Follow one task from intent to replay.

Stage 01

Intent becomes a contract.

Before any logic runs, the task is wrapped in five envelopes. Semantic says what it means. Temporal sets logical time and checkpoints. Lineage records what it came from. Policy states what it may do. Execution states how it must run, including replay tolerance. Once validated, they bind the task for its whole life.

Five envelopes bind each operation before dispatch.

Stage 02

Rules run before the hardware does.

Checks run as in-process calls in a fixed order: global, sovereign-domain, tenant, temporal, accelerator eligibility, exceptions. An audit record is written before release. If a rule fails, no execution assignment is created.

Fig. 01 · Choose a rule to fail, then run the task.

Zero-hop describes the kernel boundary: validation, policy, scheduling and audit sit in one process. It does not mean zero physical network or device transfers.

Stage 03

The same lane, every time.

Routing considers task meaning and the published capabilities of eligible hardware. When several devices qualify, a fixed tie-break decides: rule priority, tenant locality, memory-latency class, then lexicographic device ID. Shuffle the arrival order and the assignment does not change.

Fig. 02 · Shuffle the arrival order.

Stage 04

Every step leaves a sealed trace.

Each transformation becomes a node in a directed graph. Its SHA-256 anchor covers the node ID, its sorted parent IDs, its envelope ID, its policy-decision ID and its temporal checkpoint. Alter one record and its anchor no longer matches. Fig. 03 below also chains each anchor to its parents' anchors, so the change shows downstream.

Stage 05

Replay, then compare.

An isolated replay context reconstructs past state without touching production. It then compares the execution assignment, policy outcomes, event sequence and lineage hashes. Numerical drift on GPUs is judged against a recorded tolerance.

Fig. 03 · Run the record, replay it, then alter one entry.

Hashes give a reference for comparison. They do not by themselves prove that inputs were true, establish independent custody or show that a model was correct. Replay is an evaluation objective within a pinned environment.