A ComputeImage is one file. It is divided into six inspectable strata. Each stratum has a purpose. Each is named.

1. Metadata — the model's identity, the source digest, the schema version, the producer, the producer's commit. The metadata is what makes the artifact a content-addressed object, not a renamed checkpoint.
2. Logical tensors — the model's logical view. Shapes, dtypes, semantic families, roles. The logical tensor is the model's claim.
3. Physical layouts — how the tensors are stored. Tile sizes, alignment, padding, memory tier. The physical layout is what the runtime reads.
4. Execution views — how a backend consumes the layouts. Lane, provider, dispatch path. The execution view is the contract by which the runtime executes.
5. Plan and receipts — the admitted plan and the receipts that admit it. The plan is durable. The receipts are durable. Together they are the evidence the runtime has the authority to execute.
6. Payloads — the tensor bytes, aligned to a 16 KB page boundary so the runtime can mmap them without parsing. The payloads are the data.

The six strata are not layers in a stack. They are aspects of one artifact. A cimage has all six or it is not a cimage.

A model is not a single identity. Three identities are preserved across the journey, and the site names all three because conflating them is a category error.

• `source_artifact_digest` is the content hash of the source package as it arrived. The bytes. The same bytes, the same digest.
• `semantic_graph_identity` is the content hash of the canonical computational graph, recovered under a declared schema, declared operator semantics, and a declared normalization policy. The declaration is part of the identity.
• `computeimage_artifact_digest` is the content hash of the produced .cimage bytes. The artifact. The same plan, the same payload bytes, the same digest.

The site does not use the phrase "same effective computation." The site names the declaration that produced each identity. A reviewer who wants to verify an identity can do so against the declaration, the source, and the artifact bytes.

A plan is legal for a target if every operator it references admits a legal lowering for that target, every tile size is admitted, every memory tier is admitted as a target property, and every KV-cache policy is permitted. Legality is decided at compile time. The artifact carries the legalization report.

The capability map names which targets are implemented in the source, which qualify against fixture tests, and which have been validated end-to-end with a receipt. A cimage is legal for a target in the sense of "this cimage was admitted with a legalization report for that target." It is not legal in the sense of "this cimage will run correctly" until execution has produced a receipt.

The v1 corpus does not contain a legalization report. The site names the gap. The cimage on the Specimen page has no execution target — it is a test fixture, sanitized.

A cimage is admitted by a receipt. The receipt is durable. The receipt carries identity, fencing generation, deadline, artifact digest, numerical policy, route, and outcome.

The v1 corpus does not contain an execution receipt. The ExecutionReceipt type is bound to a follow-on ADR (per ADR-033, ADR-035). The site does not display a receipt it does not have. The Evidence page names the gap.

A failure receipt is the same shape. The failure_class field is required by the type, not optional. The constitutional principle — failure is first-class — is enforced by the type, not by a convention.

The cimage file format is described in the engine's cimage module. A developer who wants to inspect a cimage opens it. The header is JSON. The payloads are 16 KB-aligned. The file is mmap-able.

A reviewer reads the header. The header names the source, the model family, the tensor records, the legalization report (if any), the compilation events, the search trace, the model manifest (if multi-model), the execution plan (if heterogeneous). Each named field is auditable without reopening the source model.

The Specimen page (/computeimage/specimen/) is the artifact browser. The ComputeImage page is the chapter. The chapter teaches the concept. The Specimen exposes the bytes.