@gonzih/meet-the-one-ai 1.0.0

package/basis/layers/L2_CONVERSATION.md

@@ -0,0 +1,159 @@
+# Layer 2 — Conversation (Intake LLM)
+
+**Goal:** User gets a call from an AI within minutes of verifying. Conversation feels warm, curious, unhurried. Not a form. Not a questionnaire. A real conversation that reaches past the social mask.
+
+---
+
+## Calling Infrastructure
+
+**Twilio Voice + Twilio Media Streams.**
+
+Twilio handles the PSTN (real phone network) — calls actual phones globally. Media Streams pipes bidirectional audio in real time over WebSocket.
+
+Why Twilio over alternatives (Vonage, Plivo, Bandwidth):
+- Best real-time audio streaming support for AI integration
+- Most mature webhook architecture
+- Global PSTN coverage including developing markets
+- Handles both inbound (user calls system) and outbound (system calls user) on same infrastructure
+- One vendor, one integration, one billing relationship
+
+---
+
+## Voice AI Stack
+
+**Vapi.ai for v0.1. Twilio + OpenAI Realtime API for v1.**
+
+### v0.1 — Vapi.ai
+Purpose-built voice AI platform. Abstracts Twilio + STT + LLM + TTS into one product with a single API.
+
+What it handles:
+- Outbound and inbound calls
+- Real-time STT → LLM → TTS pipeline
+- Interruption handling (user cuts AI off mid-sentence — works naturally)
+- Call recording
+- Conversation history across sessions
+- System prompt configuration
+
+Why start here: ship in days not weeks. One integration. Good enough latency (~600ms). Full control later.
+
+Cost: ~$0.10–0.15/minute of conversation. 30-minute intake = ~$4. Acceptable at early scale.
+
+### v1 — Twilio + OpenAI Realtime API
+Direct integration when we need full control, lower latency, or Vapi pricing becomes a factor.
+
+```
+User phone ←→ Twilio Media Streams (WebSocket) ←→ OpenAI Realtime API ←→ LLM
+```
+
+OpenAI Realtime API handles STT + LLM + TTS in one bidirectional WebSocket. Single round-trip. Latency under 500ms. Voice interruptions handled natively.
+
+---
+
+## Conversation LLM
+
+**Claude claude-sonnet-4-6 (Anthropic)** as the underlying model.
+
+Why Claude over GPT-4o for conversation:
+- Superior at nuanced, sensitive psychological content (desire, fear, sexuality, trauma) without triggering or becoming clinical
+- Better at holding a persona across long context
+- Longer context window — carries full multi-session history
+- Less likely to deflect on edgy topics that are directly relevant here (kink, open relationships, sexual desire)
+
+System prompt contains:
+- Conversation persona (warm, curious, genuinely interested, non-leading)
+- Projective prompt library (rotated, not scripted — AI picks the right moment)
+- Behavioral signal extraction instructions (indirect, never asked as a list)
+- World-view bifurcation questions (embedded naturally into conversation flow)
+- Hard rule: never ask directly "what kind of relationship are you looking for?"
+- Hard rule: never ask about modality preferences directly
+- Instruction to note time-of-day and flag late-night sessions
+
+Model sees: full conversation history for THIS user across all sessions. Nothing from any other user. Ever.
+
+---
+
+## Isolation Principle
+
+This LLM has zero access to:
+- Other users' profiles
+- The matching engine
+- Compacted identities
+- Modality weights
+
+It knows one person. It talks to one person. The conversation is a safe container. Users can say anything — it stays in their lane only.
+
+---
+
+## What the Conversation Captures
+
+### Projective elicitation (reach the unconscious)
+Not asked as a list. Woven into conversation naturally when the moment is right:
+- "Close your eyes for a second. Imagine a big white empty space. What's the first thing that comes to mind?"
+- "Imagine a big empty black space. What do you feel?"
+- "Tell me about a moment you felt completely free."
+- "What are you most afraid of in another person?"
+- "What would your perfect Saturday look like — be specific."
+- Symbols that emerge: scary or joyful? Open or constrictive?
+
+### Indirect behavioral signals
+Never asked directly. Come up in natural conversation:
+- Travel history, passport (open vs closed worldview signal)
+- Communication style (do they texting in exclamation points? ellipses? full sentences?)
