Verify a Claim Across Sources

from gildea_sdk import Gildea

client = Gildea()

# 1. Start with a claim you found interesting
# (Get the unit_id from a signal detail or search result)
source_unit_id = "clm_01JABCDEF987654321"  # replace with a real unit_id

# 2. Find similar verified text units across all sources
similar = client.search(similar_to=source_unit_id, limit=15)

# 3. Group by source to see consensus breadth
sources = {}
for hit in similar["data"]:
    domain = hit["citation"]["registrable_domain"]
    if domain not in sources:
        sources[domain] = []
    sources[domain].append({
        "text": hit["unit"]["text"],
        "score": hit["relevance_score"],
        "signal": hit["citation"]["signal_title"],
    })

print(f"Similar claims found across {len(sources)} distinct sources:\n")
for domain, claims in sorted(sources.items(), key=lambda x: -len(x[1])):
    print(f"  {domain} ({len(claims)} matching units):")
    for c in claims[:2]:
        print(f"    [{c['score']:.3f}] {c['text'][:100]}")
    print()

# Find the claim you're interested in
results = client.search("NVIDIA H200 shipments increased significantly", limit=1)
unit_id = results["data"][0]["unit"]["unit_id"]

# Now find cross-source consensus for that claim
consensus = client.search(similar_to=unit_id, limit=15)

# Find recent consensus — weight toward newer signals
recent_consensus = client.search(
    similar_to=unit_id,
    recency_boost=0.8,
    limit=10,
)

# Only find consensus within Infrastructure-tagged signals
consensus = client.search(similar_to=unit_id, theme="Infrastructure", limit=10)

# Only find consensus in signals mentioning NVIDIA
consensus = client.search(similar_to=unit_id, entity="NVIDIA", limit=10)

import json

# Build the assessment input
original_claim = "NVIDIA H200 shipments increased significantly in Q1 2026"

corroborating_claims = []
for domain, domain_claims in sources.items():
    for c in domain_claims:
        corroborating_claims.append({
            "text": c["text"],
            "source": domain,
            "signal_title": c["signal"],
            "similarity_score": c["score"],
        })

assessment_data = json.dumps({
    "original_claim": original_claim,
    "corroborating_claims": corroborating_claims,
    "distinct_source_count": len(sources),
    "total_matching_units": sum(len(c) for c in sources.values()),
}, indent=2)

SYSTEM_PROMPT = """You are a research analyst producing a consensus assessment for
a specific claim. You will receive the original claim and a set of semantically
similar verified text units from independent expert sources, found via Gildea's
AI market intelligence platform.

Rules:
- Your job is to assess CONSENSUS BREADTH, not truth. You're measuring how many
  independent experts say something similar, not whether it's objectively true.
- Classify consensus as:
  - STRONG CONSENSUS (4+ distinct sources with high similarity)
  - MODERATE CONSENSUS (2-3 distinct sources)
  - WEAK/ISOLATED (0-1 distinct sources)
- For each corroborating source, note whether it AGREES (says essentially the same
  thing), QUALIFIES (adds nuance or caveats), or PARTIALLY DISAGREES (makes a
  related but different claim).
- Flag any contradictions explicitly — if one source says the opposite, that's
  critical information.
- End with a CONFIDENCE recommendation: how much weight should a decision-maker
  put on this claim?
- Keep it under 300 words.

Output format (markdown):

## Consensus Assessment

**Claim:** "<the original claim>"

**Consensus:** <Strong | Moderate | Weak> ([N] sources, [M] matching units)

**Confidence recommendation:** <Act on it | Investigate further | Treat with skepticism>

### Corroborating Sources
<For each source: domain, what they say, and whether they Agree/Qualify/Partially Disagree>

### Contradicting Sources
<Any sources that say the opposite, or "None found">

### Caveats
<1-2 sentences on what's missing: time periods, specificity, potential bias in source mix>

### Bottom Line
<1 sentence: should the reader trust this claim enough to act on it?>
"""

USER_PROMPT = f"""Assess the consensus for this claim:

{assessment_data}
"""

# Pass to your LLM, or print for manual use
print("=== SYSTEM PROMPT ===")
print(SYSTEM_PROMPT)
print("=== USER PROMPT ===")
print(USER_PROMPT)

## Consensus Assessment

**Claim:** "NVIDIA H200 shipments increased significantly in Q1 2026"

**Consensus:** Strong (5 sources, 8 matching units)

**Confidence recommendation:** Act on it

### Corroborating Sources
- **semianalysis.com** — AGREES: "H200 shipments to hyperscalers exceeded internal
  targets by 15-20% in Q1" (similarity: 0.87)
- **theinformation.com** — AGREES: "NVIDIA delivered more H200 units in Q1 2026
  than all of H2 2025 combined" (similarity: 0.82)
- **bloomberg.com** — QUALIFIES: "H200 shipments grew, but concentrated in 3
  hyperscaler customers accounting for 70% of volume" (similarity: 0.78)
- **reuters.com** — AGREES: "NVIDIA's data center revenue growth driven by
  accelerating H200 adoption" (similarity: 0.74)
- **ft.com** — QUALIFIES: "While H200 volumes rose, average selling prices
  declined 8%, partially offsetting revenue gains" (similarity: 0.71)

### Contradicting Sources
None found.

### Caveats
No source provides exact unit counts — "significantly" is directionally supported
but not precisely quantified. Sources are predominantly US/Western-focused; China
shipment data is absent due to export controls. Bloomberg's concentration note
is important context: growth may be less broad-based than the headline implies.

### Bottom Line
This claim has strong multi-source backing. Act on the directional conclusion
(shipments up materially) but qualify any specific quantity claims with the
Bloomberg concentration caveat.