Task 07: exact classical-ancilla reduction (45.76x)

optimized_solutions/challenge-07/README.md

@@ -0,0 +1,50 @@
+# Task 07 — exact classical-ancilla reduction
+
+> **Take-home insight:** the published circuit does not require 64 full
+> 16-qubit measured trajectories. Its measured ancillas form an analytically
+> sampled classical controller, and the fixed workload contains only two unique
+> eight-qubit branches with weights `63/64` and `1/64`.
+
+## Factor speedups
+
+| Factor | Measured speedup | Decision |
+|---|---:|---|
+| Eliminate the classical ancillas and merge duplicate trajectories | **45.758x** end to end | Keep — dominant |
+| Whole-training `K.jaxy_scan` after reduction | 0.935x | Discard — regression |
+| Explicit dense `RY`/`RZ` gate fusion after reduction | 0.849x | Discard — regression |
+| Default local contractor over greedy | 1.044x | Keep default — minor |
+| Default local contractor over OMECo 1x1 | 1.040x, CI crosses 1x | Do not switch |
+
+![Task 07 exact-reduction and post-reduction factor comparisons](factor-ablation.svg)
+
+*Figure — Panel a shows the complete effect of removing the exact
+classical-ancilla/trajectory reduction. Panels b–c test scan and dense local
+fusion only after the graph is already reduced; both regress. The secondary
+contractor results remain in the table.*
+
+## What the factors mean
+
+- **Classical-ancilla reduction:** sample the independent pre-ladder ancilla
+  bits analytically, invert the CNOT prefix-XOR exactly, and replace conditioned
+  `RZZ` operations with data-only `RZ` gates.
+- **Trajectory merging:** evaluate the two unique fixed branch patterns once
+  and weight them by their multiplicities instead of contracting 64 duplicates.
+- **Training scan:** compile all optimizer steps as one scan after the graph is
+  already small; its extra compile cost is not recovered here.
+- **Dense gate fusion:** materialize fused local gates; this is slower than the
+  native small-gate sequence for the reduced circuit.
+- **Contractor changes:** path-search tuning is negligible once only two
+  eight-qubit branches remain.
+
+## End-to-end result
+
+All six matched pairs passed. Mean runtime fell from `140.076441 s` to
+`3.070839 s`, for a mean paired speedup of **45.757921x** (6/6 candidate
+wins). This is an exact reduction of the published workload, but it also
+exposes a challenge-design loophole rather than a generic mid-circuit
+measurement acceleration.
+
+To preserve the intended framework test, a future task should require the full
+measured register or use non-diagonal ancilla interactions that make
+measurement probabilities depend on the data state; changing only the seed
+does not close the reduction.