Soul

Soul is Sadie’s working theory of you. It is not a profile you fill out once. It is a living model that accumulates evidence from sources, rewrites, calibration choices, and conversation behavior, then promotes the most reliable signals into confirmed items. You can confirm, edit, or reject any of it at any time.

Soul is used everywhere. Every chat response, every brief, and every discourse opportunity loads Soul as context before generation. Keeping it honest keeps Sadie honest.

Three layers

Soul state. Atomic claims about you, each typed by kind: audience, voice, obsession, allergy, recurring_theme, goal, taste, angle_style. Each item has a confidence score, a state (candidate, confirmed, edited, rejected), and evidence references pointing back to the sources that justified it. Items that stop getting reinforced decay over time.

Each item also shows its origin, review status, evidence count, and last signal date in Memory. Expand Evidence to inspect the refs behind a claim. Expand “Why this confidence” to see the deterministic rationale for its current state: whether it is confirmed, still learning, manually corrected, rejected, single-source, multi-evidence, decayed, or recently reinforced.

Memory also shows a Learning ledger: a recent audit trail of preference candidates, rewrite events, and direct Soul corrections Sadie has seen. It includes pending, accepted, reinforced, dismissed, voice-tier signals, manual additions, confirmations, edits, rejections, and confidence changes, so you can tell what is shaping Soul even after a signal is no longer waiting in the review queue. Posted-angle outcome rows are called out as landed examples or prediction misses, with observed metrics and pattern tags, so audience feedback is legible as learning rather than a hidden calibration table. Pending signals also expose a compact “Why this signal” rationale before review: threshold/readiness, source trail, active Soul relation, and freshness. When you reject a Soul item or dismiss a pending learning signal, you can choose a reason and leave a correction note; both are preserved in the ledger as counter-evidence.

When a pending learning signal appears to contradict an active Soul item, Memory shows that conflict on the affected item as well as in the queue. The item card names the candidate, its signal count, and the overlapping terms so you can decide whether to reject the signal, edit the belief, or let the new evidence revise it.

Each Soul item also shows its latest review action in place. If you corrected, confirmed, rejected, added, adjusted confidence, accepted a new signal into, or reinforced that belief, the item card shows the newest action, date, confidence delta when relevant, and any rejection reason or note. You do not have to scan the global ledger to understand why a particular belief changed.

Voice portrait. An 8-dimension model of how you write: argument architecture, sentence rhythm, register range, characteristic moves, vocabulary fingerprint, avoidances, emotional texture, and framing habits. Each dimension carries a description, confidence, evidence count, and any accumulated rewrite adjustments from you editing Sadie’s output. Memory shows those rewrite adjustments on the affected dimension, including direction, signal count, strength, last signal date, and recent notes. Voice portrait compiles into a system prompt that grounds every generation.

Soul narrative. A prose portrait in six sections: work context, personal context, top of mind, and three tiers of brief history (recent, earlier, long term). This is the conversational surface of Soul, the version Sadie reads when it needs to ground a response in your life rather than your claims.

The confirm flow

New signals land as candidates. You review them in Memory. You confirm (they become confirmed and carry full weight), edit (they become edited and keep a record of your correction), or reject (they are logged as counter-evidence, including your rejection reason when provided, so Sadie knows not to propose them again).

candidate  ->  confirmed      kept with full weight
           ->  edited         kept with your correction
           ->  rejected       suppressed, counter-evidence logged

Recent Soul reviews also shape Today generation. Confirmed and edited items are rendered as positive synthesis guidance; rejected or lowered-confidence items are rendered as avoid guidance, so manual corrections affect future why you and candidate-angle choices instead of only changing the Memory screen.

Behavioral inference

Soul grows mostly from what you do. Source uploads and weightings. Rewrite patterns. Discourse dismissals (flagged by reason, like wrong_audience or stale_topic). Source curation is visible learning: representative sources teach voice, reference-only choices teach boundaries, and cautious sources become allergy/noise evidence. Onboarding uses a distillation loop of contrastive probes, rewrite drills, and scenario validation rather than asking you to describe yourself. See Self-evolution for how operational defaults tune on top of Soul.

Soul vs. policies

Soul allergies are probabilistic. They nudge the model. Policies are deterministic. They are rules Sadie cannot break. Put behavioral aversions into Soul. Put unconditional bans into Policies.