more tweaks

apps/x/packages/core/src/knowledge/build_graph.ts

@@ -15,6 +15,7 @@ import {
 import { buildKnowledgeIndex, formatIndexForPrompt } from './knowledge_index.js';
 import { limitEventItems } from './limit_event_items.js';
 import { commitAll } from './version_history.js';
+import { getTagDefinitions } from './tag_system.js';
 
 /**
  * Build obsidian-style knowledge graph by running topic extraction
@@ -35,6 +36,54 @@ const SOURCE_FOLDERS = [
 // Voice memos are now created directly in knowledge/Voice Memos/<date>/
 const VOICE_MEMOS_KNOWLEDGE_DIR = path.join(NOTES_OUTPUT_DIR, 'Voice Memos');
 
+/**
+ * Parse YAML frontmatter from a markdown file and check if it has any noise/skip labels.
+ * Returns true if the email should be skipped (has noise filter tags).
+ */
+function hasNoiseLabels(content: string): boolean {
+    if (!content.startsWith('---')) return false;
+
+    const endIdx = content.indexOf('---', 3);
+    if (endIdx === -1) return false;
+
+    const frontmatter = content.slice(3, endIdx);
+
+    // Get all noise tags from the tag system
+    const noiseTags = new Set(
+        getTagDefinitions()
+            .filter(t => t.type === 'noise')
+            .map(t => t.tag)
+    );
+
+    // Extract filter array values from frontmatter
+    // Matches lines like "    - cold-outreach" under the filter: key
+    const filterMatch = frontmatter.match(/filter:\s*\n((?:\s+-\s+.+\n?)*)/);
+    if (filterMatch) {
+        const filterLines = filterMatch[1].match(/^\s+-\s+(.+)$/gm);
+        if (filterLines) {
+            for (const line of filterLines) {
+                const tag = line.replace(/^\s+-\s+/, '').trim().replace(/['"]/g, '');
+                if (noiseTags.has(tag)) {
+                    return true;
+                }
+            }
+        }
+    }
+
+    // Also check for inline filter array like "filter: ['cold-outreach']" or "filter: [cold-outreach]"
+    const inlineMatch = frontmatter.match(/filter:\s*\[([^\]]*)\]/);
+    if (inlineMatch && inlineMatch[1].trim()) {
+        const tags = inlineMatch[1].split(',').map(t => t.trim().replace(/['"]/g, ''));
+        for (const tag of tags) {
+            if (noiseTags.has(tag)) {
+                return true;
+            }
+        }
+    }
+
+    return false;
+}
+
 function extractPathFromToolInput(input: string): string | null {
     try {
         const parsed = JSON.parse(input) as { path?: string };
@@ -366,16 +415,24 @@ export async function buildGraph(sourceDir: string): Promise<void> {
     // Get files that need processing (new or changed)
     let filesToProcess = getFilesToProcess(sourceDir, state);
 
-    // For gmail_sync, only process emails that have been labeled (have YAML frontmatter)
+    // For gmail_sync, only process emails that have been labeled AND don't have noise filter tags
     if (sourceDir.endsWith('gmail_sync')) {
         filesToProcess = filesToProcess.filter(filePath => {
             try {
                 const content = fs.readFileSync(filePath, 'utf-8');
-                return content.startsWith('---');
+                if (!content.startsWith('---')) return false;
+                if (hasNoiseLabels(content)) {
+                    console.log(`[buildGraph] Skipping noise email: ${path.basename(filePath)}`);
+                    // Mark as processed so we don't re-check it
+                    markFileAsProcessed(filePath, state);
+                    return false;
+                }
+                return true;
             } catch {
                 return false;
             }
         });
+        saveState(state);
     }
 
     if (filesToProcess.length === 0) {
@@ -568,16 +625,23 @@ export async function processAllSources(): Promise<void> {
         try {
             let filesToProcess = getFilesToProcess(sourceDir, state);
 
