@opendirectory.dev/skills 0.1.41 → 0.1.42

package/skills/vc-finder/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: vc-finder
-description: Takes a startup product URL or description, detects the industry and funding stage, identifies 5 comparable funded companies, searches who invested in those companies (Track A), finds VCs who publish investment theses about this space (Track B), and returns a ranked sourced list of relevant investors with deep-dives and outreach hooks. Use when asked to find investors for a startup, identify which VCs fund products like mine, research who backs companies in my space, build a VC target list, or find investor-market fit. Trigger when a user says "find VCs for my startup", "who invests in my space", "build me a VC list", "which funds should I pitch", "find investors for my product", "who backed companies like mine", or "help me find venture capital".
+description: 'Takes a startup product URL or description, detects the industry and funding stage, identifies 5 comparable funded companies, searches who invested in those companies (Track A), finds VCs who publish investment theses about this space (Track B), and returns a ranked sourced list of relevant investors with deep-dives and outreach hooks. Use when asked to find investors for a startup, identify which VCs fund products like mine, research who backs companies in my space, build a VC target list, or find investor-market fit.'
 compatibility: [claude-code, gemini-cli, github-copilot]
 ---
 
@@ -10,7 +10,11 @@ Take a product URL or description. Detect industry and stage. Find 5 comparable
 
 ---
 
-**Critical rule:** Every VC in Track A must include the specific comparable company they backed as evidence. Every VC in Track B must include the exact article or post title where they stated their thesis. If a VC name did not appear in Tavily search results, do not include them. No hallucinated fund names.
+**Zero-hallucination policy:** Every fact in the output must be traceable to a specific Tavily search result or the fetched product page. This applies to:
+- Comparable company names: must appear in Tavily search results, not AI training knowledge
+- VC fund names: must appear verbatim in Tavily search results
+- Check sizes, stage focus, portfolio companies: must come from search snippets, not AI knowledge
+- Fund overviews and thesis summaries: extracted from search snippets only. If a detail is not in the search data, write "not found in search data" -- do not fill from training knowledge.
 
 ---
 
@@ -19,25 +23,24 @@ Take a product URL or description. Detect industry and stage. Find 5 comparable
 | The agent will want to... | Why that's wrong |
 |---|---|
 | Add a16z or Sequoia because they are famous | A famous VC without evidence is noise. Only include VCs that appear in Tavily search results for this specific product. Name-dropping wastes the founder's time. |
-| Continue when all 5 Track A searches return 0 results | Zero Track A results means the comparables were wrong or too obscure. Stop, regenerate comparables with broader known names, and retry. Continuing produces an evidence-free list. |
+| Generate comparable companies from training knowledge | Comparables must come from Tavily search results (Step 6). AI knowledge of companies is not evidence -- a company suggested from memory may have wrong funding status or may not be a true comparable. |
+| Continue when all 5 Track A searches return 0 results | Zero Track A results means the comparables were wrong or too obscure. Stop, re-run Step 6 with broader search queries, and retry. |
 | Include a Track B VC without citing the article or post | Thesis without a source is indistinguishable from hallucination. The founder cannot verify it and the list loses all credibility. |
-| Detect stage from website aesthetics ("site looks polished") | Stage must come from the specific CTA signals detected in Step 4. Aesthetic guessing sends founders to wrong-stage investors. |
-| Write generic outreach hooks like "highlight your traction" | Every outreach hook must name this specific product's differentiator and a specific VC portfolio signal. Generic hooks are removed by the QA step. |
-| Skip the URL fetch when the user also provides a description | Always fetch the URL. The live page often reveals stage signals (pricing CTAs, customer logos, job openings) that the user's description omits. |
+| Fill in fund overview from training knowledge | Fund overviews must come from Tavily snippet text only. If the snippets don't describe the fund, write "not found in search data". |
+| Detect stage from website aesthetics | Stage must come from the specific CTA signals detected in Step 4. |
+| Write generic outreach hooks | Every outreach hook must name this specific product's differentiator and a specific VC portfolio signal or thesis quote from the search data. |
+| Skip the URL fetch when the user also provides a description | Always fetch the URL. The live page often reveals stage signals that the user's description omits. |
 
 ---
 
 ## Step 1: Setup Check
 
 ```bash
-echo "GEMINI_API_KEY: ${GEMINI_API_KEY:+set}"
-echo "TAVILY_API_KEY: ${TAVILY_API_KEY:+set}"
+echo "TAVILY_API_KEY:    ${TAVILY_API_KEY:+set}"
 echo "FIRECRAWL_API_KEY: ${FIRECRAWL_API_KEY:-not set, Tavily extract will be used as fallback}"
 ```
 
-**If GEMINI_API_KEY is missing:** Stop. Tell the user: "GEMINI_API_KEY is required for product analysis and VC synthesis. Get it at aistudio.google.com. Add it to your .env file."
-
-**If TAVILY_API_KEY is missing:** Stop. Tell the user: "TAVILY_API_KEY is required to research VC investments and theses. There is no fallback for this. Get it at app.tavily.com. Free tier: 1000 credits/month (about 125 full runs). Add it to your .env file."
+**If TAVILY_API_KEY is missing:** Stop. Tell the user: "TAVILY_API_KEY is required to research VC investments and theses. There is no fallback for this. Get it at app.tavily.com -- free tier: 1000 credits/month (about 125 full runs). Add it to your .env file."
 
 **If only FIRECRAWL_API_KEY is missing:** Continue silently. Tavily extract will be used for the URL fetch.
 
@@ -120,7 +123,7 @@ Proceed to Step 5 using the pasted description as `product_content`.
 
