roadmap: Phase 0 exit-gate — the dogfood criteria that unlock Phase 1 #73

Goal

A single, visible checklist that decides when Phase 0 is done — and therefore
when we are allowed to start implementing Phase 1 ("make it a brain"). Per
§13, the gate is not a
dashboard metric — it is "do you reach for it without being told to."

Why now

The "we don't know what we're building" problem is really "the backlog doesn't
express the phase plan." The architecture is explicit: the loop must earn the
right to the brain. This issue makes the boundary concrete so Phase 1 design
issues (the type: architecture set this was filed alongside) don't get built
prematurely — design is allowed now; implementation waits for this gate.

Exit criteria (§13.2 — verbatim, made checkable)

• Founder uses it daily, unprompted, for 14 consecutive days.
• capture → recallable in < 10s (spoken thought findable before you put the phone down).
• At least one reminder fires reliably and on time, end-to-end, for a week with zero misses.
• ACK latency p95 < 400ms — capture never feels like it's "thinking".
• You have asked it something you'd forgotten and it answered correctly at least once (the "oh, it actually works" moment).

What this gate explicitly does NOT require (deliberately faked in Phase 0)

No tiering · no triage (one dumb reminder rule) · no graph · no consolidation ·
no multi-agent · no encryption tier. These are Phase 1/2 and are tracked by the
type: architecture design issues this gate blocks.

How to do it

• Add a short ## Phase 0 status section to README.md (or architecture/roadmap)
  tracking the five criteria with dates.
• Wire a tiny ACK-latency log/metric on POST /capture so the p95 < 400ms claim
  is measured, not asserted (the capture path is in
  apps/api/src/modules/memory/).
• Keep a running note of daily-use streak + the first correct "forgotten" recall.

Acceptance criteria

• All five criteria are tracked somewhere durable in the repo.
• ACK latency is actually measured (not guessed).
• When all five are met, this issue is closed — and that close is the signal
  to start Phase 1 implementation.

Sequencing

This gates implementation of every Phase 1 architecture issue below. The
design work in those issues can proceed in parallel; only coding against them
waits on this close.

🔗 Sequencing

• 🚧 Blocks: feat(memory): encoder stage — real 5-type classifier (replace hardcoded episodic) #96, feat(memory): encoder stage — structured summary generation + re-embed off summary #97, feat(memory): encoder stage — entities[] + tags[] extraction #98, feat(memory): encoder stage — derive real salience + confidence (replace constants) #99, feat(memory): encoder stage — populate per-type typeAttrs #100, feat(memory): encoder stage — attention gate (filler / self-cancel / near-dup) #101, feat(memory): encoder stage — multi-object split (one capture → N objects) #102, feat(memory): encoder_version stamping + versioned re-encode backfill #103, feat(platform): model router core — class config (fast|strong) over the LLM seam #105, feat(platform): per-user + per-route budget caps (reject over-budget calls) #106, feat(platform): per-call cost computation + telemetry in /admin/stats #107, feat(platform): wire encoder→fast, recall→strong + assert ingest is fast-only #108, feat(notifications): unified delivery pipeline + idempotent delivery key #123, feat(notifications): triage scorer + class thresholds (INTERRUPT/TIMELY/DIGEST/SILENT) #124, feat(notifications): quiet hours / circadian hold (default 22:00–08:00 local) #125, feat(notifications): token-bucket backpressure (refill ~4/hr, cap 6, INTERRUPT exempt) #126, feat(notifications): digest coalescing into one morning-slot (08:30 local) ping #127, feat(notifications): deadline-decay urgency ramp ([7d,1d,3h,30m]) + completed cancel #128, feat(notifications): feedback loop — outcome → salience delta + 3-dismissal mute #129, feat(memory): per-type half-life heat function (replace global 7-day stub) #130, feat(memory): recompute heat on access — instant re-warm to HOT #131, feat(memory): scheduled per-user heat recompute + re-tier job (pg-boss) #132, feat(memory): hysteresis dead-band + per-tier dwell (anti-thrash) #133, feat(memory): forgetting GC — TTL-accelerated decay, never deletes source #134, feat(recall): keyword doorway — Postgres FTS + pg_trgm retriever (§05.3 β) #136, feat(recall): temporal doorway — B-tree range + last/first-occurrence retriever (§05.3) #137, feat(recall): graph doorway — entity/edge model + recursive-CTE associative walk (§05.3 γ) #138, feat(recall): query planner — classify {lookup|temporal|summarize|open} → retrieval plan (§05.2) #160, feat(recall): RRF fusion (K=60) over the heterogeneous doorway lists (§05.4) #161, feat(recall): cross-encoder rerank (top ~30) + δ·heat bonus, behind a reranker seam (§05.4) #162, feat(recall): grounded reconstruction — answer from top-k only, cite every claim (§05.5) #163, feat(recall): citation verifier — strip uncited claims; fold in #67 (§05.5) #164, feat(recall): confidence bands (≥.75 answer / .45–.75 hedge / <.45 miss) — replace RECALL_MIN_SCORE (§05.6) #165, feat(recall): spreading activation on a miss — ≤3 iters, graph walk + widen + re-fuse (§05.7) #166, feat(recall): recall-is-rewrite — re-warm heat + optional summary rewrite, async (§05.8) #167, feat(recall): tiny held-out eval — prove graph-walk recall beats pure-vector (§05; Phase 1 exit) #168 — these wait on this one.