+- Living situation, pets, routines
+- Energy level — depleted vs running hot
+- Yes-person or no-person
+
+### World-view bifurcations (the big sorts — asked directly, once, naturally)
+- "Do you think the world is basically a dangerous place or an adventurous one?"
+- "When you meet someone new, is your default to trust them or to wait and see?"
+
+These are the non-negotiable compatibility splits. Asked once. Answer weighted heavily.
+
+### Natural conversation signals
+Everything else the person brings up unprompted. Stories about exes. What they're proud of. What they regret. What they want. What they're afraid they want but won't admit.
+
+The AI doesn't steer. It listens and follows the thread.
+
+---
+
+## The 4AM Principle
+
+System actively encourages off-hours contact. Specifically: late night, early morning, moments of vulnerability.
+
+SMS sent after first call: "You can call anytime — even at 3 or 4 in the morning. That's actually when the most interesting conversations happen."
+
+Late-night sessions tagged in metadata. Analysis pipeline weights those sessions more heavily — masks are down, defenses are off, signal is richer.
+
+---
+
+## Inbound Calls
+
+User can call the system number at any time. Twilio routes inbound to same WebSocket pipeline. System picks up immediately — no hold music, no IVR menu, no "press 1 for..."
+
+System recognizes the phone number, loads conversation history, continues where they left off.
+
+"Hey, you're back. Last time you were telling me about..."
+
+---
+
+## Async Text Input
+
+Between calls, users can submit text:
+- Written reflections
+- Dream journals
+- Things they forgot to say
+- Books / podcasts they want to flag as signal
+
+Simple web form (same URL, phone-verified). Text submitted → stored as session with `source: text` → processed in next analysis run.
+
+System may reference it in next call: "You mentioned you'd been reading [X] — what drew you to that?"
+
+---
+
+## Session Continuity
+
+Every call is a session. Sessions are numbered. History is cumulative.
+
+The AI never makes the user re-explain themselves. It builds. It remembers. The longer someone engages, the more it knows them — and the better the eventual match.
+
+This is a relationship, not a form.

package/basis/layers/L3_RECORDING_STORE.md

@@ -0,0 +1,104 @@
+# Layer 3 — Recording Store
+
+**Goal:** Every interaction captured at full fidelity. Immutable. Cheap to store forever. Re-analyzable as models improve. The recordings are the ground truth.
+
+---
+
+## What Gets Stored
+
+Per session:
+- **Raw audio** — full fidelity `.wav` or `.mp3`
+- **Transcript** — auto-generated, word-level timestamps, speaker-separated
+- **Metadata** — timestamp, duration, source, time-of-day bucket
+
+For text submissions:
+- Raw text content
+- Submission timestamp
+- No audio file
+
+---
+
+## Audio Storage
+
+**Cloudflare R2** (S3-compatible, cheaper egress than AWS S3).
+
+Structure:
+```
+/{user_id}/{session_id}/audio.mp3
+/{user_id}/{session_id}/transcript.json
+/{user_id}/{session_id}/metadata.json
+```
+
+Access: private only. No public URLs. Presigned URLs generated on demand for internal processing. Users never directly access their audio files (they get narrative summaries instead).
+
+Retention: indefinite. This is the user's identity asset.
+
+Cost math: 1 hour of audio ≈ 7MB as MP3. 10,000 users × 5 hours average = 350GB = ~$1.50/month on R2. Cost is irrelevant at any realistic scale.
+
+---
+
+## Transcription
+
+**Deepgram Nova-2** — runs async after each call ends.
+
+Why Deepgram over Whisper (OpenAI):
+- 10–20x faster on long audio
+- Better accuracy on conversational speech (not just dictation)
+- Speaker diarization built in — separates system voice from user voice cleanly
+- Word-level timestamps — useful for future analysis refinement
+- Cheaper at scale: $0.0043/minute (1-hour call = $0.26)
+- Handles accents, informal speech, filler words better than Whisper
+
+Transcript stored as JSON:
+```json
+{
+  "duration": 1847,
+  "words": [
+    {"word": "yeah", "start": 0.24, "end": 0.48, "confidence": 0.99, "speaker": 1},
+    ...
+  ],
+  "paragraphs": [...],
+  "summary": "...",
+  "topics": [...]
+}
+```
+
+Speaker 0 = system AI. Speaker 1 = user. All downstream analysis processes speaker 1 only.
