Benchmarks
SEC-scored fundamentals, with the caveats kept visible.
On 2026-07-07, the same 42 tickers and 10 latest-FY headline fields were fetched from StockContext, financialdatasets.ai, and Tiingo. Each source had 420 cells scored against SEC's own companyconcept API.
The headline is not "everyone else is wrong." The scored run found 0 served-wrong cells across all three sources. The difference is what happens when SEC cannot verify a field.
Methodology
Same cells, same SEC oracle.
The harness matched each source's latest annual value to the SEC period it reported, then classified the result as correct, wrong, refused, null, divergent, or unverifiable. The provenance model is described in Core concepts.
42
Tickers
Cross-shape sample
10
Fields
Latest-FY headline cells
420
Cells per source
Scored against SEC companyconcept
Results
Exact-match rates are reported only on the verifiable subset.
The verifiable subset excludes served cells where SEC did not expose a comparable tagged value. Definitional divergences are tracked separately, not counted as served-wrong.
| Source | Exact on verifiable served cells | Served wrong | Note |
|---|---|---|---|
| StockContext | 97.2%317 exact of 326 verifiable served cells | 0 | typed refusals when a value could not be proven |
| financialdatasets.ai | 89.5%299 exact of 334 verifiable served cells | 0 | higher raw fill, more unverifiable served cells |
| Tiingo | 82.9%291 exact of 351 verifiable served cells | 0 | filled every scored cell in this run |
The JSON ledger contains 94 cases where StockContext refused while at least one vendor served a value. Of those, 86 were SEC-untagged and unverifiable; 8 were divergent-definition cases. StockContext also kept native ADR currency for TSM TWD, SAP EUR, BABA CNY, and NVS USD in this run.
Honesty gap
When SEC cannot verify the field, the refusal is the product.
These are examples from the scored ledger. StockContext refused the field; the vendors served values that SEC companyconcept could not verify for the same benchmark cell.
GS operating_income
- StockContext
- not_reported
- financialdatasets.ai
- -17.434B
- Tiingo
- +21.852B
MS operating_income
- StockContext
- not_reported
- financialdatasets.ai
- 1.612B
- Tiingo
- 22.303B
ARCC revenue
- StockContext
- not_applicable_for_shape
- financialdatasets.ai
- -34M
- Tiingo
- +3.052B
Literal-zero exhibit
One field, three surfaces.
For Costco research and development in the FY2025 annual period ending 2025-08-31, the captured raw payloads show three materially different contracts.
{ "dataCode": "rnd", "value": 0.0 }{ "research_and_development": null }{
"research_and_development": {
"available": false,
"reason": "not_reported",
"recovery": { "action": "use_full_document" }
}
}Limits
What this benchmark does not prove.
It is one dated fundamentals sample, not a promise about every ticker, every period, every endpoint, latency, prices, filings, insider data, ownership, or future vendor behavior.
The run was dated 2026-07-07 and covered 42 tickers by 10 fields. "Exact on the verifiable subset" means served cells with a comparable SEC companyconcept value; SEC-untagged cells are disclosed separately because SEC cannot prove them right or wrong.
A separate 2026-07-03 production dogfood run used 4 independent LLM agents and found zero wrong values, but that was a live agent-experience audit, not this scored benchmark.
Reproducibility
The harness is the receipt.
The benchmark harness fetches the same cells from all three sources, scores each against SEC companyconcept, and caches raw responses for inspection. Harness available on request.