 ## Step 4: Detect Stage Signals Locally (No API)
 
-Parse the fetched markdown with regex before any API call. This gives Gemini anchored evidence rather than asking it to guess from aesthetics.
+Parse the fetched markdown with regex before the analysis step.
 
 ```bash
 python3 << 'PYEOF'
@@ -129,25 +132,20 @@ import re, json
 content = open('/tmp/vc-product-raw.md').read().lower()
 stage_signals = []
 
-# Pre-seed signals
 if re.search(r'join\s+(the\s+)?waitlist|sign\s+up\s+for\s+beta|early\s+access|request\s+(an?\s+)?invite|get\s+notified', content):
     stage_signals.append({'signal': 'waitlist or beta CTA', 'stage_hint': 'pre-seed'})
 
-# Seed signals
 if re.search(r'start\s+(your\s+)?free\s+trial|try\s+(it\s+)?for\s+free|request\s+a?\s+demo|book\s+a?\s+demo|schedule\s+a?\s+demo', content):
     stage_signals.append({'signal': 'free trial or demo CTA', 'stage_hint': 'seed'})
 
-# Series A signals
 if re.search(r'contact\s+sales|talk\s+to\s+(our\s+)?sales|see\s+pricing|view\s+pricing|plans\s+and\s+pricing', content):
     stage_signals.append({'signal': 'pricing or sales CTA', 'stage_hint': 'series-a'})
 if re.search(r'case\s+stud(y|ies)|customer\s+stor(y|ies)|trusted\s+by\s+[\d,]+|used\s+by\s+[\d,]+', content):
     stage_signals.append({'signal': 'case studies or customer count', 'stage_hint': 'series-a'})
 
-# Series A/B signals
 if re.search(r'enterprise\s+(plan|pricing|tier)|we.?re\s+hiring|join\s+our\s+team|open\s+positions', content):
     stage_signals.append({'signal': 'enterprise tier or job openings', 'stage_hint': 'series-a-or-b'})
 
-# Funding announcement -- extract directly if present
 funding_match = re.search(
     r'raised\s+\$[\d,.]+\s*[mk]?|series\s+[abc]\s+round|seed\s+round|(\$[\d,.]+\s*[mk]?\s+(?:seed|series\s+[abc]))',
     content
@@ -155,7 +153,6 @@ funding_match = re.search(
 if funding_match:
     stage_signals.append({'signal': f'funding text: {funding_match.group(0).strip()}', 'stage_hint': 'announced'})
 
-# Determine dominant stage
 if not stage_signals:
     dominant = 'unknown'
 elif any(s['stage_hint'] == 'announced' for s in stage_signals):
@@ -181,87 +178,340 @@ PYEOF
 
 ---
 
-## Step 5: Product Analysis with Gemini
+## Step 5: Product Analysis (Taxonomy, Stage, ICP)
+
+Print the product content and stage signals:
+
+```bash
+python3 -c "
+import json
+content = open('/tmp/vc-product-raw.md').read()[:6000]
+signals = json.load(open('/tmp/vc-stage-signals.json'))
+print('=== PRODUCT PAGE (first 6000 chars) ===')
+print(content)
+print()
+print('=== DETECTED STAGE SIGNALS ===')
+print(json.dumps(signals, indent=2))
+"
+```
+
+**AI instructions:** Analyze the product page content above. Generate the taxonomy, ICP, and stage classification only -- do NOT generate comparable companies yet (that is done via live search in Step 6).
+
+Write to `/tmp/vc-product-analysis.json`:
+
+- `product_name`: from the page
+- `one_line_description`: what it does, for whom, core value prop. Under 20 words. No marketing language.
+- `industry_taxonomy`: `l1` (top-level: fintech / healthtech / developer tools / consumer / etc.), `l2` (sector: sales technology / logistics software / etc.), `l3` (specific niche: outbound prospecting / last-mile routing / etc.). Vague labels like "technology" or "software" alone are not acceptable.
+- `icp`: `buyer_persona` (job title), `company_type`, `company_size`
+- `detected_stage`: pre-seed / seed / series-a / series-b / unknown
+- `stage_confidence`: high / medium / low
+- `stage_evidence`: one sentence citing exactly which CTA or text on the page drove this. Write "no clear signals found" if unknown.
+- `geography_bias`: US / Europe / global / unclear
+- `comparable_companies`: leave as empty array `[]` -- will be filled in Step 6
 
 ```bash
 python3 << 'PYEOF'
 import json
 
-product_content = open('/tmp/vc-product-raw.md').read()[:6000]
-stage_signals = json.load(open('/tmp/vc-stage-signals.json'))
+analysis = {
+    # FILL from your analysis above
+    "comparable_companies": []
+}
 
