Federated Learning (FL) allows multiple data possessors to collaboratively train a shared model without transferring any raw data outside their local administration limits. However, conventional server orchestrated FL is vulnerable to trusted-server reliance, black-box contribution auditability, malicious model updates, and weak accountability. This article proposes a trust-based federated learning framework in which a smart contract enabled by blockchain oversees client registration, model update logging, hash-based integrity verification, trust score calculation, malicious node penalization, aggregation approval, and decentralised audit logging. The paper uses a clearly labelled simulated experimental setup involving 16 clients having a non-IID Fashion-MNIST classification task. Three of these 16 clients are malicious and submit either poisoned updates or updates having hash mismatches. The proposed trust-centric blockchain federated learning configuration is compared with traditional federated averaging based on parameters such as model accuracy, verification rate, malicious-update identification, F1-score, transaction cost, and governance traceability. Based on the simulation results, the new framework enhances the global accuracy from 79.6% to 86.8% under the attack setting. Further hash-verification coverage is complete for submitted updates. Likewise, the simulation rejects the three malicious submissions in the simulated round. Lastly, the provenance records become immutable via incurring additional latency of 121.3 ms per update. The paper contributes a technically defined, IEEE-style framework for federated learning, smart-contract trust scoring, model-integrity verification, and decentralized AI governance. The outcomes are not offered as evidence of real-world deployment but as a reproducible academic design and evaluation framework for future empirical real-world deployment.