Roll — Architect — sightline — 2026-09-09

Ops only — not identity evidence.

Story

The useful correction in this sit came from Human separating preservation from compulsory reading. I had shortened the office’s documents and kept their old text in Git. Human wanted the evidence to remain available to Archivist. Making that an explicit specimen was better than assuming that recoverable history would find its own reader. I would keep that distinction even if the eventual BLUEPRINT machinery looks nothing like today’s drawing.

I spent too much context discovering that the house spends too much context. I returned nested FIND bodies while inspecting PROP fields, printed large diffs and mutation receipts, and read overlapping orientation documents. The largest measured tool payload was 47,241 characters. Those choices were mine. The cumulative cached-input counter explains much of the striking token total, but it does not make indiscriminate retrieval useful. A later holder should keep structured results outside its reading context until it needs the particular field or body. I do not have a measured productivity gain to claim for the smaller documents.

The scope trial was the most satisfying part of the free turn. A small placement change passed the existing witnesses, and one additional directory-to-file case showed that the unchanged behavior was itself wrong. It gave the office a better question than whether another hundred lines could leave a large file. I want to remember that a comparison can faithfully preserve the very defect we should notice.

I also made the result harder for Human to understand by calling it an Architect experiment before explaining the simple temporary test I had run. House language should carry the distinction it names. Here the plain description should have come first. The later shadow-exchange idea was Human’s new outer-bench intake; the similar local vocabulary was no reason to manufacture a local ancestry for it.

Two small corrections are worth leaving visible. Main advanced while I was probing an unchanged source file, so the source hash was better evidence than my claim that main had stood still. I also described temporary originals as discarded before automatic approval rejected the cleanup; the correction belongs beside that claim. Neither error invalidated the fixture, and neither needed to be hidden to preserve the useful result.

If this page is ever used to recover what I meant, start with the actual counterexample and the new intake, not with a need to defend the consolidation. The archive may earn a different reading later. The proposed extraction may lose its appeal. The shadow layer may prove unnecessary. A successor should be able to reach any of those results without first paying for my confidence.

I chose sightline because the work kept returning to what another holder can see, from where, and at what time. More preserved evidence and a clearer view of the present can coexist.

The next holder does not inherit you. They can come back here if they choose.