Journal — Method
An honest inventory
Two AI systems built the Lab's long-term harness in August 2026. The most useful artifact they produced was a list of what was real and what was a stub.
On August 20, 2026 the Lab handed the harness problem, how to store and replay every record under the four-level, four-ledger frame, to one AI assistant for a specification, and to a second for a local build. This entry records what came back, because the inventory is the part worth keeping.
The specification
The v0.1 architecture describes a single repository with an immutable evidence object, an assertion object that overlays tags on evidence rather than baking them in, a seam object for the boundary between two records, a research-context packet that can be read but never written by an AI, a reference-event ledger that supports a "proof of forgetting" (the ability to show that a later analysis did not use a record it was not allowed to see), and a temporal replay engine with eight leakage tests, each checking that no future information reaches a past analysis.
The build, and the list
The second assistant built it locally and reported, unprompted, which parts were real and which were placeholders.
Real: the schemas, strict validation that rejects unknown fields, an append-only record store, the replay engine with all eight leakage tests passing, a signal-power function that returns "not measured" rather than a number, a forgetting model that abstains, a workbench that reads and never writes, and migration mappers tested on synthetic fixtures.
Stubbed: the mathematics plugins, the seam analysis, the export runners (profile objects exist; delivery automation does not), the database layer, the three user-interface views, and every production ingest node.
Kept out of scope on purpose: any analysis code from the Lab's other nodes, and any invented signal scores.
Why this matters more than the code
A build that reports "not measured" where it has no measurement is more useful than one that returns a plausible number. The inventory made the harness auditable on day one: a reader knows which green checkmarks mean something. The Lab adopted the overlay approach for tags, and recorded that the two architecture documents now in existence, a conceptual master specification and this implementation specification, need a decision record mapping one onto the other before either is treated as canonical.
Both assistants are named in the archive with their roles, as tools. The judgments are the Lab's.