Algorithm 1: Adaptive Bridge Controller (ABC)
[1]	Input: Self-denoised tensors ($F_m$), Mid-level semantic tensors ($S_m$)	$m \in \left\{L, C \right\}$
[2]	Output: Semantic guidance for CMD	//Flow from clean sensor to noisier sensor
1:	begin ABC module
2:	for each input do
3:	Common latent space
4:	$H_m=W[F_m\Vert S_m]$	$m \in \left\{L, C \right\}$
5:	Bidirectional Cross-attention	//Parallel operation
6:	$A_{L\to C}=Softmax\left({\displaystyle \frac{(H_L{W}_Q\left)\right(H_C{W}_K{)}^T}{\sqrt{d_k}}}\right)H_C{W}_V$
7:	$A_{C\to L}=Softmax\left({\displaystyle \frac{(H_C{W}_Q\left)\right(H_L{W}_K{)}^T}{\sqrt{d_k}}}\right)H_L{W}_V$
8:	Calculate reliability scores for both modalities
9:	$R_L= \mathit{cos}(H_L, A_{C\to L})$	//Reliability for LiDAR
10:	$R_C=\mathit{cos}(H_C,{ A}_{L\to C})$	//Reliability for Camera
11:	$\Delta R=|R_L-R_C|$	//Reliability difference
12:	$g=\sigma (M(\Delta R\left)\right)$	//gating function
13:	if $g > \tau$ then	//LiDAR is more reliable
14:	guidance mode ← “LiDAR to Camera”	//LiDAR guides Camera
15:	semantic tensor ← $S_L$	//LiDAR semantic tensor as cleane
16:	noisy tensor ← $F_C$	//Camera as noisy modality
17:	else
18:	guidance mode ← “Camera_to_LiDAR”	//If Camera is more reliable
19:	semantic tensor ← $S_C$	//Use Camera semantic tensor
20:	noisy tensor ← $F_L$	//LiDAR is noisy
21:	end if
22:	The guidance (semantic flow decision)
23:	$Guide=\left\{\begin{array}{ll}L\to C,& g>\tau \\ C\to L,& g\le \tau \end{array}\right.$
24:	return guidance
25:	end for
26:	End