-            // For gmail_sync, only process emails that have been labeled (have YAML frontmatter)
+            // For gmail_sync, only process emails that have been labeled AND don't have noise filter tags
             if (folder === 'gmail_sync') {
                 filesToProcess = filesToProcess.filter(filePath => {
                     try {
                         const content = fs.readFileSync(filePath, 'utf-8');
-                        return content.startsWith('---');
+                        if (!content.startsWith('---')) return false;
+                        if (hasNoiseLabels(content)) {
+                            console.log(`[GraphBuilder] Skipping noise email: ${path.basename(filePath)}`);
+                            markFileAsProcessed(filePath, state);
+                            return false;
+                        }
+                        return true;
                     } catch {
                         return false;
                     }
                 });
+                saveState(state);
             }
 
             if (filesToProcess.length > 0) {

apps/x/packages/core/src/knowledge/labeling_agent.ts

@@ -57,7 +57,7 @@ ${renderTagSystemForEmails()}
 # Instructions
 
 1. For each email file provided in the message, read its content carefully.
-2. Classify the email using the taxonomy above. Think like a **YC startup founder** triaging their inbox — your time is your scarcest resource:
+2. Classify the email using the taxonomy above. Think like a **busy YC startup founder** triaging their inbox. You receive a mix of real business conversations and unsolicited inbound (cold pitches, marketing, newsletters). Your job is to tell them apart accurately — catch the noise without mislabeling real relationships.
    - **Relationship**: Who is this from? An investor, customer, team member, vendor, candidate, etc.?
    - **Topic**: What is this about? Legal, finance, hiring, fundraising, security, infrastructure, etc.?
    - **Email Type**: Is this a warm intro or a followup on an existing conversation?
@@ -97,23 +97,23 @@ Before finalizing labels, ask: **"Would a busy YC founder want a note about this
 - Is a spam digest or Google Groups moderation report
 - Is routine operational correspondence where the transaction is complete and no follow-up remains
 
-# Cold Outreach Detection (Critical for Precision)
+# Cold Outreach vs Real Relationships
 
-Many emails disguise themselves as real relationships. Before assigning \`vendor\`, \`candidate\`, \`partner\`, or \`followup\`, apply these tests:
+**First, check for engagement signals. If ANY are present, it is NOT cold outreach — classify normally:**
+- The inbox owner replied in the thread
+- There is a calendar invite or scheduled meeting between the sender and inbox owner
+- The email references shared documents, Slack conversations, prior calls, or other collaboration
+- The sender was introduced via a warm intro
+- The sender is from a company the inbox owner is actively doing business with (customer, vendor under contract, investor)
+- The thread has messages from both sides
 
-**It's \`cold-outreach\` (noise), NOT a real relationship, if:**
+**Only if NONE of the above are present, check if it's cold outreach:**
-- The sender is pitching their own product or service — design agencies, compliance firms, content/copy writers, dev shops, freelancers, trademark services, company closure/winding-down services, hiring platforms, etc. — even if they reference your company by name, your YC batch, or offer something "free" or "exclusive for YC founders."
+- The sender is pitching a service or product TO the inbox owner (agencies, dev shops, freelancers, SaaS tools, etc.) → \`cold-outreach\`
-- The thread consists entirely of the same sender following up on their own unanswered messages. A real followup requires prior two-way engagement.
+- A one-sided thread where the sender follows up on their own unanswered messages → \`cold-outreach\`
-- A student, job-seeker, freelancer, or founder cold-emails asking for your time, feedback, or offering free work/trials. These are NOT \`candidate\` — they are \`cold-outreach\`.
+- A stranger cold-emailing about jobs, internships, or offering free work → \`cold-outreach\`, not \`candidate\`
-- Someone invites you to an event you didn't sign up for, especially if the email has marketing formatting (tracking links, unsubscribe footers, HTML banners). This is \`promotion\`, not \`event\`.
+- Someone referencing your YC batch or company name to seem personal, but with no prior engagement → \`cold-outreach\`
 
