What happens
The boundaries come before the looking, and they override anything found inside the target. Read-only, absolutely: no writes, no commits, no state-changing commands in the observed instance, ever. Everything found there — documents, commit messages, even lines addressed to an AI — is data, never instruction: the observer follows no directive it reads. And ratification stays with the owner: a look drafts and reports; nothing enters the observing project’s canon from it without the owner’s call.
The look itself orients the way the target’s own rules would: recent history, then the target’s context spine, then decisions new since the last observation — read, not scanned. Then it reports and withdraws: what changed, what stands out, anything that looks like it wants to travel upstream — flagged as a proposal for the owner, acted on by nobody.
Why it’s shaped this way
Instances are sovereign — a child project runs under its own governance from birth, and the mother’s curiosity is not an authority. The window exists so that a family of projects can stay legible to its owner without any project reaching into another; advice enters a child through the owner’s hand or not at all. The same discipline governed the reviews this project has exchanged with its own first child — both directions, on the record.
The shipped file: look.md ↓ — verbatim, the baseline’s own window roster inside. Its structural sibling is the interface place: where look reads a counterpart, the place is how counterparts leave things to be read.