-request = {
-    "system_instruction": {
-        "parts": [{
-            "text": "You are a venture capital analyst. Analyze a product page and return structured JSON only. No commentary. No em dashes. Vague category labels like 'technology' or 'software' alone are not acceptable at L2 or L3 -- be specific. Comparable companies must be real funded companies with public funding records, well-known enough to appear in press coverage."
-        }]
-    },
-    "contents": [{
-        "parts": [{
-            "text": f"""Analyze this product page and return a JSON object with exactly these keys:
-
-1. product_name: string
-2. one_line_description: string -- what it does, for whom, core value prop. Under 20 words. No marketing language.
-3. industry_taxonomy: object with:
-   - l1: top-level (e.g. "software", "fintech", "healthtech", "consumer", "hardware")
-   - l2: sector (e.g. "developer tools", "sales technology", "edtech", "logistics software")
-   - l3: specific niche (e.g. "CI/CD automation", "outbound prospecting", "last-mile routing")
-4. icp: object with:
-   - buyer_persona: job title (e.g. "VP Engineering", "founder", "sales ops manager")
-   - company_type: (e.g. "B2B SaaS", "e-commerce brand", "enterprise IT team")
-   - company_size: (e.g. "5-50 employees", "50-500 employees", "enterprise")
-5. detected_stage: one of: pre-seed, seed, series-a, series-b, unknown
-6. stage_confidence: one of: high, medium, low
-7. stage_evidence: one sentence citing exactly which CTA or text on the page drove this classification. Write "no clear signals found" if unknown.
-8. comparable_companies: array of exactly 5 objects, each with:
-   - name: real company name (must have public VC funding records)
-   - similarity_reason: one sentence why this company is comparable to the product
-   - estimated_stage: their funding stage as of your knowledge cutoff
-9. geography_bias: one of: US, Europe, global, unclear -- infer from page text
-
-Stage signals detected from the page (use as input to your stage classification):
-{json.dumps(stage_signals, indent=2)}
-
-Product page content:
-{product_content}"""
-        }]
-    }],
-    "generationConfig": {
-        "temperature": 0.2,
-        "maxOutputTokens": 3000
-    }
+json.dump(analysis, open('/tmp/vc-product-analysis.json', 'w'), indent=2)
+print('Product analysis written.')
+PYEOF
+```
+
+Verify:
+
+```bash
+python3 -c "
+import json
+a = json.load(open('/tmp/vc-product-analysis.json'))
+print('Product:', a['product_name'])
+print('Industry:', a['industry_taxonomy']['l1'], '>', a['industry_taxonomy']['l2'], '>', a['industry_taxonomy']['l3'])
+print('Stage:', a['detected_stage'], '(' + a['stage_confidence'] + ' confidence)')
+"
+```
+
+---
+
+## Step 5b: Curated Pre-Match Against Verified Fund Dataset
+
+Run the product taxonomy against a curated dataset of 25 verified VC funds (sourced from fund websites). Produces zero-hallucination fund matches and seed comparables for Track A -- no Tavily credits consumed.
+
+Print product analysis for tag mapping:
+
+```bash
+python3 -c "
+import json
+a = json.load(open('/tmp/vc-product-analysis.json'))
+print('Taxonomy:', a['industry_taxonomy']['l1'], '>', a['industry_taxonomy']['l2'], '>', a['industry_taxonomy']['l3'])
+print('Stage:', a['detected_stage'])
+print('Geography:', a['geography_bias'])
+"
+```
+
+**AI instructions:** Map the product taxonomy to the standard tags used in the fund dataset. Available tags:
+`DevTools`, `Infrastructure`, `Open Source`, `B2B SaaS`, `AI`, `Data`, `FinTech`, `HealthTech`, `Enterprise`, `Consumer`, `Marketplaces`, `E-commerce`, `Crypto`, `DeepTech`, `Cybersecurity`, `Generalist`
+
+Pick 2-4 tags that describe this product. Map `detected_stage` to: `Pre-seed`, `Seed`, `Series A`, or `Growth`. Map `geography_bias` to: `US`, `Europe`, `India`, or `Global`.
+
+Write product context:
+
+```bash
+python3 << 'PYEOF'
+import json
+
+# FILL based on taxonomy analysis above
+context = {
+    "extracted_tags": ["TagA", "TagB"],  # 2-4 tags from the list above
+    "stage_hint": "Seed",               # Pre-seed / Seed / Series A / Growth
+    "geography_hint": "US"              # US / Europe / India / Global
 }
+json.dump(context, open('/tmp/vc-product-context.json', 'w'), indent=2)
+print('Product context:', context)
+PYEOF
+```
+
+Run scoring against the embedded curated dataset:
+
+```bash
+python3 << 'PYEOF'
+import json
+
+context = json.load(open('/tmp/vc-product-context.json'))
+
+VC_FUNDS = [
+  {"fund_name":"Y Combinator","thesis":"We provide seed funding for startups. We invest in deeply technical teams building massive companies across all domains.","check_size":"$500k","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","B2B SaaS","DevTools","AI"],"geography_focus":["Global"],"notable_portfolio":["Stripe","Airbnb","GitLab"],"website":"https://www.ycombinator.com"},
+  {"fund_name":"boldstart ventures","thesis":"Day one partner for developer first, crypto, and SaaS founders. We love deeply technical founders solving hard infrastructure problems.","check_size":"$1M - $3M","stage_focus":["Pre-seed","Seed"],"industry_tags":["DevTools","Infrastructure","Crypto"],"geography_focus":["Global","US"],"notable_portfolio":["Snyk","Blockdaemon","Superhuman"],"website":"https://boldstart.vc"},
+  {"fund_name":"Heavybit","thesis":"The leading investor in developer-first startups. We help technical founders launch, gain traction, and build enterprise-ready companies.","check_size":"$1M - $5M","stage_focus":["Seed","Series A"],"industry_tags":["DevTools","Infrastructure","Open Source"],"geography_focus":["Global","US"],"notable_portfolio":["PagerDuty","Sanity","Netlify"],"website":"https://www.heavybit.com"},
+  {"fund_name":"Amplify Partners","thesis":"We invest in technical founders building the next generation of IT infrastructure, developer tools, and data platforms.","check_size":"$2M - $8M","stage_focus":["Seed","Series A"],"industry_tags":["DevTools","Infrastructure","AI","Data"],"geography_focus":["US"],"notable_portfolio":["Datadog","OCTO","dbt Labs"],"website":"https://www.amplifypartners.com"},
+  {"fund_name":"OSS Capital","thesis":"We exclusively back early-stage founders building Commercial Open Source Software (COSS) companies.","check_size":"$500k - $2M","stage_focus":["Pre-seed","Seed","Series A"],"industry_tags":["Open Source","DevTools"],"geography_focus":["Global"],"notable_portfolio":["Cal.com","Appsmith","Hoppscotch"],"website":"https://oss.capital"},
+  {"fund_name":"Sequoia Capital","thesis":"We help the daring build legendary companies, from idea to IPO and beyond. Sequoia is an early-stage and growth-stage investor.","check_size":"$1M - $10M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Enterprise","Consumer","AI"],"geography_focus":["Global"],"notable_portfolio":["Apple","Google","WhatsApp"],"website":"https://www.sequoiacap.com"},
+  {"fund_name":"Andreessen Horowitz (a16z)","thesis":"We invest in software eating the world. We back bold entrepreneurs building the future through technology.","check_size":"$1M - $50M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Crypto","Enterprise","Consumer","AI"],"geography_focus":["Global","US"],"notable_portfolio":["Facebook","Coinbase","Figma"],"website":"https://a16z.com"},
+  {"fund_name":"Point Nine Capital","thesis":"We are a seed-stage venture capital firm focused on B2B SaaS and B2B marketplaces globally.","check_size":"$1M - $3M","stage_focus":["Seed"],"industry_tags":["B2B SaaS","Marketplaces"],"geography_focus":["Europe","Global"],"notable_portfolio":["Zendesk","Typeform","Docplanner"],"website":"https://www.pointnine.com"},
+  {"fund_name":"Cherry Ventures","thesis":"We champion founders in Europe from their earliest days. We are generalist seed investors.","check_size":"$1M - $4M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","Consumer","B2B SaaS"],"geography_focus":["Europe"],"notable_portfolio":["FlixBus","Auto1 Group","Forto"],"website":"https://www.cherry.vc"},
+  {"fund_name":"First Round Capital","thesis":"We are the seed-stage firm that builds the most supportive community for founders.","check_size":"$1M - $4M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","B2B SaaS","Consumer"],"geography_focus":["US"],"notable_portfolio":["Uber","Notion","Roblox"],"website":"https://firstround.com"},
