Part 7 · 7.2 Analytical Rationale
The analytical frame: terms of reference and sign-off
How do we define and fill systemic gaps — what a Cell gives an analytical practice that an undifferentiated log would not.
§The standard move this mirrors
Hand a defined sub-task to someone with a written brief; let them work inside their own scope; have an independent reviewer grade the result against the brief before anyone downstream is allowed to rely on it; only then is the finished piece released, carrying its own record of who did it, who checked it and who signed it off. This is not a NestVM invention — it is how a newsroom handles a fact-check, how an audit firm handles a workpaper, how an intelligence shop handles a tasked sub-report. The vocabulary is standard:
| standard term | its Cell equivalent |
|---|---|
| terms of reference / brief | the Charter — title, purpose, the question, and criteria with weights and minimums |
| scope of engagement | the policy — exactly what the Cell may hear (accepts), what it may do (capabilities), and how much (budget) |
| the analyst's draft | the Cell's result — the producer bind's one delegated answer |
| independent quality review | the assessment — a different bind grades the result against the Charter's criteria |
| partner sign-off / release | attestation — the human's cell.reviewed decision |
| the released deliverable, with its audit trail | the accepted export — the result in an envelope naming the Charter, the assessment and the review, by their log positions |
§The gap this fills
An undifferentiated log has no natural way to say "this text came from a controlled sub-task with a stated brief, was graded, and was accepted by a named decision" — without a Cell, a delegated answer is just another fact, indistinguishable in kind from the person's own typed words or from a bind's own local computation. Two systemic gaps a Cell closes:
Gap 1 — provenance of a delegated claim. In most working documents, if you ask "where did this sentence come from", the honest answer is often "somewhere in the chat history, we're not sure exactly which prompt." A Cell makes the question answerable by construction: every accepted finding names its Charter (charterRef), the offset of its assessment, and the offset of its review. You are never more than three lookups from the brief that was given and the person who signed off.
Gap 2 — the line between "the model said something" and "the study accepted it". Language models produce text continuously; a study should not. A Cell's lifecycle enforces a hard stop between the world answered (running → awaiting_assessment) and the study may use this (sealed), with a human decision as the only door between them. Nothing about a Cell being open, answered, or even graded 0.9 makes its content usable elsewhere — only the review does.
§Decomposition, barriers, and independence — three moves this section keeps meeting
A Cell is where three moves recognisable from ordinary research method become mechanical constructs:
- Decomposition into a controlled sub-task. The four vectors of a press-release study (Part 4) each open their own Cell — the study does not read the whole release once and hope for a coherent answer; it asks one narrow question per Cell, with one brief per question.
- A barrier before synthesis. A Cell's own inner report — its evaluator bind — cannot fire until both the Charter and the result exist; the study's outer report cannot fire until all four vectors have returned. Nothing is written up early.
- Independence of the reviewer. The machine enforces, not merely recommends, that the bind grading a Cell's result differs from the bind that planned it and the bind that produced it (
assessment: the evaluator must be independent of planner and producer). A single actor cannot mark its own homework — this is checked, not just written in a style guide.
§What is still the analyst's job, not the machine's
The machine grades the existence and shape of a result against the Charter's criteria; it does not grade whether the finding is actually right. In the sessions behind this book the evaluator is a deterministic stand-in that scores any non-empty, well-shaped result 0.9 and an empty one 0.2 — a liveness check, not a fact-check. Reading the findings, catching a status that looks wrong (an announced number marked "happened" when it is really a reported number), noticing a missing quote — that is exactly what attestation is for, and it stays the person's to do, walked through concretely in 4.3 and 7.4.