Sathi · notes from the build

Decision log 0 — written down to be argued with

2026-08-09 · the builder

Most of what this system is was settled by a handful of early decisions. Writing them down does two things: it lets outsiders attack the reasoning instead of the marketing, and it stops us from quietly drifting away from our own rules. This is the format future posts will mostly take.

1 · Archive first; understanding is regenerable. Originals — messages, documents, exports — are kept unchanged, and everything the assistant believes is derived from them. When better models arrive, the same life gets understood better without anyone re-uploading anything. The archive is the asset; the intelligence is a lens.
2 · Deletion beats immutability, permanently. When a person deletes something, it leaves the archive too — provably. An append-only ledger of someone's life sounds like engineering rigor and is actually a trap. This carve-out is the one exception to decision 1 and it is eternal.
3 · One understanding, many surfaces. The phone, the voice call, the future speaker — every surface reads and teaches the same understanding. No per-device memory, ever: a device adds senses, never a second brain. Nobody should repeat themselves to a "mode."
4 · Nothing executes unconfirmed. The assistant drafts the email, prepares the calendar change — and a human taps yes. Authority is a ladder climbed per-domain with explicit, revocable grants and a permanent audit row for every action. Trust is the slow part; we refuse to fake it with defaults.
5 · Operator access is granted, never assumed. We dismantled our own god-mode: the administrator can open a member's content only after that member grants support access — revocable, recorded, visible in their own account. Running the service never requires reading anyone's life. We built this against ourselves first, because operator character is a risk category, and we are operators.
6 · Exit is a tested code path, not a promise. Export produces a sealed, signed archive that can rebuild an account; leaving issues a receipt of what was destroyed and what remains, with the consent basis for each retention; returning with the archive reactivates what peers never revoked. We rehearse it. A right you cannot exercise is décor.
7 · Open weights first — the exit engineered before the entrance. The primary model is open-weights on a rented endpoint: pointing the same system at hardware we operate is a configuration change, not a migration. Stronger closed models exist; we pay that gap for independence, and we say so plainly.

If one of these is wrong, we would rather hear it now, while the blast radius is one household. That is the entire editorial policy of this blog.


Drafted with Claude; decided and lived by us. Argue at info@sathi.ai.