+  {"fund_name":"Bessemer Venture Partners","thesis":"BVP helps entrepreneurs lay strong foundations to build and forge long-standing companies.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Enterprise","Consumer","FinTech"],"geography_focus":["Global"],"notable_portfolio":["LinkedIn","Twilio","Shopify"],"website":"https://www.bvp.com"},
+  {"fund_name":"Index Ventures","thesis":"We back the best and most ambitious entrepreneurs across all stages to build category-defining businesses.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","FinTech","Consumer","B2B SaaS"],"geography_focus":["Europe","US","Global"],"notable_portfolio":["Dropbox","Slack","Figma"],"website":"https://www.indexventures.com"},
+  {"fund_name":"Lightspeed Venture Partners","thesis":"We invest globally in enterprise, consumer, and health founders who are shaping the future.","check_size":"$1M - $25M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Enterprise","Consumer","FinTech"],"geography_focus":["Global"],"notable_portfolio":["Snap","Rippling","MuleSoft"],"website":"https://lsvp.com"},
+  {"fund_name":"Accel","thesis":"We partner with exceptional founders from inception through all phases of private company growth.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","B2B SaaS","Consumer","DevTools"],"geography_focus":["Global"],"notable_portfolio":["Facebook","Atlassian","Spotify"],"website":"https://www.accel.com"},
+  {"fund_name":"Bain Capital Ventures","thesis":"From seed to growth, we back founders building legendary infrastructure, fintech, application, and commerce companies.","check_size":"$1M - $50M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Infrastructure","FinTech","B2B SaaS"],"geography_focus":["US","Global"],"notable_portfolio":["DocuSign","SendGrid","Redis"],"website":"https://www.baincapitalventures.com"},
+  {"fund_name":"Greylock Partners","thesis":"We partner with early-stage founders to build enterprise and consumer software companies that define new categories.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["Enterprise","Consumer","Cybersecurity","AI"],"geography_focus":["US"],"notable_portfolio":["Workday","Palo Alto Networks","LinkedIn"],"website":"https://greylock.com"},
+  {"fund_name":"Unusual Ventures","thesis":"We provide a breakthrough level of support for early-stage founders building enterprise tech.","check_size":"$1M - $5M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Enterprise","DevTools","B2B SaaS"],"geography_focus":["US"],"notable_portfolio":["Arctic Wolf","Harness","Vivun"],"website":"https://www.unusual.vc"},
+  {"fund_name":"Crane Venture Partners","thesis":"We back deep tech and enterprise founders in Europe solving hard problems with data and code.","check_size":"$1M - $4M","stage_focus":["Seed"],"industry_tags":["Enterprise","DeepTech","Data","AI"],"geography_focus":["Europe"],"notable_portfolio":["Onfido","Tessian","Forto"],"website":"https://crane.vc"},
+  {"fund_name":"Founder Collective","thesis":"We are a seed-stage venture capital fund, built by founders, for founders. We back weird, wonderful, and wild startups.","check_size":"$500k - $2M","stage_focus":["Seed"],"industry_tags":["Generalist","Consumer","B2B SaaS"],"geography_focus":["US","Global"],"notable_portfolio":["Uber","Airtable","BuzzFeed"],"website":"https://www.foundercollective.com"},
+  {"fund_name":"Benchmark","thesis":"We are a partnership of equal partners. We back mission-driven founders at the earliest stages and walk beside them for the long haul.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["Generalist","Marketplaces","Enterprise","Consumer"],"geography_focus":["US","Global"],"notable_portfolio":["Uber","Twitter","eBay","Snapchat"],"website":"https://www.benchmark.com"},
+  {"fund_name":"Accel India","thesis":"We partner with exceptional founders from inception through all phases of private company growth in the Indian ecosystem.","check_size":"$1M - $15M","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","B2B SaaS","Consumer","FinTech","E-commerce"],"geography_focus":["India"],"notable_portfolio":["Flipkart","Swiggy","Freshworks"],"website":"https://www.accel.com/india"},
+  {"fund_name":"Blume Ventures","thesis":"We are a seed and pre-seed venture fund that backs startups with both funding and active mentoring.","check_size":"$500k - $3M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","B2B SaaS","Consumer","DeepTech","HealthTech"],"geography_focus":["India"],"notable_portfolio":["Unacademy","Purplle","GreyOrange"],"website":"https://blume.vc"},
+  {"fund_name":"Elevation Capital","thesis":"We partner with visionary founders in India across early stages to help them build category-defining businesses.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["Generalist","Consumer","FinTech","B2B SaaS","HealthTech"],"geography_focus":["India"],"notable_portfolio":["Paytm","Swiggy","Meesho"],"website":"https://elevationcapital.com"},
+  {"fund_name":"Peak XV Partners","thesis":"Formerly Sequoia India & SEA, we partner with founders across early, growth, and public stages to build enduring companies.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Consumer","FinTech","B2B SaaS","DevTools","AI"],"geography_focus":["India","South Asia"],"notable_portfolio":["Zomato","Pine Labs","Cred"],"website":"https://www.peakxv.com"},
+  {"fund_name":"Nexus Venture Partners","thesis":"We are a US-India venture capital firm backing extraordinary founders building product-first companies.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["B2B SaaS","Enterprise","DevTools","Consumer"],"geography_focus":["India","US"],"notable_portfolio":["Postman","Hasura","Zepto"],"website":"https://nexusvp.com"}
+]
+
+STAGE_ORDER = {"Pre-seed": 0, "Seed": 1, "Series A": 2, "Growth": 3}
+
+def score_fund(fund, ctx):
+    score = 0
+    fund_tags = fund.get("industry_tags", [])
+    extracted_tags = ctx.get("extracted_tags", ["Generalist"])
+    tag_points = 0
+    matched_tags = []
+    for tag in extracted_tags:
+        if tag in fund_tags:
+            tag_points += 5 if tag == "Generalist" else 20
+            matched_tags.append(tag)
+    tag_points = min(tag_points, 60)
+    score += tag_points
+    stage_hint = ctx.get("stage_hint")
+    fund_stages = fund.get("stage_focus", [])
+    if not stage_hint:
+        score += 10
+    elif fund_stages:
+        if stage_hint in fund_stages:
+            score += 20
+        elif stage_hint in STAGE_ORDER:
+            hint_idx = STAGE_ORDER[stage_hint]
+            if any(f in STAGE_ORDER and abs(STAGE_ORDER[f] - hint_idx) == 1 for f in fund_stages):
+                score += 10
+    geo_hint = ctx.get("geography_hint")
+    fund_geo = fund.get("geography_focus", ["Global"])
+    if not geo_hint or geo_hint == "Global":
+        score += 10
+    elif fund_geo == ["India"] and geo_hint == "US":
+        pass
+    elif geo_hint in fund_geo:
+        score += 20
+    elif "Global" in fund_geo:
+        score += 15
+    if geo_hint == "US" and "India" in fund_geo and "US" not in fund_geo and "Global" not in fund_geo:
+        score = max(0, score - 30)
+    if fund_tags and extracted_tags and fund_tags[0] not in extracted_tags and tag_points <= 20:
+        score = max(0, score - 15)
+    return score, matched_tags
+
+scored = []
+for fund in VC_FUNDS:
+    score, matched_tags = score_fund(fund, context)
+    tier = "High" if score >= 70 else ("Medium" if score >= 40 else "Low")
+    scored.append({
+        "fund_name": fund["fund_name"],
+        "thesis": fund["thesis"],
+        "check_size": fund["check_size"],
+        "stage_focus": fund["stage_focus"],
+        "industry_tags": fund["industry_tags"],
+        "geography_focus": fund["geography_focus"],
+        "notable_portfolio": fund["notable_portfolio"],
+        "website": fund["website"],
+        "source": "verified (fund website)",
+        "score": score,
+        "confidence": tier,
+        "matched_tags": matched_tags
+    })
+
+scored.sort(key=lambda x: (-x["score"], x["fund_name"]))
+relevant = [m for m in scored if m["confidence"] in ("High", "Medium")]
 
