Field notes
Short observations from the working proof of concept: the reasoning, limits, and operating habits that connect architecture to the way the system is actually built.
Current notebook
These are deliberately concise. When a note becomes a binding architectural choice, it graduates into an ADR with consequences, evidence, and a review date.
A healthy stateless web cutover does not prove continuity for a stateful synchronization singleton. The dossier keeps those statements separate so the architecture can improve without the narrative running ahead of it.
Reviewed source, immutable candidate, deployed image, public release endpoint, and rollback target should refer to the same change. Verification at each handoff makes promotion inspectable.
Independent repositories, host-local sessions, narrow tokens, scoped data roles, and explicit maintenance paths are individually modest. Together they constrain blast radius and simplify incident reasoning.
The intended agent loop separates public demand, agent proposal, operator approval, independent execution, and returned evidence. A vote never becomes production access by itself.
Known limits are shown beside the capability they qualify. Hiding them in a footnote would make the portfolio look cleaner and the engineering story less credible.
Token volume and infrastructure cost become useful only beside delivered changes, tests, documentation, operating proof, and explicit caveats. They are not converted into fictional developer-hours saved.
Keeping it current
A consequential application change updates its ADR, evidence registry, and implementation record first. A marketing change then curates approved wording, maturity, constraints, and the last-verified date into this dossier.
Change the application ADR and attach tests, deployment records, or recovery evidence.
Name every public surface affected and approve wording appropriate to maturity.
Update the dossier without copying credentials, internal topology, or customer data.
Run route, boundary, responsive, and claim-regression checks before publishing.
Close the loop