# Phase 1 Epic — Trusted Memory Loop

## Epic summary

Deliver a private, text-first experience in which a user can complete a guided interview, receive source-grounded memory proposals, review them, and see accepted items on a timeline. This phase validates trust and usefulness before advanced learning or social features.

## Goal and non-goals

**Goal:** One user can reliably create and control their own first structured life-memory record from a conversation.

**Non-goals:** audio recording, public sharing, automatic confirmation, capability scoring, integrations, and autonomous follow-up.

## Personas

- **Memory owner:** wants to preserve experiences without losing nuance or control.
- **Cautious reviewer:** wants to inspect sources and correct AI suggestions before they become part of a timeline.
- **Product/operator:** needs clear job states and auditability without reading private content.

## User journey

```text
Consent & topic → guided interview → save raw turns → extraction queued
→ candidate review with source evidence → accept/edit/reject/defer
→ confirmed event/person/relationship → personal timeline
```

## Work packages

### P1-01 — Project and privacy foundations

**Outcome:** authenticated private workspace, consent capture, ownership checks, audit trail, and local/development data policy.

**Acceptance criteria:**

- A user cannot read or change another user's records.
- Consent is captured before the first interview and is tied to a policy version.
- Key actions create audit events without logging full private text by default.
- Test fixtures contain synthetic, not personal, data.

### P1-02 — Interview workspace and raw source capture

**Outcome:** text interview with create, save/resume, end, and manual-entry capabilities.

**Acceptance criteria:**

- The user can set a topic and skip/stop at any time.
- Each turn has stable sequence/order and remains available after session completion.
- Host messages follow `AI_HOST_SPEC.md` policy and display no claims as confirmed.
- A completed session is eligible for extraction without changing original turns.

### P1-03 — AI Host policy and model gateway

**Outcome:** versioned, observable AI host and structured-output gateway.

**Acceptance criteria:**

- Host output validates against a schema and is resilient to malformed/provider-failure responses.
- Context is scoped to the current authorized user and necessary session data.
- Policy/model versions and request outcome metrics are recorded.
- Prompt-injection and unsafe-overreach test cases pass the agreed evaluation bar.

### P1-04 — Source-grounded extraction pipeline

**Outcome:** queued extraction creates event, person, relationship, and time candidates with evidence.

**Acceptance criteria:**

- Jobs are idempotent, retryable, and expose pending/succeeded/failed state.
- Every candidate has kind, payload schema version, confidence, and at least one source span.
- Extraction never creates confirmed timeline records directly.
- Fuzzy dates retain original expressions and precision.

### P1-05 — Candidate review and correction

**Outcome:** grouped review queue with Accept, Edit & Accept, Reject, and Later actions.

**Acceptance criteria:**

- Reviewers can open supporting text for each candidate.
- An edit creates a traceable user-authored revision; original candidate/source remains intact.
- Rejected/deferred candidates are excluded from default timeline/search projections.
- Same-name people are never merged automatically.

### P1-06 — Confirmed timeline and memory details

**Outcome:** private timeline and detail pages for confirmed events, people, and relationships.

**Acceptance criteria:**

- Timeline handles unknown and approximate dates without false precision.
- Detail views show origin/source links and latest review state.
- Users can directly correct a confirmed record and preserve history.
- Only confirmed or user-authored content appears by default.

### P1-07 — Quality, safety, and release readiness

**Outcome:** tested core loop and a small controlled beta readiness decision.

**Acceptance criteria:**

- End-to-end test covers interview → extraction → review → timeline.
- Authorization, provenance, candidate lifecycle, deletion/export workflow, and job failure tests pass.
- Usability test confirms the Phase 1 success measures in `MASTER_PLAN.md` or documents exceptions.
- Open decisions blocking beta are resolved or explicitly accepted as risk.

## Dependency order

`P1-01 → P1-02 → P1-03 → P1-04 → P1-05 → P1-06 → P1-07`

P1-03 may prototype alongside P1-02, but production extraction cannot begin until P1-01 consent/audit foundations are in place.

## Definition of done

The epic is complete when the entire user journey works in a controlled environment with real authorization boundaries, source-grounded candidate proposals, explicit confirmation, stable timeline projections, operational visibility, and the Phase 1 success criteria met or consciously waived in `DECISIONS.md`.