+
+---
+
+## Database Record
+
+PostgreSQL `sessions` table (Supabase):
+
+| Column | Type | Notes |
+|--------|------|-------|
+| `session_id` | UUID | Primary key |
+| `user_id` | UUID | Foreign key → users |
+| `started_at` | timestamptz | Call start |
+| `ended_at` | timestamptz | Call end |
+| `duration_seconds` | integer | |
+| `source` | enum | `voice` / `text` / `whatsapp` |
+| `time_of_day_bucket` | enum | `morning` / `afternoon` / `evening` / `night` / `late_night` |
+| `session_number` | integer | nth session for this user |
+| `audio_r2_key` | text | Path in R2 |
+| `transcript_r2_key` | text | Path in R2 |
+| `transcription_status` | enum | `pending` / `processing` / `complete` / `failed` |
+| `analysis_status` | enum | `pending` / `processing` / `complete` |
+| `late_night_flag` | boolean | True if time_of_day = late_night |
+
+---
+
+## Immutability Principle
+
+Raw store is append-only. Analysis pipeline reads from it, never writes back to it. Raw recordings are never modified or deleted (unless user explicitly requests data deletion under GDPR/CCPA).
+
+Why this matters: analysis models will improve. When we upgrade from Claude claude-sonnet-4-6 to whatever is better in 6 months, we re-run analysis on all existing raw transcripts and get better identity profiles retroactively. The recordings are the permanent asset. The analysis is just the current best interpretation of them.
+
+---
+
+## Processing Trigger
+
+After call ends → Vapi.ai (or Twilio) webhook fires → event pushed to job queue → transcription job queued → on completion, analysis job queued.
+
+User is completely unaware of this pipeline. They hung up. The system does the rest.

package/basis/layers/L4_ANALYSIS_PIPELINE.md

@@ -0,0 +1,257 @@
+# Layer 4 — Analysis Pipeline (Identity Compaction)
+
+**Goal:** Take raw transcripts → produce a compacted identity profile + relationship modality weights. Runs async, invisible to user. Re-runnable as models improve.
+
+---
+
+## Trigger
+
+Every time a session's transcription completes, an analysis job is queued.
+
+**Inngest** for job orchestration — serverless background functions, no Redis or worker infrastructure to operate. Handles retries, delays, fan-out cleanly.
+
+Job: `compact_identity(user_id, session_id)`
+
+Runs in background. User never waits for this. They hung up, it happens.
+
+---
+
+## Analysis LLM
+
+**Claude claude-opus-4-6 (Anthropic)** — most capable model for this step.
+
+This is an offline batch process. Latency doesn't matter. Quality does. Opus over Sonnet here.
+
+---
+
+## Two-Pass Process Per Session
+
+### Pass 1 — Session Signal Extraction
+
+Input: transcript (user speech only — speaker 1 from Deepgram diarization)
+
+Prompt instructs model to extract:
+
+**Domain signals:**
+- Relationships — attachment cues, trust language, conflict references, depth vs breadth indicators
+- Desire — explicit and inferred. Kink signals, vanilla signals, shame vs sovereignty language, body relationship
+- Money — scarcity vs abundance framing, risk language, ambition signals, what they said about work/success/security
+- Health — energy descriptors, physicality references, self-care language, mortality framing
+
+**World-view signals:**
+- Responses to the two bifurcation questions (dangerous/adventurous, good/bad people)
+- Symbolic responses from any projective prompts used in this session
+- Fear vs joy valence of any imagery that emerged
+
+**Behavioral signals:**
+- Travel mentions, passport references
+- Communication style descriptions
+- Living situation, pets, routines mentioned
+- Energy level as described or implied
+
+**Modality signals:**
+- Any explicit statements about relationship type preferences
+- Inferred modality lean from language patterns, stories told, desires expressed
+- Divergence between stated and implied (itself high signal)
+
+**Framework signals (invisible to user):**
+- VALS type indicators — which segment language patterns match
+- Eight Games — which games appear active (Game 1 binary morality language, Game 8 mechanism language, etc.)
+
+Output: `session_signals.json` — structured, per-domain, with direct transcript quotes as evidence for each signal.
+
+---
+
+### Pass 2 — Identity Merge
+
+Input:
+- Existing compacted identity (if any — null for session 1)
+- New `session_signals.json`
+- Session metadata (session number, time-of-day, duration)
+
+Prompt instructs model to:
+- Integrate new signals with existing identity
+- Weight recent sessions slightly higher than older ones
+- Flag contradictions or evolution (person said X in session 1, implies Y in session 4 — that's data)
+- Update modality weights based on accumulated signals
+- Note confidence level per domain (some domains may have sparse signal still)
+
+**Outputs:**
+
+---
+
+## Output A — Base Identity Profile
+
+Stored as structured JSON. Narrative descriptions + underlying signal data.
