Verification API · EvidenceBound
Early AccessIntercept wrong financial numbers before your AI ships them.
EvidenceBound is the verification layer for financial AI products. It checks generated GAAP numbers and filing-based claims against SEC filings — per claim, with citations and an audit trail — before your customers see them. Ship the copilot. Don't ship the hallucinated number.
Agents can generate. We make their work verifiable.
API & MCP · GAAP coverage today · non-GAAP on the roadmap
The Problem
Financial copilots hallucinate numbers. Reviewers catch them late, or not at all.
When a financial AI gets a number wrong, the error is invisible at the output layer. It flows into a research note, a client deck, or a downstream model before a human reads it. By the time someone notices, it has your brand on it. EvidenceBound sits between the model and the customer — every numeric claim checked against its filing, or flagged.
Numbers that don't reconcile
Models state confident figures that don't match the filing; analysts transcribe the wrong line item. A single wrong number moves a model or misleads a client.
Citations that don't exist
AI-generated answers reference sources that were never filed or point at the wrong document. An unsourced claim is an unverifiable claim.
Errors that propagate silently
A bad figure is invisible at the output layer. It flows into a memo, a deck, or a downstream model before anyone catches it.
Pipeline
How EvidenceBound works.
Decompose
The answer is parsed into discrete claims — factual assertions and numeric values — each treated as a separate unit of verification.
Ground
Every claim is traced to its primary source: SEC filings and a structured financial-fact corpus built on EDGAR/XBRL. Unsourced assertions are flagged, not assumed.
Check
Reported figures are matched against the filing to a configurable tolerance — exact-to-the-cent for banking — and contradictions between a claim and its source are surfaced.
Authorize & audit
Each claim returns a deterministic verdict — supported, contradicted, or outside scope — with a reason code, the cited evidence, and a replayable, hash-backed audit record. No confidence estimate; the check either authorized the claim or it did not.
API
One endpoint. Per-claim verdicts, citations, and a top-level verdict.
Send a ticker and the generated answer. Get back a verdict per claim — supported, contradicted, or needs_review — each tied to the filing it was checked against. Narrative claims are returned for human review, not silently graded.
POST /api/verify
{
"ticker": "AAPL",
"memo": "AAPL revenue for fiscal year 2025 was $416,161,000,000.00.
AAPL diluted EPS for fiscal year 2025 was $7.46.
Apple has a durable consumer brand advantage."
}200 OK
{
"verdict": "needs_review",
"elapsed_ms": 38,
"claims": [
{
"status": "supported",
"text": "AAPL revenue for fiscal year 2025 was $416,161,000,000.00.",
"reason": "Matched authoritative filing data.",
"checked_against": {
"ticker": "AAPL",
"metric": "Revenue",
"period": "FY 2025",
"value": 416161000000.0,
"source": "AAPL 10-K (2025-09-27)",
"filed_at": "2025-10-31",
"is_primary": true
}
},
{
"status": "needs_review",
"text": "Apple has a durable consumer brand advantage.",
"reason": "Narrative claim — no primary source to ground against."
}
]
}Response trimmed for illustration. The live endpoint returns the full claim set with status, reason code, and the complete checked_against evidence block. Run the demo for the real shape.
Capabilities
Available nowWhat EvidenceBound does today.
Claim decomposition
Break any answer into individually verifiable claims — factual assertions and numeric values — instead of grading a paragraph as one blob.
Primary-source grounding
Every claim is traced to the SEC filing behind it, on a corpus built with the same EDGAR/XBRL lineage we ran in production at OTC Markets. No source, no pass.
Numeric verification
Reported numbers are checked against the filing to a configurable tolerance, from loose rounding down to exact-to-the-cent. Off-by-a-cent figures are caught, not waved through.
Contradiction detection
Where a claim conflicts with its source — or with the reported figure it cites — the discrepancy is flagged with a plain-language explanation.
Scoped verdicts with reason codes
Each claim gets a deterministic status — supported, contradicted, or outside scope — plus a reason code and the cited evidence. A reviewer sees exactly why a verdict was authorized, not a confidence estimate to interpret.
Audit trail
What was checked, against which filing, at what time, with what result — logged and exportable for internal review or model-governance documentation.
Limitations
What EvidenceBound does not do. Stated plainly.
In verification, imprecise claims erode trust. These are the boundaries of the current release. If your use case needs something beyond them, we'd rather tell you that up front than discover it in production.
Narrative claims are not auto-verified
Qualitative assertions (e.g. 'durable brand advantage', 'regulatory headwinds') are isolated for human review, not graded against the filing. EvidenceBound verifies what can be grounded in a primary source; it does not opine on judgment calls.
GAAP only, for now
Coverage is US-listed GAAP financials from 10-K and 10-Q filings. Non-GAAP and adjusted metrics from 8-K earnings exhibits are on the roadmap, not shipped.
