Validate a Thesis

import numpy as np
from gildea_sdk import Gildea

client = Gildea()

# 1. Your thesis
thesis = "Anthropic's moat is research depth, not distribution."

# 2. Embed it
thesis_vec = np.array(client.embed(thesis)["embedding"])

# 3. Search for related verified claims
results = client.search(thesis, limit=10)

# 4. Score each result against your thesis
print(f"Thesis: {thesis}\n")
for hit in results["data"]:
    claim_text = hit["unit"]["text"]
    # If the search result has an embedding, use it for similarity
    # Otherwise use the embed endpoint
    claim_vec = np.array(client.embed(claim_text)["embedding"])
    sim = float(np.dot(thesis_vec, claim_vec))
    
    label = "SUPPORTS" if sim > 0.75 else "RELATED" if sim > 0.6 else "TANGENTIAL"
    print(f"[{sim:.3f}] [{label}] {claim_text}")
    print(f"  Source: {hit['citation']['signal_title']} ({hit['citation']['registrable_domain']})")
    print()

# Only check against reasoning about a specific company
results = client.search(thesis, entity="org:/anthropic", limit=10)

# Only check against a specific market dynamic
results = client.search(thesis, theme="Competitive Dynamics", limit=10)

# Search broadly, then let an LLM classify support vs. contradiction
related_claims = []
results = client.search("Anthropic competitive moat defensibility", limit=15)

for hit in results["data"]:
    claim_vec = np.array(client.embed(hit["unit"]["text"])["embedding"])
    sim = float(np.dot(thesis_vec, claim_vec))
    if sim > 0.5:
        related_claims.append({
            "text": hit["unit"]["text"],
            "source": hit["citation"]["signal_title"],
            "domain": hit["citation"]["registrable_domain"],
            "similarity": sim,
        })

import json

# Build the LLM prompt
evidence_block = json.dumps(related_claims, indent=2)

SYSTEM_PROMPT = """You are a senior analyst producing a thesis validation brief.
You will receive a hypothesis and a set of verified expert claims with source
attribution. Your job is to assess whether the evidence supports, contradicts,
or is insufficient to evaluate the thesis.

Rules:
- Only cite claims actually provided. Never fabricate sources.
- Distinguish between SUPPORTING evidence (directionally agrees with the thesis),
  CONTRADICTING evidence (directionally disagrees), and CONTEXTUAL evidence
  (relevant but neither supports nor contradicts).
- Be direct about the strength of evidence. "3 sources agree" is stronger than
  "1 source agrees." Say so.
- End with a clear VERDICT: Supported, Challenged, Mixed, or Insufficient Evidence.
- Keep the brief under 400 words.

Output format (markdown):

## Thesis Validation Brief

**Thesis:** <the hypothesis being tested>

**Verdict:** <Supported | Challenged | Mixed | Insufficient Evidence>

**Confidence:** <High | Medium | Low> (based on volume and diversity of evidence)

### Supporting Evidence
<bulleted list of claims with source attribution, or "None found">

### Contradicting Evidence
<bulleted list of claims with source attribution, or "None found">

### Contextual Evidence
<bulleted list of relevant-but-neutral claims, or "None found">

### Assessment
<2-3 sentence synthesis: what does this mean for the decision at hand?>

### Recommended Next Steps
<1-2 specific actions based on the evidence gaps>
"""

USER_PROMPT = f"""Thesis to validate:
"{thesis}"

Verified expert claims (from Gildea intelligence database):
{evidence_block}
"""

# Pass SYSTEM_PROMPT and USER_PROMPT to your LLM of choice.
# Example with Anthropic SDK:
#
# import anthropic
# llm = anthropic.Anthropic()
# response = llm.messages.create(
#     model="claude-sonnet-4-20250514",
#     max_tokens=1024,
#     system=SYSTEM_PROMPT,
#     messages=[{"role": "user", "content": USER_PROMPT}],
# )
# brief = response.content[0].text

# Or print the prompts to use in any LLM interface:
print("=== SYSTEM PROMPT ===")
print(SYSTEM_PROMPT)
print("=== USER PROMPT ===")
print(USER_PROMPT)

## Thesis Validation Brief

**Thesis:** Anthropic's moat is research depth, not distribution.

**Verdict:** Mixed

**Confidence:** Medium (6 relevant claims from 4 distinct sources)

### Supporting Evidence
- "Anthropic's constitutional AI research represents a differentiated technical
  approach that competitors have not replicated" (The Information, theinformation.com)
- "Research-first culture attracts top ML talent, creating a self-reinforcing
  advantage" (Stratechery, stratechery.com)

### Contradicting Evidence
- "Amazon's $4B investment gives Anthropic distribution through AWS Bedrock that
  rivals OpenAI's Microsoft channel" (Bloomberg, bloomberg.com)
- "Claude's API adoption in enterprise is growing faster than its research
  publications" (Semianalysis, semianalysis.com)

### Assessment
The evidence suggests research depth IS a real advantage, but the thesis
understates Anthropic's emerging distribution moat through Amazon/AWS.
The binary framing ("research, not distribution") is too narrow — both
are building simultaneously.

### Recommended Next Steps
- Pull entity co-occurrence between Anthropic and Amazon to quantify how
  tightly the distribution narrative is growing
- Search for "Anthropic enterprise adoption" claims to size the distribution
  counter-evidence