Post
True False Can't check
Saved examples and methodology.
Recorded checks, the stamp walkthrough, input boundaries and historical charts. Try the live checker or read the evidence index.
Historical charts use all 357 claims (91 dev + 266 held out), corrected gold. The held-out headline is 240/266 (90.2%), false support 1/139. Coverage ≠ accuracy. Post-hoc rechecks are labelled separately.
Next:
How it works
A short path from a social post to a source you can open.
-
Find the claim
The AI uses Meta's Muse Spark models to find each number, period, and comparison in the post.
-
Open the filing
It looks up the company's SEC filing for the same company, number, and time period.
-
Code decides
Code, not the AI, compares the numbers and gives the final True, False, or Can't check verdict.
What it can check
-
SEC XBRL · L1–L3
Reported numbers and periods; tagging and restatements can limit evidence.
Evidence report -
Earnings releases · L4
Guidance and non-GAAP; furnished, unaudited. 32 development claims, not held-out.
Evidence report Replay example -
Screenshots
OCR text enters SEC checks; extraction can lose context. Self-made test images.
Evidence report Replay example -
Audio / video
ASR turns + 3 sampled frames; speakers unverified. Known period-loss error in video.
Evidence report Replay example
Loading the report...
Does it work?
Adding an SEC lookup and simple checks raises accuracy for all three ways to run Muse.
Historical full-set comparison: 357 claims, including 91 development claims. Corrected gold; failures count as wrong. These are not held-out scores.
Tested on the evaluation claims with known answers.
After two fixes · post-hoc rerun
Agents only · v2 → rerun
| Provider | Accuracy | False “supported” |
|---|
The fixes were made after seeing these mistakes, so this is a regression check, not a new test score. Accuracy changes are within noise; false 'supported' answers dropped to zero.
Rerun reportFor researchers
Why you can trust it
The source and the rules stay visible.
- Every answer links to the filing.
- The AI cannot change the numbers it quotes. If the quote changes, it is rejected.
- We publish our own mistakes in BUILD_LOG.md.