Playground
Structured Schema Validator for Finance
Structured schema validator finance — paste LLM JSON, see missing fields, type mismatches, enum errors, and plausibility flags. Client-side.
- Inputs
- Paste + configure
- Runtime
- 1–15 s
- Privacy
- Client-side · no upload
- API key
- Not required
- Methodology
- Open →
1 · Pick a schema
LLM research note on an investable idea. Must declare a probability, a thesis, and explicit invalidation conditions — every research output that cannot be falsified is just narrative.
Schema definition
| Field | Type / bounds | Req |
|---|---|---|
| probability | number [0 … 1] Subjective probability the thesis plays out. 0..1 inclusive. | required |
| thesis | string ≥50 chars Plain-English thesis. Minimum 50 chars — forces substance over one-liners. | required |
| conviction | enum (low|medium|high) Qualitative conviction bucket. | required |
| invalidation_conditions | array ≥1 items Explicit events that would kill the thesis. ≥1 item. | required |
| source_citations | array ≥1 items URLs or document IDs backing the claims. ≥1 item. | required |
Valid example
{
"probability": 0.62,
"thesis": "Issuer SYNTHETIC_A's cloud-revenue mix is re-rating vs peers after three consecutive quarters of operating-margin expansion, while the sell-side model still assumes the legacy mix.",
"conviction": "medium",
"invalidation_conditions": [
"Next quarter cloud-revenue growth prints below 20% YoY",
"Operating margin contracts by more than 150 bps"
],
"source_citations": [
"10-Q FY2026 Q1, p.14",
"Earnings call transcript 2026-02-11"
]
}2 · Paste your LLM JSON
Validation runs in your browser as you type. Nothing is uploaded.
Schema validity
100%
5 / 5 fields passed · overall pass
Schema: Research Output
3 · Validation results
| Field | Message | |
|---|---|---|
| ✓ | probability | ok |
| ✓ | thesis | ok |
| ✓ | conviction | ok |
| ✓ | invalidation_conditions | ok |
| ✓ | source_citations | ok |
What gets checked
- Field presence, JSON type, enum values, numeric bounds, minimum string / array sizes.
- Cross-field sanity (long stops below take-profit, leverage above retail limits, duplicate peer tickers).
- Soft plausibility bands drawn from real-world issuer filings — out-of-band values warn, not fail.
- Missing unit / GAAP-basis heuristics for revenue / EPS / EBITDA field names.
See methodology for why each schema is shaped the way it is.
How to use
Step-by-step
- 1
Pick a reference schema (trade order, risk report, transaction extraction, KPI extraction, earnings summary) or upload your own.
- 2
Paste a sample LLM output to validate.
- 3
Read pass/fail with per-field error details: missing required fields, type mismatches, enum violations, cross-field consistency failures.
- 4
For batch validation, upload a directory of outputs. Read aggregate failure rates and top-3 most common errors.
- 5
Use the failure-mode aggregate to debug your prompt or schema — repeated failures often point to ambiguous prompt language.
For agents
Use in an agent
Same math, same result shape as the UI above — as a static ES module. No HTTP request, no auth, no rate limit.
import { compute } from "https://aifinhub.io/engines/structured-schema-validator-finance.js";Contract: /contracts/structured-schema-validator-finance.json Full agent guide →
Glossary references
Terms used by this tool
Questions people ask next
FAQ
What schemas does the validator check?
Common finance-domain LLM output schemas: trade orders, risk reports, transaction extraction, KPI extraction, and earnings summary. The validator includes 12 reference schemas pre-built from common production patterns. You can also upload your own.
Why use this instead of plain JSON Schema validation?
Two things plain JSON Schema misses: (1) finance-specific constraints (e.g., 'side' must be 'buy' or 'sell', 'symbol' must match the regex for valid ticker symbols, 'amount' must be positive), (2) cross-field consistency (e.g., 'limit price' should only appear when 'order type' is 'limit'). The validator wraps both.
What does it report?
Pass/fail per output, plus per-failure details: which field, which constraint failed, and a suggested fix. For a batch of LLM outputs, the tool aggregates failure modes — top-3 most common errors, useful for debugging your prompt or schema.
Can it validate streaming outputs?
Partially — it can validate at end-of-stream once the JSON is complete. Mid-stream validation requires knowing whether the partial output is supposed to be an array or object; the tool exposes a 'streaming mode' that does best-effort partial validation but flags incomplete documents.
How strict should I make the schema?
Strict enough to catch real errors, loose enough to not reject legitimate variation. Common pattern: tight validation on machine-consumed fields (numbers, enums, IDs); looser validation on human-readable fields (descriptions, summaries). The methodology page documents this principle.
Related deep dive
All articles →Read further
Long-form context behind the tool output.
- Pillar · Guide·14 min
Reading Financial Filings With LLMs: 2026 Playbook
A map of eight filing tasks — extraction, summarization, peer comparison, Q&A, classification, sentiment, forecasting input, compliance — with model.
Read - Tutorial · Runnable·11 min
Numeric Precision in LLM Filing Extraction
Six precision traps — units, currency, GAAP vs non-GAAP, diluted vs basic shares, restatements, rounding — and the structured-output pattern that fixes them.
Read - Tutorial · Runnable·12 min
Prompt Patterns for Earnings Calls
Five copy-paste patterns — speaker attribution, hedged-guidance confidence, multi-quarter delta, risk aggregator, forward-outlook separator.
Read
Complementary tools
Users of this tool often explore
Hallucination Detector
Paste a source document + an LLM's extraction. Every numeric claim in the output is checked against the source. Client-side. Catches silent fabrication.
Agent Skill Tester for Markets
Paste a SKILL.md definition + sample input + your Anthropic API key. See structured extraction, token cost, and latency — all in your browser. No signup.
Prompt Regression Tester
Run the same prompt against multiple models (Claude 4.5/4.6/4.7, GPT-5, Gemini 2.5) with your own keys. Diff outputs, score drift, catch regressions.