Trust and authorization

"Trust" names two separate things, and design keeps them apart. Trust to act is how much the box may do without asking, along the ladder below. Trust to keep is the promise that the box never throws your things away, whatever it does with them.

Paperwork lives on as schema process-fields

The early design wanted every outbound action to carry "paperwork" — impact, confirmation of user intent, risk. Command cards (which carried explicit authorization fields) are gone, but the idea did not die (ruling 9): it lives in the schemas. Schemas include fields that must be filled in not because they are essential information, but because filling them is essential process — deliberation the agent is forced through before the artifact is valid. A question card's memo (why am I asking?) and directive (what happens with the answer?) are the pattern; new action-shaped schemas should use it. (2026-10: both fields are optional in src/schemas/question.ts, so the question card states the pattern but does not enforce it; a new schema that wants the deliberation should make its fields required.) The rest of the trust surface is the question/confirm gates and the commit history (durability-and-provenance.md).

Trust to act: question → confirmation → automatic

The system's relationship with the user evolves (design intent, correct but not well filled out in the implementation — ruling 10):

1. Question — "what should I do with this?" No idea; need guidance.
2. Confirmation — "I think I should do X. OK?" Have an idea; need approval.
3. Automatic — just do it. Confident, trusted pattern.

Transitions happen through answers that teach: a one-time "yes, send it," a "yes, and don't ask again for this sender" escalation, or an explicit standing instruction. Learned trust gets recorded in rules/guide material so the escalation persists. (Shipped 2026-07: a question's learning: field names a guide, briefing, or personality sink, and the follow-up job records the answer there as a basis: user-stated belief — ../questions.md.) Keep this as the umbrella principle when building approval flows; it is design to fill, not history.

A question, the first rung, is never auto-answered, because silence is not consent. It is nudged once and then expires, so an unanswered ask stops claiming attention. When the person hands a question off ("Decide for me"), or it waits too long, the box may decide it conservatively and reversibly, recorded as provisional: not an answer, never the person's words, and replaced by their answer whenever they give one (../questions.md, "Figuring it out"). A dismissed or expired question stays answerable, because not answering now is not a decision (../questions.md).

Trust to keep: nothing is thrown away

A thought is never dropped: input the box cannot place is stored and waits for a person (spirit.md's "nothing should just disappear"; ../chat/quick-chat.md). The record is kept so an action can be undone or moved again, which is what lets the box act forward on some things without asking (representation.md, durability-and-provenance.md). A record's reviewed and archived are the boxholder's judgments, so they are set only on the boxholder's say-so (src/schemas/record.tsx).

Quick chat is a compromise, not a pattern to copy. It posts a thought to the chosen conversation above a confidence threshold, with no confirmation step, and the post cannot be undone. That gives up the trust-to-act ladder for frictionless capture, a trade made for another reason: capture must cost nothing in the moment, or the thought is not captured. It is acceptable only because of trust to keep: the thought is never lost, and the record lets the person move it (../chat/quick-chat.md).

The confidence vocabularies

Three scales answer "how sure/authorized is the agent," each domain-specific:

• Trust to act (above) — permission to act.
• Triage confidence — confident / probable / guess; classification certainty, gating whether an item files silently or raises a question (../triage.md).
• Retro belief confidence — hypothesis / low / medium by recurrence; how firmly an inferred belief is held (../glossary.md, retrospective).

(2026-10: the three above are not the complete set. Other shipped scales: belief basis user-stated > feedback > inferred > default; guide confidence confirmed|high|medium|low|hypothesis, with hypothesis excluded from compiled guidance; progress evidence observed|inferred|self-report, with no level recorded without evidence (src/plugins/courseware/progress.ts); procedure-run outcome completed|failed|inconclusive (src/schemas/procedure-run.ts); question states pending|answered|dismissed|expired; notification loudness dot|quiet|loud (src/core/notification/intent.ts).)

No ruling unifies them; treat them as instances of one instinct (act only with warrant proportional to certainty) with deliberately separate vocabularies until a real need forces a merge.