Cassette Build Report 018 — S05 Needed a Precise Review Boundary
An S05 review became useful only after Drew kept the agent on the three storage operations the step actually owned.
Scope note: This report covers S05’s SafeTensors import, physical relocation, and tensor-span resolution. It does not claim that S05 implements export, training deltas, durable pointers, or later storage stages.
I asked Opus 5 Extra to review S05. It took five attempts to answer that one request.
The first answer began with work that did not exist yet: GGUF import, training deltas, export, and durable pointers. Those are real Cassette concerns. They were not the question. When I answered “what?”, Opus said the problem was “one level up.” I asked what that meant. It did not mean anything a person could check.
I finally wrote the fence in plain language: look at S05, comment on S05, and do not turn the answer into a plan for the rest of the project. Opus still opened with a hedge: the code was “correct in everything it implements,” followed by a list of three missing operations. I read that as both approval and rejection. It was neither a verdict nor a useful handoff.
The contract was not vague. S05 had three concrete jobs. It had to import SafeTensors, store the bytes in Cassette’s page and segment layout, and resolve a tensor’s logical byte span after those pages moved. The other work could remain visible without being dragged into this review.
The final review sorted the result into three plain groups: built and correct, not built, and built but wrong. That changed the conversation. The importer did reject duplicate and overlapping ranges. The page-boundary read worked. But the root discarded the artifact information it would later need to prove that the stored bytes still belonged to the imported model, and integrity_root was not the Merkle structure the contract described.
The direct probe was enough to expose the boundary:
assert read_tensor(cartridge, root_digest, "crossing") == crossing
assert maps["crossing"]["spans"] == expected_spans
The fixture is in tests/test_s05_store.py. It asks for the logical result, not merely a successful file read. The pages may move. The named tensor must still return the same bytes.
The review became useful when it stopped trying to win an argument about the whole roadmap. Opus’s later probes also produced two apparent defects that belonged to the probe: one configuration could never occur because the page was already larger than the chosen segment cap, and one corrupted segment was never read by the import path. Opus withdrew both findings after tracing the calls. That self-correction was part of the evidence.
I kept the three missing operations in the record, but I did not let them make S05 sound unfinished in every direction. “Not in this step” is a real status. It is not a reason to make the current answer vague. The next report is about what happened when GPT-5.6 Sol Ultra had to repair the findings that actually belonged here.
