Case studies

A case study here is a replay, not a story. Each one names the version it ran against and separates what was observed from what was only tested.

#Case 1 — A deck request that started no process

This is the case that justifies the preflight. The same request against version 0.0.20 consumed 9.4 million tokens before failing, because nothing checked whether an engine existed for the requested format before work began.

LimitWhat this case does not show: a successful deck. At the version it records, that format had no engine, and the refusal **is** the case — it is what preflight exists for. Version 0.0.23 has since added a deck engine, a memo engine and an atomic bundle; this run predates them, and rewriting it to match today would turn a dated observation into a claim 1.

#Case 2 — A desk that ran, and what it cost

The second header state is the honest one: when the vendor CLI does not publish its quota, the application writes that it does not know rather than inventing a number. No dollar figure appears anywhere in the product.

#Case 3 — A workbook that was never produced by an agent

The workbook path is complete in code: an MCP call deposits a typed contract, the core composes the file, writes it, reopens it with a second library and finds its formulas again. Every step of that chain has a test.

Still missingNo agent has produced a workbook in a live run. A .xlsx file does exist in a data room from an earlier version — it was written by a Python agent in the 0.0.20 run, not by the engine. A file found on disk is not evidence until you know what wrote it 2.

What would close this case: replay a concrete workbook request on a stable screen, watch the accounting role call the tool, and watch the module appear on the canvas. That is the first item on the next OS pass.