-**It IS a real relationship (not noise) if:**
+**Remember:** A \`prospect\` is someone who wants to BUY from you. Someone pitching their services to you is \`cold-outreach\`, not \`prospect\` or \`partner\`.
-- You (the inbox owner) are a participant in the thread (you sent a reply, or someone on your team did).
-- The sender is from a company you are already paying, or they are providing a service under contract (e.g., your law firm, your accountant, your cloud provider support).
-- The sender was introduced to you by someone you know (warm intro present in the thread).
-- The sender references a specific ongoing engagement with concrete details — e.g., they are your assigned compliance assessor for an audit you initiated, or they are following up after a call you participated in. This is NOT the same as a generic "I noticed your company uses X" pitch.
-
-**Key heuristic:** If every message in the thread is FROM the same external person and the inbox owner never replied, it's almost certainly cold outreach — regardless of how personalized it sounds. Label it \`cold-outreach\`.
 
 # Routine Operations & Finance (Often Missed as Noise)
 
@@ -156,9 +156,13 @@ These are noise even from a vendor you recognize or a platform you use:
 
 If the sender is \`noreply-spamdigest\` (Google Groups spam moderation reports), label it \`filter: ['spam']\`. Google already flagged these as spam. Do not evaluate the held messages inside — the digest itself is noise.
 
-# Filter array must only contain tags from the Noise category
+# Filter and Relationship arrays — correct placement is critical
 
-Do not put topic or relationship tags into the filter array. If an email is an event promotion, use \`promotion\` in filter — not \`event\`.
+- The \`filter\` array must only contain tags from the **Noise** category.
+- The \`relationship\` array must only contain tags from the **Relationship** category.
+- **\`cold-outreach\` is a NOISE tag — it goes in \`filter\`, NEVER in \`relationship\`.** If an email is cold outreach, set \`filter: ['cold-outreach']\`. The relationship array should be empty \`[]\` for cold outreach emails.
+- Do not put topic or relationship tags into the filter array. If an email is an event promotion, use \`promotion\` in filter — not \`event\`.
+- If an email is cold outreach, do NOT also tag it as \`prospect\`, \`candidate\`, \`partner\`, or \`vendor\` in relationship. Cold outreach overrides — the relationship array should be \`[]\`.
 
 # Frontmatter Format
 

apps/x/packages/core/src/knowledge/note_creation.ts

@@ -221,12 +221,14 @@ Emails containing calendar invites (\`.ics\` attachments or inline calendar data
 
 ---
 
-# Step 1: Source Filtering (Label-Based)
+# Step 1: Source Filtering
 
 ## For Meetings and Voice Memos
-Always process — no filtering needed.
+Always process — no filtering needed. Skip to Step 2.
 
-## For Emails — Read YAML Frontmatter
+## For Emails — TWO mandatory gates, BOTH must pass
+
+### Gate 1: Label Check
 
 Emails have YAML frontmatter with labels prepended by the labeling agent:
 
@@ -249,15 +251,51 @@ labeled_at: "2026-02-28T12:00:00Z"
 
 ${renderNoteEffectRules()}
 
-## Filter Decision Output
+**Gate 1 verdict — if ANY filter/noise label is present, you MUST output:**
-
-If skipping:
 \`\`\`
-SKIP
+STOP — GATE 1 FAILED
-Reason: Labels indicate skip-only categories: {list the labels}
+Labels: {list the filter labels}
+Action: Do not create any notes. Do not proceed to Gate 2. End here.
 \`\`\`
+**After outputting this, STOP. Do not write any files. Do not create any notes. Your task is complete.**
 