-json.dump(request, open('/tmp/vc-analysis-request.json', 'w'))
+curated_comparables = []
+for m in relevant:
+    for company in m.get("notable_portfolio", []):
+        if company not in curated_comparables:
+            curated_comparables.append(company)
+
+output = {
+    "high_medium_matches": relevant,
+    "curated_comparables": curated_comparables[:6]
+}
+json.dump(output, open('/tmp/vc-curated-matches.json', 'w'), indent=2)
+print(f'Curated matches: {len(relevant)} High/Medium confidence funds')
+for m in relevant[:8]:
+    print(f'  {m["confidence"]:6} ({m["score"]:3}) {m["fund_name"]}')
+print(f'Seed comparables from portfolio: {curated_comparables[:6]}')
 PYEOF
+```
 
-curl -s -X POST \
-  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent?key=$GEMINI_API_KEY" \
-  -H "Content-Type: application/json" \
-  -d @/tmp/vc-analysis-request.json \
-  | python3 -c "
-import sys, json
-d = json.load(sys.stdin)
-text = d['candidates'][0]['content']['parts'][0]['text'].strip()
-if text.startswith('\`\`\`'):
-    text = '\n'.join(text.split('\n')[1:-1])
-analysis = json.loads(text)
-json.dump(analysis, open('/tmp/vc-product-analysis.json', 'w'), indent=2)
-print('Product analysis complete.')
-print('Product:', analysis['product_name'])
-print('Industry:', analysis['industry_taxonomy']['l1'], '>', analysis['industry_taxonomy']['l2'], '>', analysis['industry_taxonomy']['l3'])
-print('Stage:', analysis['detected_stage'], '(' + analysis['stage_confidence'] + ' confidence)')
-print('Comparables:', ', '.join(c['name'] for c in analysis['comparable_companies']))
+---
+
+## Step 6: Discover Comparable Companies via Tavily
+
+Load curated portfolio companies from Step 5b as seed comparables:
+
+```bash
+python3 -c "
+import json
+matches = json.load(open('/tmp/vc-curated-matches.json'))
+curated = matches.get('curated_comparables', [])
+print(f'Curated portfolio comparables ({len(curated)}): {curated}')
+need = max(0, 5 - len(curated))
+print(f'Tavily will supplement with up to {need} more')
 "
 ```
 
-**If Gemini returns empty or JSON parsing fails:** Retry once with `maxOutputTokens` reduced to 2000. If retry also fails: Stop. Tell the user: "Product analysis failed. Please paste a direct description (3-5 sentences: what it does, who it is for, current stage) and run again."
+**Do not use AI training knowledge to generate comparable companies.** Curated portfolio companies (above) are already zero-hallucination comparables from verified fund data. Tavily supplements with L3-niche-specific companies.
+
+```bash
+python3 << 'PYEOF'
+import json, os, urllib.request
+
+analysis = json.load(open('/tmp/vc-product-analysis.json'))
+l2 = analysis['industry_taxonomy']['l2']
+l3 = analysis['industry_taxonomy']['l3']
+tavily_key = os.environ.get('TAVILY_API_KEY', '')
+
+queries = [
+    f'"{l3}" startup raised funding venture capital seed series',
+    f'"{l2}" companies venture backed funded startup'
+]
+
+all_results = []
+for query in queries:
+    payload = json.dumps({
+        "api_key": tavily_key,
+        "query": query,
+        "search_depth": "advanced",
+        "max_results": 8,
+        "include_answer": True
+    }).encode()
+
+    req = urllib.request.Request(
+        'https://api.tavily.com/search',
+        data=payload,
+        headers={'Content-Type': 'application/json'},
+        method='POST'
+    )
+
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            result = json.loads(resp.read())
+            all_results.append({
+                'query': query,
+                'answer': result.get('answer', ''),
+                'results': [
+                    {'title': r.get('title',''), 'url': r.get('url',''), 'content': r.get('content','')[:500]}
+                    for r in result.get('results', [])
+                ]
+            })
+            print(f'Comparable search: {len(result.get("results", []))} results for "{query[:60]}"')
+    except Exception as e:
+        print(f'Comparable search FAILED: {e}')
+        all_results.append({'query': query, 'answer': '', 'results': [], 'error': str(e)})
+
+json.dump(all_results, open('/tmp/vc-comparable-search.json', 'w'), indent=2)
+PYEOF
+```
+
+Print results for AI selection:
+
+```bash
+python3 -c "
+import json
+results = json.load(open('/tmp/vc-comparable-search.json'))
+for r in results:
+    print(f'Query: {r[\"query\"]}')
+    print(f'Answer: {r.get(\"answer\",\"\")[:400]}')
+    for item in r.get('results', []):
+        print(f'  - {item[\"title\"]} | {item[\"url\"]}')
+        print(f'    {item[\"content\"][:200]}')
+    print()
+"
+```
+
+**AI instructions:** Combine the curated portfolio companies from `/tmp/vc-curated-matches.json` with the Tavily search results above. Pick exactly 5 comparable companies. Prioritize curated portfolio companies (already verified -- they are real portfolio companies of matched VC funds). Supplement with Tavily-discovered companies to reach 5 if needed.
+
+For each comparable write:
+- `name`: company name
+- `similarity_reason`: one sentence explaining the fit (for curated: reference the fund that backed them; for Tavily: cite the snippet)
+- `source_url`: portfolio fund website for curated companies, Tavily result URL for discovered ones
+- `estimated_stage`: from curated data or snippet text -- write "not in search data" if unknown
+- `source_type`: `"curated_portfolio"` or `"tavily_discovered"`
+
+Update `/tmp/vc-product-analysis.json` with the `comparable_companies` array:
+
+```bash
+python3 << 'PYEOF'
+import json
+
+analysis = json.load(open('/tmp/vc-product-analysis.json'))
+
+analysis['comparable_companies'] = [
+    # FILL 5 companies -- curated_portfolio first, then tavily_discovered
+    # Each: {"name": str, "similarity_reason": str, "source_url": str, "estimated_stage": str, "source_type": str}
+]
+
+json.dump(analysis, open('/tmp/vc-product-analysis.json', 'w'), indent=2)
+print('Comparables written:', ', '.join(c['name'] for c in analysis['comparable_companies']))
+PYEOF
+```
+
+**If fewer than 3 comparable companies appear in the search results:** Broaden the queries. Run a third search: `"[l1] startup" funding round venture capital`. If still thin, proceed with what is available and flag in `data_quality_flags`.
 
 ---
 
-## Step 6: Track A -- Who Invested in Comparable Companies
+## Step 7: Track A -- Who Invested in Comparable Companies
 
-Run 5 Tavily searches, one per comparable. Save all results to a single file.
+Run 5 Tavily searches, one per comparable.
 
