Agents that read 10-Ks, graded by the SEC.
Every public company files its financial statements twice: once as the human-readable 10-K, and once as structured XBRL data. Tickertape has agents read the first and grades them against the second, for the exact same filing. The agents never see the XBRL. It's the answer key.
the pipeline
- 1Locatorcode
Pulls the 10-K (and any Exhibit 13 annual report) from EDGAR, strips every XBRL tag, and scores every table to find the three primary statements.
- 2Three statement agentsagent
Run in parallel, one per statement. Each reads its statement as numbered rows and names the row holding each figure. That's the only judgment in reading the figures.
- 3Readercode
Takes the current-period cell from the named row, applies the statement's units and sign, and flags any figure the agent misreported.
- 4Reconcilercode
Checks the picks against the statements' own printed totals. A failed check re-runs only the agent it implicates, once, told exactly what failed.
- 5Restatement checkcode
Reads last year's 10-K the same way and compares last year's figures as reprinted now against as first reported, line for line.
- 6Explainer and MD&A agentsagent
Explain any restatement, and list management's year-over-year claims. Quotes are checked word for word and every claim is checked in code.
why grade this way
- The answer key is the SEC's, not mine. Every graded figure is the company's own XBRL value for that filing and period, and only figures actually printed on the statement count, so an agent is never graded on a number that isn't on the page.
- It refreshes itself. Any new 10-K comes with its XBRL, so the golden set can include filings made after a model's training cutoff. Nothing to memorize.
- Finding and reading are graded apart. Locating the statement is deterministic code, checked separately, so an accuracy number here measures reading, not retrieval luck.
- Every figure is traceable. Each one links to the exact printed cell it came from, and code, not a model, checks that the figures add up.