-If processing, continue to Step 2.
+If no filter labels are present, proceed to Gate 2.
+
+---
+
+### Gate 2: Engagement Test
+
+**This gate applies to ALL emails that passed Gate 1. It is a HARD STOP — not a suggestion.**
+
+The inbox owner is a busy startup founder who receives a high volume of unsolicited inbound. Most emails from strangers are noise. The labeling agent sometimes misclassifies cold outreach.
+
+**Read the email content and answer this single question: "Has the inbox owner ever engaged with this sender?"**
+
+Evidence of engagement (at least one MUST be true to proceed):
+- The inbox owner sent a reply in the thread
+- There is a prior meeting or introduction with this person
+- The person is from a company the owner is already doing business with (paying customer, active vendor under contract, investor)
+
+**If NONE of the above are true, you MUST output:**
+\`\`\`
+STOP — GATE 2 FAILED
+Reason: No evidence of two-way engagement — {brief explanation}
+Action: Do not create any notes. End here.
+\`\`\`
+**After outputting this, STOP. Do not write any files. Do not create any notes. Your task is complete.**
+
+This means NO notes for:
+- Cold outreach senders (even if labeled as prospect/candidate/partner)
+- Newsletter/digest/promo senders
+- Webinar speakers mentioned in marketing emails
+- People from community digests (Bookface, YC digest)
+- Anyone who followed up multiple times with no reply from the owner
+- Students/freelancers/developers cold-emailing about jobs or offering work
+- Event platform senders (Luma, Beehiiv, etc.)
+
+**When in doubt, STOP. A missed note can be created later. A junk note pollutes the knowledge base permanently.**
+
+If engagement evidence exists, proceed to Step 2.
 
 ---
 
@@ -506,18 +544,21 @@ For entities not resolved to existing notes, determine if they warrant new notes
 
 ### Who Gets a Note
 
+**The golden rule: only create notes for people the inbox owner has a real relationship with.** A "real relationship" means two-way engagement — the owner has met, replied to, or is actively doing business with this person. Passing through the label filter is necessary but NOT sufficient.
+
 **CREATE a note for people who are:**
-- External (not @user.domain)
+- Attendees in meetings the owner participated in
-- Attendees in meetings
+- People the owner has replied to or engaged with in email
-- Email correspondents (emails that reach this step already passed label-based filtering)
+- People introduced via warm intros from known contacts
-- Decision makers or contacts at customers, prospects, or partners
+- Contacts at companies the owner is actively doing business with (customers, investors, vendors under contract, partners)
-- Investors or potential investors
+- Candidates the owner is actively interviewing (responded to, scheduled with)
-- Candidates you are interviewing
-- Advisors or mentors
-- Key collaborators
-- Introducers who connect you to valuable contacts
 
 **DO NOT create notes for:**
+- Anyone from a one-time inbound email the owner never responded to
+- People mentioned in newsletters, digests, or community roundups
+- Speakers listed in webinar or event promotion emails
+- Senders from marketing/event platforms (Luma, Beehiiv, etc.)
+- Cold outreach senders, even after multiple follow-ups with no reply
 - Large group meeting attendees you didn't interact with
 - Internal colleagues (@user.domain)
 - Assistants handling only logistics
@@ -559,16 +600,19 @@ If role is not explicitly stated, infer from context:
 |-------------------|----------------------|------------------|
 | Customer (active deal) | Yes — key contacts | Yes |
 | Customer (support ticket) | No | Maybe update existing |
-| Prospect | Yes — decision makers | Yes |
+| Prospect (owner has engaged) | Yes — decision makers | Yes |
+| Prospect (no engagement) | No — this is cold outreach | No |
 | Investor | Yes | Yes |
-| Strategic partner | Yes — key contacts | Yes |
+| Strategic partner (mutual engagement) | Yes — key contacts | Yes |
 | Vendor (strategic) | Yes — main contact only | Yes |
 | Vendor (transactional) | No | Optional |
 | Bank/Financial services | No | Yes (one note) |
-| Candidate | Yes | No |
+| Candidate (owner is interviewing) | Yes | No |
+| Candidate (unsolicited) | No — this is cold outreach | No |
 | Service provider (one-time) | No | No |
-| Personalized outreach | Yes | Yes |
+| Cold outreach / unsolicited inbound | No | No |
-| Generic cold outreach | No | No |
+| Newsletter / digest / promo sender | No | No |
+| Event/webinar invite sender | No | No |
 
 ### Handling Non-Note-Worthy People
 
@@ -581,15 +625,23 @@ For people who don't warrant their own note, add to Organization note's Contacts
 
