REA review What model headers and file hashes can prove
Twelve source-provider checks examine REA GGUF and safetensors metadata, tensor totals, header identity and refusal paths without loading a model.
AgentIndex · Published 2026-11-05 · Updated 2026-11-05
REA connects agents to reverse-engineering tools through a CLI and MCP server. Its range includes application structure, native binaries and runtime observation. For teams handling model artifacts, it also offers a smaller entry point: inspect GGUF or safetensors declarations without loading the model or installing a decompiler.
We tested that file-inspection provider. It recovered known declarations from artificial files, distinguished header identity from full-file identity, marked an incomplete metadata preview and refused selected malformed inputs. The useful selection question is what those facts establish before a model enters another workflow. A parameter total does not establish model quality, and a matching header does not identify the weights.
Fixed source and twelve provider checks
The review pins commit c7fce66035a4eba490157b4ec3f43b1a7515aa99, with package version 6.8.0, on November 5, 2026. The repository uses MIT for its own code; external tools have separate requirements. Its declared Node versions are 22.x from 22.23.4, 24.x from 24.22.1, or 26.12.1 and later. Native-analysis providers are additional dependencies for other workflows. Project requirements
Tests ran on macOS arm64 with a temporary Node.js 26.12.1 runtime. The fixture transpiles selected TypeScript modules and calls the real ModelArtifactProvider, including its real file reads and output validation. It does not run agent setup, the CLI wrapper, MCP transport or a native analysis engine. No model code or inference was executed. The reproduction files record dependencies, imported-source fingerprints and observed results.
| Specimen or operation | Observed result |
|---|---|
| 152-byte safetensors fixture without a model extension | Detected format; six F32 parameters, 24 tensor bytes |
| Header-only inspection | Read 128 bytes; full-file digest remained null |
| Full-file inspection | Read 152 bytes; digest matched independent SHA-256 |
| Change one payload byte | Same header digest; different full-file digest |
| 344-byte GGUF v3 fixture | Six parameters; 312-byte header; tensor data at offset 320 |
| Ten tokenizer entries | Eight values returned; preview explicitly incomplete |
| Duplicate JSON keys or inconsistent tensor extent | Refused |
| GGUF version 999, short pickle marker or truncated extent | Refused with a diagnostic |
| Signal cancelled before inspection | Returned cancellation |
The specimens were tiny constructed format examples, not usable trained models. They make expected values independently calculable, but do not measure performance on a multi-gigabyte checkpoint.
Header identity and weight identity are different
For the safetensors specimen, header mode read the eight-byte prefix and padded JSON header, totaling 128 bytes. The file contained another 24 bytes of tensor payload. The result recorded a header digest, a header read scope and no full-file SHA-256. Full mode read all 152 bytes and matched a separately computed digest. Provider implementation
We then changed the last payload byte while preserving file size and declarations. Header identity stayed the same; full-file identity changed. This is a direct reason to choose the hash mode according to the question. Header mode can inspect declarations with less reading. To record which exact file a later stage should consume, a header digest alone is insufficient.
A full digest still does not explain what the weights mean or demonstrate that a model will load in a particular runtime. Its value is identity: another independently obtained digest can be compared against it. This review did not compare against a publisher's checksum or establish provenance for any real model.
Derived totals are declarations with constraints
Both specimens declared an F32 tensor shaped 2 by 3. The provider reported six parameters and 24 tensor bytes, matching independent arithmetic. Shapes and parameter totals use decimal strings in the report, which avoids treating potentially large counts as rounded floating-point values. Model result schema
Changing the safetensors extent to 20 bytes while retaining the same shape caused refusal. That supports the selected size-consistency boundary. It does not prove semantic correctness of a tensor name, training history or compatibility with a particular architecture implementation. A declaration can be structurally consistent while still unsuitable for the intended application.
The GGUF header ended at byte 312, while aligned tensor storage began at byte 320. The header digest matched our independently constructed prefix and excluded the alignment gap. Consumers should retain the reported offsets and scope together with a digest rather than relabeling every digest as a whole-file checksum.
A vocabulary preview is not a vocabulary export
The artificial GGUF metadata contained ten tokenizer strings. REA retained a declared count of ten, returned eight preview values and set preview_complete to false. Report coverage identified array values as a preview. A consumer that reads only the returned array length would incorrectly conclude that the tokenizer has eight entries.
This matters when agents summarize large artifacts. Ask the agent to preserve the declared count, completeness flag and limitation when it presents a result. An omitted value is unknown from this response, rather than absent from the file. The same principle applies when exporting evidence into another report or indexing pipeline. Model-file inspection semantics
We did not test large vocabularies, every metadata type or unknown tensor layouts. The eight-value observation belongs to the selected array path and pinned version, not a universal model-format limit.
Refusals are useful adoption evidence
Duplicate safetensors keys and a shape/byte mismatch were rejected. An unsupported GGUF version preserved a diagnostic instead of being accepted as a known layout. A short pickle marker was refused, and an overlarge declared header in a truncated file failed without constructing that requested extent in our fixture. Safetensors reader
These are selected negative tests, not a comprehensive parser audit. Likewise, the cancellation case begins with an already-cancelled signal; it does not establish recovery from cancellation during a large full-file read. Report a refusal as a workflow outcome, retain its reason and decide whether another supported format or tool is required.
Where REA fits
REA is worth evaluating when an agent needs structured, attributable observations across artifact types. This particular provider is useful for inventorying model declarations and deciding whether a full identity check is needed. Deep native decompilation, browser capture and Android observation have different prerequisites and effects; success here does not validate those routes.
For an initial trial, inspect a known artifact, compare expected declarations and choose the hash scope explicitly. Agent-visible results may still be sent to the configured model provider even though inspection runs locally. For generated-artifact contracts, see the datamodel-code-generator review. For the storage side of a persistent agent workspace, see the PI-Desktop review.