 ```bash
 python3 << 'PYEOF'
@@ -318,15 +568,13 @@ print(f'Track A complete. Comparables with results: {sum(1 for r in all_track_a
 PYEOF
 ```
 
-**If all 5 Track A searches return 0 results:** Tell the user: "No funding data found for the comparable companies. This usually means the comparables are too early-stage or obscure for public press coverage. I will retry with broader comparable names." Then re-run Step 5 with a note to Gemini to choose "well-funded companies with significant press coverage" and retry Step 6.
-
-If the retry also returns 0 results: proceed to Track B only, and flag this in `data_quality_flags`.
+**If all 5 Track A searches return 0 results:** Re-run Step 6 with broader queries. Retry with well-covered companies (those with significant press coverage). If still 0: proceed to Track B only and flag in `data_quality_flags`.
 
 ---
 
-## Step 7: Track B -- VCs With Investment Theses About This Space
+## Step 8: Track B -- VCs With Investment Theses About This Space
 
-Run 3 Tavily searches using the L2 and L3 taxonomy from Step 5.
+Run 3 Tavily searches using L2 and L3 taxonomy from Step 5.
 
 ```bash
 python3 << 'PYEOF'
@@ -339,18 +587,9 @@ stage = analysis['detected_stage']
 tavily_key = os.environ.get('TAVILY_API_KEY', '')
 
 queries = [
-    {
-        'name': 'thesis_l3',
-        'query': f'venture capital investment thesis "{l3}" investing 2023 OR 2024 OR 2025'
-    },
-    {
-        'name': 'thesis_l2',
-        'query': f'VC fund "{l2}" investment thesis portfolio companies'
-    },
-    {
-        'name': 'stage_space',
-        'query': f'{stage} investors "{l3}" startup venture capital fund'
-    }
+    {'name': 'thesis_l3', 'query': f'venture capital investment thesis "{l3}" investing 2023 OR 2024 OR 2025'},
+    {'name': 'thesis_l2', 'query': f'VC fund "{l2}" investment thesis portfolio companies'},
+    {'name': 'stage_space', 'query': f'{stage} investors "{l3}" startup venture capital fund'}
 ]
 
 all_track_b = []
@@ -383,33 +622,29 @@ for q in queries:
             print(f"Track B - {q['name']}: {len(result.get('results', []))} results")
     except Exception as e:
         print(f"Track B - {q['name']}: FAILED ({e})")
-        all_track_b.append({
-            'query_name': q['name'],
-            'query': q['query'],
-            'answer': '',
-            'results': [],
-            'error': str(e)
-        })
+        all_track_b.append({'query_name': q['name'], 'query': q['query'], 'answer': '', 'results': [], 'error': str(e)})
 
 json.dump(all_track_b, open('/tmp/vc-trackb-results.json', 'w'), indent=2)
 PYEOF
 ```
 
-**If all 3 Track B searches return 0 results:** Proceed with Track A results only. Note in `data_quality_flags`: "No thesis-led investors found via public search. Try checking Substack manually for VC newsletters covering this niche."
+**If all 3 Track B searches return 0 results:** Proceed with Track A results only. Note in `data_quality_flags`: "No thesis-led investors found via public search."
 
 ---
 
-## Step 8: Gemini Synthesis -- Rank and Score All VCs
+## Step 9: Synthesize -- Rank and Score All VCs
+
+Print the research data:
 
 ```bash
-python3 << 'PYEOF'
+python3 -c "
 import json
 
 analysis = json.load(open('/tmp/vc-product-analysis.json'))
 track_a = json.load(open('/tmp/vc-tracka-results.json'))
 track_b = json.load(open('/tmp/vc-trackb-results.json'))
+curated = json.load(open('/tmp/vc-curated-matches.json'))
 
-# Compress results to stay within token limits
 track_a_summary = []
 for item in track_a:
     snippets = [{'title': r.get('title',''), 'url': r.get('url',''), 'content': r.get('content','')[:400]}
@@ -431,7 +666,22 @@ for item in track_b:
         'top_results': snippets
     })
 