+
+```json
+{
+  "updated_at": "2026-03-03T04:17:00Z",
+  "session_count": 4,
+  "total_minutes": 87,
+
+  "relationships": {
+    "attachment_style": "anxious-secure spectrum, leans secure with time",
+    "trust_pattern": "slow to open, deep once established",
+    "conflict_mode": "avoidant initially, expressive when safe",
+    "depth_vs_breadth": 0.8,
+    "confidence": 0.75,
+    "evidence": ["said 'I need to feel safe before I open up'", "mentioned 2 long relationships, both 4+ years"]
+  },
+
+  "desire": {
+    "expressed_desires": ["emotional intimacy before physical", "eye contact", "slow burn"],
+    "inferred_desires": ["dominance dynamic, soft", "being truly seen"],
+    "kink_vanilla_spectrum": 0.35,
+    "shame_sovereignty_score": 0.6,
+    "confidence": 0.55,
+    "evidence": [...]
+  },
+
+  "money": {
+    "scarcity_abundance_orientation": 0.65,
+    "risk_tolerance": 0.7,
+    "ambition_contentment": 0.6,
+    "financial_mythology_active": ["money = freedom", "security as prerequisite for love"],
+    "confidence": 0.5,
+    "evidence": [...]
+  },
+
+  "health": {
+    "energy_pattern": "high baseline, crashes under stress",
+    "physicality_orientation": 0.7,
+    "self_care_mode": "reactive not proactive",
+    "confidence": 0.4,
+    "evidence": [...]
+  },
+
+  "worldview": {
+    "world_danger_adventure": 0.72,
+    "people_good_bad": 0.65,
+    "vals_type": "Experiencer",
+    "vals_confidence": 0.7,
+    "games_active": ["Game 5 — meta-awareness emerging", "Game 1 residual — binary moral language occasionally"],
+    "symbolic_responses": [
+      {"prompt": "white empty space", "response": "ocean", "valence": "expansive/joyful"}
+    ]
+  }
+}
+```
+
+---
+
+## Output B — Modality Weights
+
+Updated after every session. Weighted distribution across all relationship modalities.
+
+```json
+{
+  "long-term": 0.65,
+  "open-relationship": 0.20,
+  "friends": 0.10,
+  "casual": 0.05,
+  "kink": 0.0,
+  "polyamory": 0.0,
+  "swinging": 0.0
+}
+```
+
+**Dual-track logic:**
+- Explicit statements about modality preference captured directly as signal
+- Inferred modality from language, stories, desires, fears weighted alongside
+- Stated ≠ true. Divergence between the two is its own signal, logged separately
+- Weights evolve every session — identity is never frozen
+
+**Threshold for modality engagement:** weight > 0.1 triggers modality-specific rendering.
+
+---
+
+## Output C — Modality-Specific Profiles
+
+For each modality where weight > 0.1, a third LLM pass generates a narrative profile through that modality's lens.
+
+Purpose: the matching LLM reads these, not the raw base identity. Each lens emphasizes different dimensions.
+
+**Long-term lens** emphasizes: attachment, depth, values alignment, life vision, conflict resolution style
+
+**Open-relationship lens** emphasizes: autonomy signals, boundary clarity, jealousy patterns, communication about needs
+
+**Kink lens** emphasizes: power dynamic preferences, safety language, community orientation, shame vs sovereignty score
+
+**Friends/adventure lens** emphasizes: energy level, activity orientation, spontaneity vs planning, worldview (adventurous score critical here)
+
+**Casual lens** emphasizes: physical energy, low-attachment signals, communication style, time orientation
+
+Output stored as markdown narrative per modality:
+```
+/identity/{user_id}/modality/long-term.md
+/identity/{user_id}/modality/open-relationship.md
+```
+
+---
+
+## Data Sufficiency Tracking
+
+`identities` table tracks:
+
+| Column | Notes |
+|--------|-------|
+| `session_count` | Total sessions completed |
+| `total_minutes` | Total conversation time |
+| `signal_completeness_score` | 0–1, how complete identity feels across all domains + worldview |
+| `domains_confident` | JSON — which of the four domains have confidence > 0.6 |
+| `ready_for_matching` | Boolean — true when completeness threshold met |
+
+**Completeness threshold (v0.1):** 3+ sessions OR 30+ minutes, AND at least the worldview bifurcation questions answered.
