the finaliser argument
+continuing 6, 2, 0, …
+
+**I scanned all 1046 unclaimed worklist rows for `lui`/`addiu` pairs producing either predicted base:
+ZERO rows reference them.** So the family has exactly three members, and the pattern extrapolation
+does not continue.
+
+**That is a useful negative**: it bounds the family, and it means no worker should spend time looking
+for the fourth. **A predicted member that does not exist is worth recording**, because the next
+worker will otherwise re-derive the prediction and repeat the search.