-context = {
+curated_summary = []
+for m in curated.get('high_medium_matches', []):
+    curated_summary.append({
+        'fund_name': m['fund_name'],
+        'confidence': m['confidence'],
+        'score': m['score'],
+        'matched_tags': m['matched_tags'],
+        'thesis': m['thesis'],
+        'check_size': m['check_size'],
+        'stage_focus': m['stage_focus'],
+        'notable_portfolio': m['notable_portfolio'],
+        'website': m['website'],
+        'source': 'verified (fund website)'
+    })
+
+print(json.dumps({
     'product': {
         'name': analysis['product_name'],
         'description': analysis['one_line_description'],
@@ -441,83 +691,58 @@ context = {
         'stage_confidence': analysis['stage_confidence'],
         'geography': analysis['geography_bias']
     },
+    'curated_matches': curated_summary,
     'track_a_research': track_a_summary,
     'track_b_research': track_b_summary
-}
-
-request = {
-    "system_instruction": {
-        "parts": [{
-            "text": """You are a venture capital research analyst. Synthesize investor research into a sourced, ranked list. Follow these rules exactly:
-1. Only include VCs whose names appear in the provided Tavily search results. Do not add VCs not mentioned in the data.
-2. Every Track A VC must have evidence_company: the specific comparable company they backed (required -- omit the VC if you cannot confirm this).
-3. Every Track B VC must have thesis_source_title: the exact article or page title where they stated their thesis (required -- omit the VC if you cannot confirm this).
-4. stage_fit_score 1-10: penalize 3 points if the VC's typical stage does not match the product's detected stage.
-5. space_fit_score 1-10: only give 9-10 if the VC backed 2+ companies in this specific L3 niche.
-6. check_size: use ranges from search result data only. If not found, write "not in search data".
-7. approach_method: one of -- cold email, warm intro required, AngelList, application form, Twitter/X DM. Infer from what is publicly known about this fund's intake process.
-8. outreach_hook: must reference this specific product's differentiator and a named VC portfolio signal or thesis quote. Generic hooks like 'highlight your traction' are not acceptable.
-9. No em dashes anywhere in output.
-10. No marketing language."""
-        }]
-    },
-    "contents": [{
-        "parts": [{
-            "text": f"""Synthesize this VC research for the product below. Return a JSON object with exactly these keys:
-
-1. product_summary: object with name, one_line_description, industry_l1, industry_l2, industry_l3, detected_stage, comparable_companies_used (array of names)
-
-2. track_a_vcs: array of VC objects from Track A research. Each object:
-   - fund_name, evidence_company (REQUIRED), evidence_source_url, stage_focus, check_size, thesis_summary (1-2 sentences), stage_fit_score (1-10), space_fit_score (1-10), approach_method
-
-3. track_b_vcs: array of VC objects from Track B research. Each object:
-   - fund_name, thesis_source_title (REQUIRED), thesis_source_url, stage_focus, check_size, thesis_summary (1-2 sentences), stage_fit_score (1-10), space_fit_score (1-10), approach_method
-
-4. top_5_deep_dives: array of exactly 5 objects (the 5 highest combined score VCs across both tracks). Each:
-   - fund_name, track ("A" or "B"), fund_overview (2-3 sentences), why_fit (2-3 sentences specific to this product's L3 niche), portfolio_in_space (array of 1-3 names from search data only), how_to_approach (specific steps, min 30 chars), outreach_hook (2-3 sentences, product-specific)
+}, indent=2))
+"
+```
 
-5. outreach_hooks: array of exactly 3 objects:
-   - hook_type (e.g. "portfolio overlap angle", "thesis language mirror", "comparable exit angle"), hook_text (2-3 sentences a founder would actually send), best_for (which VC type this works for)
+**AI instructions -- zero-hallucination rules:**
+
+Every field in the output must be traceable to the printed data above. Rules:
+
+1. **curated_vcs:** Use the `curated_matches` data directly. These are pre-verified -- no Tavily evidence required. `fund_overview` comes from the `thesis` field in the curated data. `check_size` and `stage_focus` come from the curated data fields. Do NOT fill from training knowledge even for these funds.
+2. **VC names (Track A / B):** Only include a fund if its name appears verbatim in the snippet text or title. No exceptions.
+3. **evidence_company (Track A):** The comparable company they backed -- must be stated in the snippet text, not inferred.
+4. **thesis_source_title (Track B):** The exact title of the article or post as it appears in the search results.
+5. **fund_overview (Track A / B):** Extract from snippet text only. Max 2 sentences. If the snippets do not describe the fund, write "not found in search data".
+6. **thesis_summary:** Close paraphrase of the snippet text. Do not add context from training knowledge.
+7. **check_size (Track A / B):** From snippet data only. Write "not in search data" if not mentioned.
+8. **portfolio_in_space:** Only companies that appear in the search snippets. Write "not found in search data" if none.
+9. **stage_fit_score 1-10:** Penalize 3 points if the VC's stated stage does not match the product's detected stage.
+10. **space_fit_score 1-10:** 9-10 only if the VC backed 2+ companies in the L3 niche per the snippets or curated data.
+11. **approach_method:** one of -- cold email / warm intro required / AngelList / application form / Twitter/X DM. Infer from snippets or fund website.
+12. **outreach_hook:** Must name a specific portfolio signal or thesis quote. Generic hooks like "highlight your traction" are not acceptable.
+13. No em dashes. No marketing language.
+
+Write to `/tmp/vc-final-list.json`:
+
+- `product_summary`: name, one_line_description, industry_l1, industry_l2, industry_l3, detected_stage, comparable_companies_used (names only)
+- `curated_vcs`: fund_name, confidence ("High"/"Medium"), matched_tags, fund_overview (from thesis field), check_size, stage_focus, website, source ("verified (fund website)"), stage_fit_score, space_fit_score
+- `track_a_vcs`: fund_name, evidence_company (REQUIRED), evidence_source_url, stage_focus, check_size, fund_overview, thesis_summary, stage_fit_score, space_fit_score, approach_method
+- `track_b_vcs`: fund_name, thesis_source_title (REQUIRED), thesis_source_url, stage_focus, check_size, fund_overview, thesis_summary, stage_fit_score, space_fit_score, approach_method
+- `top_5_deep_dives`: fund_name, track ("Curated"/"A"/"B"), fund_overview, why_fit, portfolio_in_space, how_to_approach (min 30 chars), outreach_hook
+- `outreach_hooks`: 3 objects -- hook_type, hook_text (2-3 sentences), best_for
+- `data_quality_flags`: gaps, missing fields, low-confidence areas
 
-6. data_quality_flags: array of strings noting any gaps or low-confidence areas
+```bash
+python3 << 'PYEOF'
+import json
 
-Research data:
-{json.dumps(context, indent=2)}"""
-        }]
-    }],
-    "generationConfig": {
-        "temperature": 0.3,
-        "maxOutputTokens": 6000
-    }
+result = {
+    # FILL from synthesis above
+    # Must include: product_summary, curated_vcs, track_a_vcs, track_b_vcs, top_5_deep_dives, outreach_hooks, data_quality_flags
 }
 
-json.dump(request, open('/tmp/vc-synthesis-request.json', 'w'))
-print('Synthesis request prepared.')
-PYEOF
-
-curl -s -X POST \
-  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent?key=$GEMINI_API_KEY" \
-  -H "Content-Type: application/json" \
-  -d @/tmp/vc-synthesis-request.json \
-  | python3 -c "
-import sys, json
-d = json.load(sys.stdin)
-text = d['candidates'][0]['content']['parts'][0]['text'].strip()
-if text.startswith('\`\`\`'):
-    text = '\n'.join(text.split('\n')[1:-1])
-result = json.loads(text)
 json.dump(result, open('/tmp/vc-final-list.json', 'w'), indent=2)
-print(f'Synthesis complete. Track A: {len(result.get(\"track_a_vcs\", []))} VCs. Track B: {len(result.get(\"track_b_vcs\", []))} VCs.')
-"
+print(f'Synthesis written. Curated: {len(result.get("curated_vcs", []))} VCs. Track A: {len(result.get("track_a_vcs", []))} VCs. Track B: {len(result.get("track_b_vcs", []))} VCs.')
+PYEOF
 ```
 
-**If Gemini returns empty or JSON parsing fails:** Retry once with `maxOutputTokens` reduced to 4000. If retry also fails: present whatever partial JSON was returned, mark missing sections `[INCOMPLETE]`, and tell the user: "Synthesis incomplete. The research data may have been too large. Try running again."
-
 ---
 
-## Step 9: Self-QA
-
-Run before presenting. Remove non-evidenced VCs structurally.
+## Step 10: Self-QA
 