+
+Below threshold: matching engine skips this person. System may prompt them to call again via SMS.
+
+"The more you share, the better your matches get. Call whenever you feel like it."
+
+---
+
+## Storage
+
+PostgreSQL `identities` table (Supabase):
+
+| Column | Type | Notes |
+|--------|------|-------|
+| `user_id` | UUID | 1:1 with users |
+| `updated_at` | timestamptz | Last compaction run |
+| `version` | integer | Increments each compaction |
+| `base_profile` | JSONB | Queryable, the full structured identity |
+| `modality_weights` | JSONB | Queryable, the weight distribution |
+| `signal_completeness_score` | float | 0–1 |
+| `ready_for_matching` | boolean | |
+| `embedding_id` | UUID | FK to pgvector store |
+
+Modality profile markdown files stored in R2:
+```
+/identity/{user_id}/modality/{modality}.md
+```
+
+---
+
+## Embedding
+
+After each identity update, re-embed the compacted profile for vector search.
+
+**OpenAI text-embedding-3-large** (3072 dimensions) — best semantic coverage for nuanced psychological content. Captures meaning, not just keywords.
+
+What gets embedded: the full base identity JSON serialized to text + all active modality profiles concatenated.
+
+Per-modality embeddings also generated — separate embedding per active modality profile. Matching engine uses per-modality embeddings for per-modality search.
+
+Stored in **pgvector** (PostgreSQL extension via Supabase) — vector search co-located with identity data. No separate vector DB to operate at v0.1. Handles millions of users before needing to migrate to dedicated infrastructure (Pinecone, Weaviate).
+
+---
+
+## Re-Analysis
+
+When a better model is available (Claude Opus 5, whatever comes next):
+1. Mark all sessions as `analysis_status: stale`
+2. Re-queue analysis jobs for all sessions in chronological order per user
+3. Compacted identities rebuilt from scratch with better model
+4. Better identities → better matches automatically
+
+The raw recordings + transcripts are the permanent asset. Everything else is derived and re-derivable.

package/basis/layers/L5_MATCHING_ENGINE.md

@@ -0,0 +1,164 @@
+# Layer 5 — Matching Engine
+
+**Goal:** Find compatible people globally. Run continuously as profiles update. Surface only high-confidence matches. Never show a match the system isn't confident about.
+
+---
+
+## Isolation Principle
+
+Entirely separate LLM context from the Conversation LLM. The matching engine sees compacted profiles and modality narratives only. Never raw recordings. Never conversation transcripts. Never personally identifying information.
+
+Two people reduced to their psychographic essence. Matched on that.
+
+---
+
+## When It Runs
+
+**Inngest scheduled function — every hour.**
+
+Processes all users whose identity was updated since the last run (new session compacted, new embedding generated).
+
+Not triggered by user action. Runs silently in background forever.
+
+---
+
+## Step 1 — Vector Candidate Retrieval
+
+**pgvector cosine similarity search** across all ready-for-matching users.
+
+For each user being processed:
+- Run per-modality vector search for each modality where their weight > 0.1
+- Each search returns top 50 candidates by embedding similarity in that modality's embedding space
+- Results weighted by the user's modality weight for that modality
+- Merge all modality results, deduplicate, rank by weighted score
+
+Filters applied at query level (cheap, no LLM):
+- `ready_for_matching = true` on both sides
+- `user_id != self`
+- Not already in `matches` table for this pair (in any direction or status)
+- Not in `declined_pairs` table
+
+Result: up to 150 candidates per user (50 per modality × 3 modalities max), deduplicated and ranked.
+
+---
+
+## Step 2 — Hard Filter Pass
+
+Eliminate obvious incompatibilities before touching the LLM. These are binary blocks.
+
+**World-view bifurcation gates (non-negotiable):**
+- `world_danger_adventure` delta between two people > 0.5 → eliminated
+- `people_good_bad` delta > 0.5 → eliminated
+
+Person who sees the world as dangerous and people as bad lives in a different universe than someone who sees it as adventurous and people as good. No amount of profile similarity bridges that. Eliminate before wasting compute.
