Abstract : Future 6G physical layers need to work together to improve peak power, spectrum confinement, reliability, energy, and latency when channels are not stationary and service slices are not the same. Traditional approaches for reducing PAPR and shaping spectra, such as PTS/SLM, μ-law companding, and fixed orthogonal precoding, work to some extent but are not very reliable across different types of waveforms (UFMC, OTFS) and power-domain multiplexing (NOMA). To suggest a single, AI-driven signal-shaping architecture that combines a hybrid orthogonal precoding bank (DFT/DST/SRC/ZC) with a probabilistic deep compander (PDC) and selective/clustered companding (SCC). A learning controller, such as Deep Reinforcement Learning (PPO) or a Cultural-History Optimization Algorithm (CHOA), changes the precoding choice, companding profile, cluster aggressiveness, NOMA split, besides delay–Doppler grid (for OTFS) to minimize a multi-objective loss over PAPR, out-of-band (OOB) emissions, EVM/BER, energy per bit, and latency. During exploration, safety shields make sure that power-amplifier headroom, spectrum masks, and latency budgets are followed. Synthetic but reasonable results show that improvements are consistent across waveforms. When CCDF is 10⁻³, PAPR goes down from 11.8 dB (OFDM) besides 10.6 dB (UFMC) to 6.9 dB (Proposed-UFMC) and 6.5 dB (Proposed-OTFS). The power outside of the band (OOB) goes up to −44.5 dBm. For 16-QAM, BER at 10 dB drops to 9.8×10⁻³ (Proposed-UFMC) and 7.5×10⁻³ (Proposed-OTFS). The framework has good energy-latency trade-offs: 0.62 mJ/bit at 1.3 ms (URLLC slice) and 0.45 mJ/bit at 3.8 ms (Green slice). Ablations show how each module works: taking off AI controller lowers PAPR to 8.3 dB and OOB to −39.1 dBm. These results indicate that suggested stack is a feasible 6G-level solution that integrates waveform diversity adaptive, safety-conscious optimization.