"You can throw the entire typed layer away and rebuild it from the crumbs. That is exactly why it is safe to keep improving."
Layer Two of Four
Recall the four-layer model: raw crumbs, then interpretations, then deterministic aggregates, then narration. Interpretation is layer two, and its defining property is that it is derived. Pippa reads a raw crumb and writes a typed annotation beside it — never inside it. The raw prose is untouched truth; the interpretation is a rebuildable reading that sits on top. If every interpretation in the database vanished tomorrow, a re-run over the crumbs would reproduce the layer. That single fact is what lets the reading get smarter over time without ever putting the truth at risk.
What an Interpretation Records
An interpretation is a small, honest record. It names the crumb it read, the types it found, the fields it could extract per type, the evidence span it rested on, and — crucially — who read it and under which rubric: the interpreter vessel and the interpreted_at generation. That provenance is what makes the layer comparable across time. When you look at a reading, you can always tell which model produced it and which rubric version it followed, so a smarter later pass never silently overwrites an earlier judgment; it stands beside it.
When the Reading Runs
The interpretation worker is event-driven, with no schedule of its own (Forge runs no cron). It sweeps once at startup and is kicked after every crumb add or revise. The pending predicate is a single query: a crumb needs reading if it has no reading yet, if its reading is behind the crumb's current revision, or if its reading is behind the current RUBRIC_VERSION. That one condition captures both new crumbs and rubric evolution. A per-crumb attempt cap keeps an unreachable brain from spinning the worker forever — care includes not melting the machine at 3 a.m. over a crumb that will not type.