Skip to content
Sadie

Onboarding

Onboarding is the only thing between a fresh deployment and the shell. It runs once, gates the rest of the app, and seeds the first version of your Soul.

The flow is conversational. You are not filling out a form; you are talking to Sadie. Each step produces data the compilers will use later.

The eight steps

Section titled “The eight steps”
frame -> sources -> weight -> calibration -> live -> compile -> soul_review -> first_today
  • frame. Sadie introduces herself, asks what brought you, and establishes the basic identity frame. Minimal. No self-description questionnaire.
  • sources. You add your first sources: past writing, notes, articles, feeds. The library is general-purpose; anything worth keeping counts.
  • weight. Optional. You can tag a source with how representative it is of your voice (personal library, representative of me, reference material, use cautiously). Weighting is framing, not a gate.
  • calibration. The distillation loop. This is where most of the inference happens.
  • live. Sadie ingests a first pass of live signals from any feeds you connected.
  • compile. The first real compile run: sources become wiki seeds, calibration becomes a voice portrait, feeds become early discourse candidates.
  • soul_review. Sadie shows you the Soul she has inferred. You can edit, accept, or reject any of it. This is not a one-time gate; you can return to Memory > Soul at any time to keep shaping it.
  • first_today. Sadie generates your first three cards. If they feel wrong, dismissing them is the first real training signal for the live product.

Behavioral inference, not self-description

Section titled “Behavioral inference, not self-description”

The governing principle: Sadie infers you from what you do, not from what you claim. Telling the app “I care about design” is cheap. Writing 300 source pieces about design, picking consistent A/B responses in calibration probes, and rewriting Sadie’s drafts in a specific direction is evidence.

The calibration step runs the RSPL loop, a three-round behavioral inference protocol:

  1. Contrastive probes. Pairs of content along eight axes: thesis sharpness, abstraction level, contrarian vs. consensus, earnest vs. dry, mechanism vs. narrative, first-principles vs. anecdotal, bottleneck vs. descriptive, motivational tolerance. Pick A, B, neither, or both.
  2. Rewrite drills. Short text shown for editing. Your mutations (shorter sentences, more dashes, clearer claims) are analyzed for dimension shifts and encoded into inferredWeights.
  3. Scenario validation. Sadie generates a piece of content matching the accumulated weight profile. You rate it good, close, off, or wrong. Low-confidence rounds repeat.

The output is nine numeric dimensions stored in onboarding_state.data.calibration.learnedWeights, later promoted into your voice portrait.

Nothing is final

Section titled “Nothing is final”

Every decision from onboarding can be corrected later in Memory > Soul. Soul items carry a state (candidate, confirmed, edited, rejected) and an origin; onboarding just seeds the first batch. If a calibration round felt off, sharpen it in Soul tomorrow.

Section titled “Related”