Apple's real P/E ratio is 29.77. Watch Stratum flag and correct the LLM in flight.
stratum-proxy · financial-verifier · live
$ stratum verify --stream
>
0%
trust
How it works
Four steps from claim to confidence.
01
Proxy the LLM
Point your SDK at Stratum's endpoint. We pass the request to OpenAI, Anthropic, or your own model.
02
Parse the stream
As tokens arrive, we extract atomic claims — numbers, dates, names, citations.
03
Verify in flight
Each claim is checked against authoritative sources: SEC EDGAR for financial figures, Semantic Scholar and CrossRef for citations, FDA databases for drug data, court records for legal claims. If there's no authoritative source, we don't guess — we flag it as unverifiable.
04
Correct & score
Wrong facts are rewritten with sources. Every claim gets a confidence score. You ship trust.
Use cases
Built for industries where wrong answers cost real money.
Fintech
Trade copilots that don't fabricate filings.
Verify P/E ratios, earnings, balance-sheet items against SEC EDGAR before they reach an analyst.
EdTech
Tutors that cite real papers.
Cross-check every citation against arXiv and Crossref. Strike out invented authors before students see them.
Healthcare
Drug data that matches the label.
Verify dosages, interactions, and contraindications against FDA labeling and drug interaction databases before AI-generated medical information reaches a clinician or patient.
Legal
Citations that survive a court check.
Cross-check case citations and statutory references against real court records before they reach a brief. Catch fabricated case law before a judge does.
Built for claims with a checkable answer
We catch the LLM being provably wrong.
Every claim an LLM makes falls into one of two buckets: claims with a checkable answer, and claims without one. "Apple's P/E ratio is 56.2" has a checkable answer — you can compute it. "This stock looks risky" does not. Stratum lives entirely in the first bucket. We catch the LLM being provably wrong, not arguably wrong.
What Stratum is not
Clarity matters. Here's what we don't do.
✕
Not A new foundation model.
We sit on top of yours.
✕
Not A prompt-engineering layer.
We verify the output, not the input.
✕
Not An eval framework.
We run live, in production, per request.
✕
Not A vector DB.
We query authoritative sources, not embeddings.
✕
Not A bias detector.
We verify arithmetic and factual records, not opinions, predictions, or editorial framing. Those require human judgment — and we say so.
Drop-in API
Two lines of code. Zero infrastructure.
Change one base URL. Keep your existing prompts, tools, and streaming logic. Stratum is OpenAI-compatible.
Compatible with OpenAI & Anthropic SDKs
Streaming-native (SSE)
Per-claim confidence in response headers
python · openai.py
from openai import OpenAI
client = OpenAI(
base_url="https://api.stratumverify.ai/v1",
api_key="sk-strat-...",
)
stream = client.chat.completions.create(
model="gpt-4o",
messages=[{"role":"user","content":q}],
stream=True,
extra_headers={"x-stratum-verify":"financial"},
)
# every chunk includes .verification {claim, status, confidence, source}
Dashboard
Every claim, logged and auditable.
Filter by verifier, confidence, status, or model. Export for compliance.