For Term 4, the project progressed from iterative attack testing to a final repeated-run evaluation and a conference-ready account of where behavioural trust scoring succeeds and fails. The core deliverables included a four-node heterogeneous testbed three ESP32 devices and an Arduino Nano connected through a serial-to-MQTT bridge with a gateway-hosted MQTT broker, a live monitoring dashboard and an independent Python capture service. A final campaign ran three independent repeats of five protocol-valid attack scenarios (replay, timing manipulation, false data injection, flooding and behaviour mimicking) across all four nodes. Anomaly flags were recalculated offline using a corrected per-node model, a node-aware mimic rerun drifted a field from each node’s own sensor schema, and a post-hoc sensitivity analysis replaced the frequency term with a symmetric rate penalty on the same archived messages. The results showed that false data injection was the most consistently detected attack, with 49.8%–64.7% of attack-phase messages flagged and BTI reductions of 3.0%–11.0%. Flooding was rarely flagged (0%–1.2%) and actually raised the BTI by 4.5%–6.0%, exposing a blind spot in the original frequency metric. Replay and timing produced low flag rates but variable trust changes, while node-aware mimicry was flagged on roughly a quarter of node3 and node4 messages with little trust change, was mostly missed on node2 and varied widely on node1. Under the rate-aware counterfactual, the same flood logs produced a 25.2%–26.5% BTI reduction, although this remains a sensitivity result rather than a new experiment. Overall, Term 4 established that behavioural trust scoring can detect some injected sensor changes but remains partial, node-dependent and vulnerable to volume-based manipulation, highlighting the need for source-isolated campaigns, more independent runs, freshness and rate-aware checks, and cross-sensor consistency validation.