Changelog

An independent audit found that for banks and thrifts we were publishing fee income as if it were total revenue, and labelling it clean. If you pulled revenue for a bank before this date, re-pull it.

• Banks frequently tag no total-revenue element at all. The only candidate left in our list was the ASC-606 contract-revenue tag — which for a bank is just fees. Fifth Third's FY2023 came out at $0.577B against a true $8.708B; Regions was 62× low. 71 tickers were affected.
• Worse, when a filer stopped tagging Revenuesthe series jumped without the business changing: M&T went from $9.64B to $1.54B year over year — an 84% collapse that never happened, published as clean.
• Revenue for those filers is now derived as net interest income + noninterest income — the standard definition, both components from the same 10-K, with first_filed taken as the later of the two so the sum is only knowable once both are public.
• New flag tag_switch_discontinuity: when the XBRL element changes between adjacent years and the value moves more than 50%, we say so rather than publish a fabricated trend.
• Also corrected in the same rebuild: ExxonMobil was missing entirely (SEC had repointed the ticker to a new registrant with no filings), trailing-twelve-month periods could enter as annual rows, and duplicate fiscal-year rows were collapsed. Counts moved to 311,798 rows · 5,215 companies · 18,763 restatements — companies went up because recovering orphaned registrants added more than the de-duplication removed.

Two corrections in a week is not a great look, and we would rather publish that than have you find it. Every one was found by auditing our own data against the raw filings; none was reported by a customer.

A second adversarial audit — this one run against the engine as well as the data — found four real defects. All are fixed and the full universe has been rebuilt from EDGAR. If you pulled data before this date, re-pull it.

• Assets and StockholdersEquity could carry a non-annualbalance-sheet date. A 10-K also tags quarterly equity roll-forward balances, new-standard opening balances, and subsequent-event snapshots; those were being ingested as if they were the fiscal-year close. Our “annual only” claim was therefore not true for roughly 4,300 rows. Instants are now anchored to a fiscal-year close the filing itself declares.
• Because those extra rows also consumed history slots, some companies silently had fewer than the advertised 12 fiscal years — in the worst cases a large-cap had no total-assets history at all before 2019. Recovered.
• EPSDiluted and DilutedShares are denominated in non-USD XBRL units, and the loader could pick the wrong unit array — which in some cases surfaced a decade-old figure as the current one. Each concept now requires its declared unit.
• Revenue tag coverage missed how banks, brokers, REITs and pre-ASC-606 filers tag revenue, so some companies had no revenue at all and others had a first_filed date years later than the truth. Widened.
• New: the JSON endpoints (/v1/fundamentals and /v1/snapshot) now carry data_through (our filing cutoff; the bulk download carries the same information in its x-dataset-generated-at header) and flags stale_for_as_of when the newest period we hold trails your as-of date. A value suppressed by our sanity guard is now flagged implausible_value — it can never come back marked clean again.
• Counts after the rebuild: 313,406 rows · 5,194 companies · 18,734 restatements.

Cross-sectional backtests no longer need per-ticker loops.

• GET /v1/download returns the full point-in-time dataset as one gzipped CSV (313,489 rows at the time), with an x-dataset-sha256 integrity header.
• GET /v1/snapshot?as_of= returns a whole-universe point-in-time cross-section for any date.
• Both included in the flat $49/mo plan — a feature, not a tier.

We ran an adversarial audit against our own dataset and published what it found. Fixed in this release:

• fiscal_year is now derived from period_end. This fixed an off-by-one on roughly 25,000 rows where comparative-year figures had carried the filing's year instead of their own.
• DilutedShares scale corrected for tickers whose share counts were stored in millions.
• All published counts re-synced to the live database: 313,489 rows · 5,212 companies · 18,529 restatements.
• Full-universe filing-lag stats measured and published: mean 66d · median 60d · max 120d on reliable rows.
• The free sample was re-published on GitHub, Kaggle, and Hugging Face.

• Serve-time plausibility guard on the API and company pages: a scale- or unit-corrupted value is served as null, never as a confidently wrong number.
• Anti-enumeration cap: 2,500 distinct tickers per key per day, alongside the existing 5,000 requests/day quota. Repeat calls to the same ticker don't count against it.
• API key delivery automated via Stripe webhook + email.

First public release: point-in-time US equity fundamentals from annual 10-K and 10-K/A filings, 7 concepts, one endpoint, $49/mo. Free verifiable 40-company sample on GitHub.