Verdicts are scoped, not blanket guarantees
A 'supported' verdict means the claim matched its cited filing under a fixed, inspectable rule. It is scoped to what the evidence can authorize — not a probabilistic bet and not a warranty of the broader claim's truth.
US public filers
The corpus is SEC filings — US-listed public companies across 1,500+ tickers today. Private-company data, international IFRS filings, and fund-level reporting are out of scope for this release.
Who It's For
Built for teams shipping financial AI.
The first release targets the people closest to the hallucination problem: the engineering and product teams building financial copilots, research assistants, and summarization features. They move fast, they feel the pain of manual QA, and they ship the products where a wrong number does the most damage.
CTO / Head of AI
Needs a verifiable layer between the model and the customer, with an audit trail for governance and a benchmark to justify the spend.
VP Product
Wants to ship the copilot this quarter without a manual review bottleneck, and without a hallucinated number reaching a demo or a customer.
AI engineering team
Wants a per-claim API they can call inline or as a post-processing step — not a dashboard a human has to babysit.
Phase 2 — later
Compliance, equity research, and wealth-management teams have the same verification need and higher willingness to pay, but longer sales cycles and heavier review workflows. We'll reach them through the fintech AI vendors who adopt EvidenceBound first — their customers become the reference accounts that open regulated doors.
Measured proof
Measured, not promisedThe numbers we publish — because we measured them.
A verification product lives or dies on its false-negative rate. Rather than promise one, we ran the deterministic checks against an internal corpus-derived regression benchmark and against real earnings claims. These are the results — including where the contract deliberately abstains.
300 real facts · 270 issuers
Primary-source facts drawn from SEC filings across 270 distinct issuers — the grounding corpus the checks authorize against.
3,510 adversarial structured cases
An internal corpus-derived regression benchmark of structured assertions designed to break the verifier: off-by-cent, wrong period, scaled units, metric confusion.
0 incorrect supports
Across all 3,510 adversarial cases, no wrong claim was authorized as supported.
0 incorrect contradictions
No correct claim was rejected. The deterministic rules do not flip on edge cases.
0 cross-support errors
No claim was authorized against evidence for a different metric, period, or entity.
154 real earnings claims
On 154 claims extracted from real earnings releases: 79% in-scope recall and 100% abstention outside the supported contract.
The 3,510-case set is an internal corpus-derived regression benchmark. The live demo lets you reproduce a verdict end-to-end on your own data.
Capabilities
On the roadmapWhere EvidenceBound is going.
These extend the same engine — decompose, ground, check, audit — with new sources and checks. They are in development, not shipped. We'd rather tell you that than pretend otherwise.
Non-GAAP & earnings-release grounding
Extend beyond GAAP financials to adjusted metrics and their reconciliations, sourced from 8-K earnings exhibits.
Derived-metric recompute
Recompute ratios, growth rates, and margins from a filing's own line items and check the stated value against them.
Intra-response contradiction
Detect where one part of an answer contradicts another, not just where a claim conflicts with its source.
Structured reason-code expansion
Extend the deterministic reason codes so each verdict exposes its factual, numeric, and source-quality checks as separate, inspectable signals — still rules, not estimates.
Tamper-evident audit chaining
Chain-anchor the audit hashes so the verification log is provably immutable for regulatory examination.
On-prem & air-gapped deployment
Run entirely inside your infrastructure, including fully air-gapped environments — no data leaves your perimeter.
Trust & Grounding
A verification tool has to be verifiable itself.
Primary-source only
Every verdict cites the filing it came from — not a vendor's re-derived number. The citation is the filing, so the audit stands on its own.
Exportable audit trail
Each verification is logged and exportable, suitable for internal review and model-governance documentation. Tamper-evident chaining is on the roadmap.
Deploys in your environment
Runs as an API/MCP service, with a self-hosted option so client data stays inside your perimeter. Your queries are never used to train any model.
Who's building it
Aman Grover
Co-founder · EvidenceBound
Twenty years inside the institutions this product serves — OTC Markets, Sapient, ION, Open Link, and Triple Point Technology. Built the production MLops and NLP-to-SQL pipelines on EDGAR/XBRL lineage that EvidenceBound's grounding corpus now runs on. The verification rules reflect time on desks where a wrong number moved a model before anyone read it.
LinkedIn →In flight
Four early-access workflows beginning across:
- · Corporate data
- · Equity research
- · Banking AI memory
- · Runtime authorization
Benchmark three real outputs from your product.
Send us three real answers your AI generates. We'll run them through EvidenceBound and return the per-claim verdicts, the citations, and the audit trail. That's faster than a sales call and more honest than a slide deck.
Benchmark three real outputsAPI & MCP · GAAP coverage today · non-GAAP on the roadmap