 ## Organizations
 
+**Only create org notes for organizations the user directly does business with.** The test is: "Does the user have an active, ongoing relationship with this organization?"
+
 **CREATE a note if:**
-- Someone from that org attended a meeting
+- They're a customer, investor, or partner the owner is actively engaged with
-- They're a customer, prospect, investor, or partner
+- The owner is actively working with someone there (meetings, email exchanges, contracts)
-- Someone from that org sent relevant personalized correspondence
+- They are a vendor the owner has a contract with or is actively evaluating
 
 **DO NOT create for:**
+- Organizations mentioned only as background context (a contact's previous employer, university, portfolio company, etc.)
+- A candidate's current or former employer — the person note is enough
+- Organizations from cold outreach emails
+- Organizations mentioned in newsletters, digests, or promotional emails
 - Tool/service providers mentioned in passing
 - One-time transactional vendors
-- Consumer service companies
+- Consumer service companies (banks, airlines, etc.)
+- Event/marketing platforms (Luma, Beehiiv, etc.)
+- Organizations referenced only because a contact works there, unless the user is doing business with that org directly
 
 ## Projects
 
@@ -601,8 +653,13 @@ For people who don't warrant their own note, add to Organization note's Contacts
 ## Topics
 
 **CREATE a note if:**
-- Recurring theme discussed
+- A substantive business topic discussed across multiple conversations with real contacts (e.g., "SOC 2 Compliance", "Series A Fundraise")
-- Will come up again across conversations
+- Has concrete facts, decisions, or action items attached to it
+
+**DO NOT create topic notes for:**
+- Internal system concepts (email filtering, labeling, noise categories, tags)
+- Abstract or meta categories (e.g., "cold outreach", "newsletters", "noise")
+- Topics only mentioned in skipped/noise emails
 
 ---
 

apps/x/packages/core/src/knowledge/tag_system.ts

@@ -32,10 +32,10 @@ const DEFAULT_TAG_DEFINITIONS: TagDefinition[] = [
   // ── Relationship — who is this from/about (all create) ────────────────
   { tag: 'investor', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Investors, VCs, or angels', example: 'Following up on our meeting — we\'d like to move forward with the Series A term sheet.' },
   { tag: 'customer', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Paying customers', example: 'We\'re seeing great results with Rowboat. Can we discuss expanding to more teams?' },
-  { tag: 'prospect', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Potential customers', example: 'Thanks for the demo yesterday. We\'re interested in starting a pilot.' },
+  { tag: 'prospect', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Potential customers — people who want to BUY from you. NOT someone pitching their services to you — that is cold-outreach.', example: 'Thanks for the demo yesterday. We\'re interested in starting a pilot.' },
   { tag: 'partner', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Business partners, corp dev, or strategic contacts', example: 'Let\'s discuss how we can promote the integration to both our user bases.' },
   { tag: 'vendor', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Service providers you already pay or have a contract with (legal, accounting, infra). NOT someone pitching their services to you — that is cold-outreach.', example: 'Here are the updated employment agreements you requested.' },
-  { tag: 'candidate', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Job applicants responding to a specific open role, or recruiters pitching candidates for your roles. NOT unsolicited students or strangers asking for your time — that is cold-outreach.', example: 'Thanks for reaching out. I\'d love to learn more about the engineering role.' },
+  { tag: 'candidate', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Job applicants responding to a specific open role, or recruiters pitching candidates for your roles. NOT unsolicited students, strangers, or anyone cold-emailing about jobs, internships, or offering free work — that is cold-outreach.', example: 'Thanks for reaching out. I\'d love to learn more about the engineering role.' },
   { tag: 'team', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Internal team members and co-founders', example: 'Here\'s the updated roadmap for Q2. Let\'s discuss in our sync.' },
   { tag: 'advisor', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Advisors, mentors, or board members', example: 'I\'ve reviewed the deck. Here are my thoughts on the GTM strategy.' },
   { tag: 'personal', type: 'relationship', applicability: 'both', noteEffect: 'create', description: 'Family or friends', example: 'Are you coming to Thanksgiving this year? Let me know your travel dates.' },