FLINT Deterministic Work Engine - Deterministic AI execution with validation before commit

I built FLINT around a basic engineering problem: generating an answer is not the same thing as proving what executed, what changed, what stayed fixed, and why a result was allowed to commit. This launch demonstrates that distinction as a live workstation. The operator path is: Identify source role → find source address → read typed boundary → state the question → declare candidate + invariants → run only needed evidence → compare to a frozen baseline → issue OPEN/HALT → retain the receipt. In the demonstration, the workstation applies that process to a Transformer causal-mask boundary. With the causal mask enforced, a future-token perturbation produces max_abs_diff = 0 and the workstation returns OPEN / AUTHORIZED. Bypassing the mask under the same intervention produces a nonzero delta and HALT / BLOCKED. The release also includes a deterministic pixel witness. The committed execution is projected into a canonical pixel artifact and cryptographically bound through framebuffer and tile hashes. A separate one-byte tamper control demonstrates that representation corruption can be detected independently from semantic execution failure. The larger architecture separates the planner from the work engine. A planner such as Astra can propose, sequence, and configure actions through the operator bridge, but the downstream FLINT system retains execution, validation, commit/block authority, persistent artifact state, and receipt generation. That distinction is deliberate: the existence of the Work Engine does not depend on what percentage of a successful workflow is attributable to a particular planner. Another qualified planner can address the same operator contract while the persistent execution and validation substrate remains underneath it. The public release, source, receipts, browser-validation evidence, hashes, demo, and walkthrough are archived at: https://doi.org/10.5281/zenodo.2...⁠� Related technical record: https://doi.org/10.5281/zenodo.2...⁠� I’m particularly interested in feedback from people working on AI infrastructure, agent systems, runtime verification, reproducibility, formal methods, technical assurance, and regulated or auditable AI workflows.