LedgerMind introduces a novel framework for multimodal agentic reasoning designed to address the limitations of aggregate accuracy metrics by enforcing strict provenance constraints. Rather than relying on simple final-answer validation, the system treats agent trajectories as provenance-constrained state machines. This is achieved through a Structured Evidence Ledger that normalizes all tool outputs into a formal state, ensuring that subsequent reasoning and decision claims can only cite active ledger entries. The architecture incorporates a Three-Layer Grounding Protocol, an Adaptive Dual-Path Dispatcher that scales reasoning depth based on query complexity, and an Event-Triggered Verification-and-Repair engine. Crucially, the repair mechanism operates via typed state transitions that include a formal provenance non-amplification guarantee, preventing the introduction of unverified content during error correction. By targeting specific failure modes—including unsupported intermediate reasoning, phantom grounding (citation-backed entity hallucination), over-reasoning, and repair-time amplification—LedgerMind improves both the accuracy and the trajectory-level faithfulness of multimodal large language models (MLLMs). Experimental results across various multimodal reasoning benchmarks demonstrate that this ledger-based approach mitigates the tendency of agents to rely on language priors or accidental error cancellation, instead ensuring that every step of the reasoning process is grounded in verifiable, tool-produced evidence.