Whole-brain models of Drosophila melanogaster have progressed from connectome-constrained simulation to neuromorphic implementation and closed-loop embodiment. This progress creates an ethical question that is no longer purely hypothetical: could large numbers of such emulations instantiate negatively valenced states? We argue that the answer is presently indeterminate. The target is valence - whether a state is good or bad for the system - rather than consciousness, which remains a further unresolved question. Behaviour cannot decide the issue because the interpretation of pain-like behaviour in insects is itself contested. Connectivity cannot decide it because a connectome omits intrinsic dynamics, neuromodulation and internal state, all of which can reconfigure the functional meaning of the same anatomical pathways. A simulated body also supplies mechanics without necessarily supplying physiological variables whose preservation or violation could ground welfare. Instead of treating valence as a single hidden property, we propose an architectural audit of four component groups: aversive transduction, shared valence and motivational gating, persistence and valence-gated plasticity, and grounding in system-relevant variables. The audit should be paired with one-directional comparison against biological recordings and, where learning is present, a species-appropriate judgement-bias assay. Precaution should be triggered when any of three features appears: a persistent or plastic aversive state, a functional valence signal that globally biases action, or variables that are genuinely at stake for the system. Publicly documented implementations do not yet clearly meet these conditions, but the relevant components are technically accessible. The appropriate policy is therefore neither a declaration of suffering nor an assumption of harmlessness, but auditable uncertainty with explicit scaling safeguards.