+
+**Data gate:**
+- Either person's `signal_completeness_score` < 0.5 → deferred, not eliminated. Re-check next run when more data arrives.
+
+**Geographic gate (optional, user-configurable):**
+- Default: global — no geographic filter
+- User can set a radius preference during conversation ("I'm open to long distance" vs "I really want someone in the same city")
+- If set, filter applied here
+
+Result after hard filter: typically 10–30 candidates from the original 150.
+
+---
+
+## Step 3 — LLM Validation Loop
+
+**Claude claude-sonnet-4-6** — needs to be fast enough to process volume. Sonnet not Opus here.
+
+For each candidate pair surviving hard filter, one LLM call.
+
+**Input to LLM:**
+- Person A's modality profile for the relevant modality (the markdown narrative)
+- Person B's modality profile for the relevant modality
+- Their worldview JSON (bifurcation scores, VALS type, Games active)
+- Cross-modality flag + both modality weights if this is a cross-modality candidate
+- Geographic info if relevant
+
+**LLM outputs structured assessment:**
+
+```json
+{
+  "compatible": true,
+  "confidence": 0.84,
+  "primary_modality": "long-term",
+  "cross_modality": false,
+  "cross_modality_bridge": null,
+  "resonances": [
+    "both hold world as adventurous and expansive",
+    "depth orientation strongly aligned — both slow to open, deep once in",
+    "trust-building style compatible — neither rushes intimacy"
+  ],
+  "tensions": [
+    "energy levels may differ — one runs hot, one more measured"
+  ],
+  "tension_fatal": false,
+  "consent_call_framing": "I found someone who shares the way you see the world — as a place to explore, not to fear. They take their time getting close to people too. There's something here worth a conversation.",
+  "data_gaps": ["desire domain sparse for Person B — only 1 session"]
+}
+```
+
+**Confidence thresholds:**
+- `confidence < 0.70` → discard
+- `confidence 0.70–0.84` → hold. Check if more data arrives in next 2 runs, then surface or discard
+- `confidence ≥ 0.85` → queue for consent gate immediately
+
+**Data gap handling:**
+If LLM flags data gaps for either person → downgrade confidence, add to hold queue, system prompts that person to call again.
+
+**Cross-modality matching:**
+Enabled. If Person A is 0.65 long-term and Person B is 0.55 open-relationship but their profiles are highly compatible, the LLM assesses the cross-modality bridge:
+- Is there genuine compatibility across the modality difference?
+- What's the bridge — what do they share that transcends the modality difference?
+- Flag `cross_modality: true` so consent call framing can acknowledge it without revealing the other person's modality explicitly
+
+---
+
+## Step 4 — Match Confirmation & Storage
+
+Match confirmed when:
+- `compatible: true`
+- `confidence ≥ 0.85` (or held at 0.70–0.84 and second run didn't improve)
+- `tension_fatal: false`
+- No data gaps flagged by LLM
+
+**`matches` table:**
+
+| Column | Type | Notes |
+|--------|------|-------|
+| `match_id` | UUID | |
+| `user_a_id` | UUID | |
+| `user_b_id` | UUID | |
+| `primary_modality` | text | |
+| `cross_modality` | boolean | |
+| `confidence_score` | float | |
+| `resonances` | JSONB | Array of resonance strings |
+| `tensions` | JSONB | Array of tension strings |
+| `tension_fatal` | boolean | |
+| `consent_call_framing` | text | Pre-written by LLM for consent call |
+| `status` | enum | `pending_consent` / `a_accepted` / `b_accepted` / `both_accepted` / `declined` / `introduced` |
+| `created_at` | timestamptz | |
+| `a_consented_at` | timestamptz | |
+| `b_consented_at` | timestamptz | |
+
+**Output per user:** 5–10 confirmed matches maximum in the queue at any time. System doesn't pile up 50 pending matches — it throttles. Quality over volume.
+
+---
+
+## Feedback Loop
+
+After introduction call + 24-hour follow-up SMS (one-tap thumbs up/down):
+
+Positive outcome → strengthen the signal patterns that generated this match:
+- Note which resonance signals predicted success
+- Adjust embedding weights over time
+
+Negative outcome → note what didn't work:
+- Which tensions were underweighted?
+- Was the confidence score miscalibrated?
+
+Over time: matching engine learns from outcomes, not just profiles. This is the moat — the system gets better with every introduction.
+
+V0.1: simple logging. V1: feedback fed back into embedding and confidence calibration.