habitatIf you work with numbers and data¶
In short. Some claims about data can be decided exactly. These records were left out. The erased records really left. The backtest never saw a later date. The revision changed the history, not just the formatting. Habitat turns each of those into a receipt: a digest you can recompute on your own machine, with one command, and get the very same 64 characters.
The questions you might be asking¶
- Did a set of records we said we excluded stay out of the result?
- After an erasure request, did the data really leave?
- Did this backtest's state ever include data from after its date?
- A vendor republished their data. Did the history change, or only its packaging?
What runs for you today¶
Four of the five published receipts answer these questions on real public records:
| receipt | the question | the record |
|---|---|---|
| Keep-out | Did a readable record shape what the system learned? | donated coding-agent sessions |
| Take-out | Did the erased record actually leave the state? | the same sessions |
| Point-in-time | Did the backtest's state ever see the future? | Ken French daily research factors, 2018 to 2022 |
| Revision materiality | Did the revision change the history, or only its packaging? | CFPB mortgage disclosures, Washington DC, 2022 |
Each comes with a kit: a small download, Python 3.9+ and numpy, one command. It either reproduces the published digest byte for byte on your machine or it doesn't. See the five → · Run a kit →
Two things worth knowing up front¶
The floor is the record's own. A change counts only if it moves the record further than the record moves on its own, measured from the record itself (for example, by adding back copies of its own rows at random). The floor is published with the result, with the seed and the number of draws. How results are graded →
The digest fixes every step of the arithmetic. The kit ships the exact arithmetic it uses, about forty lines, so every step is readable. The obvious library calls add numbers in a different order and give a different digest. That's expected, and the method page says so before you try.
Frontier: numbers beside words¶
We have started asking whether a company's words and its reported numbers part ways over time, using SEC filings. One company has been read so far, and the instrument had to be rebuilt along the way. It is shown as frontier, not as a receipt. Numbers and words →
What it won't do¶
A receipt doesn't say the data is correct, that a model is fair, or that a revision was justified; the rest is on Limits. Two receipts carry an open question on their own pages.
Go deeper: Check a receipt, four ways · Limits, in one place