 ```bash
 python3 << 'PYEOF'
@@ -540,6 +765,18 @@ removed_b = original_b - len(result['track_b_vcs'])
 if removed_b > 0:
     failures.append(f'Removed {removed_b} Track B VC(s) missing thesis_source_title')
 
+# Remove deep dives for VCs that were stripped from all tracks
+valid_funds = (
+    {v['fund_name'] for v in result.get('curated_vcs', [])} |
+    {v['fund_name'] for v in result.get('track_a_vcs', [])} |
+    {v['fund_name'] for v in result.get('track_b_vcs', [])}
+)
+original_dives = len(result.get('top_5_deep_dives', []))
+result['top_5_deep_dives'] = [d for d in result.get('top_5_deep_dives', []) if d.get('fund_name') in valid_funds]
+removed_dives = original_dives - len(result['top_5_deep_dives'])
+if removed_dives > 0:
+    failures.append(f'Removed {removed_dives} deep dive(s) for funds stripped during QA')
+
 # Check top 5 deep dives
 dives = result.get('top_5_deep_dives', [])
 if len(dives) < 5:
@@ -548,25 +785,31 @@ for dd in dives:
     if not dd.get('how_to_approach') or len(dd.get('how_to_approach', '')) < 30:
         dd['how_to_approach'] = 'Approach method not determinable from search data. Check the fund website directly for application instructions.'
         failures.append(f"Fixed: '{dd.get('fund_name')}' had missing how_to_approach")
+    if not dd.get('fund_overview') or dd.get('fund_overview') == '':
+        dd['fund_overview'] = 'not found in search data'
 
 # Check outreach hooks count
 if len(result.get('outreach_hooks', [])) != 3:
     failures.append(f"Expected 3 outreach hooks, got {len(result.get('outreach_hooks', []))}")
 
 # Check for em dashes
-if ':' in json.dumps(result):
-    result_str = json.dumps(result).replace(':', ':')
-    result = json.loads(result_str)
-    failures.append('Fixed: em dash characters removed from output')
+full_text = json.dumps(result)
+if '—' in full_text:
+    result = json.loads(full_text.replace('—', '-'))
+    failures.append('Fixed: em dash characters replaced with hyphens')
 
 # Check for forbidden words
 forbidden = ['powerful', 'robust', 'seamless', 'innovative', 'game-changing', 'streamline', 'leverage', 'transform']
-full_text = json.dumps(result).lower()
+full_text_lower = json.dumps(result).lower()
 for word in forbidden:
-    if word in full_text:
+    if word in full_text_lower:
         failures.append(f"Warning: forbidden word '{word}' found in output -- review before presenting")
 
-# Ensure data_quality_flags exists
+# Flag any "not found in search data" entries so user knows coverage is incomplete
+not_found_count = json.dumps(result).count('not found in search data')
+if not_found_count > 0:
+    failures.append(f'INFO: {not_found_count} field(s) marked "not found in search data" -- verify directly before outreach')
+
 if 'data_quality_flags' not in result:
     result['data_quality_flags'] = []
 result['data_quality_flags'].extend(failures)
@@ -582,7 +825,7 @@ PYEOF
 
 ---
 
-## Step 10: Save and Present Output
+## Step 11: Save and Present Output
 
 ```bash
 DATE=$(date +%Y-%m-%d)
@@ -603,13 +846,23 @@ Date: [today] | Stage: [detected_stage] ([stage_confidence] confidence) | Geogra
 What it does: [one_line_description]
 Industry: [l1] > [l2] > [l3]
 Buyer: [buyer_persona] at [company_type], [company_size]
-Comparable companies used for research: [comma-separated list]
+Comparable companies used: [comma-separated list, noting source_type for each]
+
+---
+
+### Curated Matches (Verified)
+
+*Funds matched from a verified dataset of 25 VC funds sourced from fund websites. Zero hallucination -- details come directly from the dataset.*
+
+| Fund | Confidence | Stage Focus | Check Size | Matched Tags |
+|---|---|---|---|---|
+[one row per curated VC, sorted by confidence then score]
 
 ---
 
 ### Track A: VCs Who Backed Similar Companies
 
-*These investors have already written a check in this space.*
+*These investors have already written a check in this space. Evidence from live Tavily search.*
 
 | Fund | Backed Comparable | Stage Focus | Check Size | Fit Score | Approach |
 |---|---|---|---|---|---|
@@ -629,15 +882,15 @@ Comparable companies used for research: [comma-separated list]
 
 ### Top 5 Deep Dives
 
-#### [N]. [Fund Name] (Track [A/B])
+#### [N]. [Fund Name] (Track [Curated/A/B])
 
-Overview: [fund_overview]
+Overview: [fund_overview -- from dataset or search data only]
 Why it fits: [why_fit]
-Portfolio in this space: [names, or "Not found in search data"]
+Portfolio in this space: [from dataset or search data, or "not found in search data"]
 How to approach: [how_to_approach]
 Outreach hook: "[outreach_hook]"
 
-[repeat for all 5]
+[repeat for all available deep dives]
 
 ---
 
@@ -657,7 +910,7 @@ Saved to: docs/vc-intel/[PRODUCT_SLUG]-[DATE].md
 Clean up temp files:
 
 ```bash
-rm -f /tmp/vc-product-raw.md /tmp/vc-stage-signals.json /tmp/vc-analysis-request.json \
-      /tmp/vc-product-analysis.json /tmp/vc-tracka-results.json /tmp/vc-trackb-results.json \
-      /tmp/vc-synthesis-request.json /tmp/vc-final-list.json /tmp/vc-qa-result.json
+rm -f /tmp/vc-product-raw.md /tmp/vc-stage-signals.json /tmp/vc-product-analysis.json \
+      /tmp/vc-product-context.json /tmp/vc-curated-matches.json /tmp/vc-comparable-search.json \
+      /tmp/vc-tracka-results.json /tmp/vc-trackb-results.json /tmp/vc-final-list.json
 ```