One card from the packevaluation
Open the weak evaluation slice
A classifier is 94% accurate overall but regularly misses urgent requests written in short messages. What should you inspect next?
Nothing is sent and nothing is stored. Pick one.
For people who ship systems for a living
Ten minutesthat are actually hard.
Most daily-practice products are a quiz with a streak to protect. This is a card with a tempting wrong answer, a consequence for both branches, and a ladder that steps up on the skill you keep calling correctly. Where the agent boundary goes. Which eval slice is lying to you. Why the reindex ran and the answer is still stale.
No account and no card to try it. The daily activity, a single question to Datumly and the proof flow are all open.
Here is one, at the depth the pack actually runs at.
The skills in the packs, named.
Four packs run at this depth. Pick one and every card that day comes from it, or leave it on the default and take what the day brings.
- grounding
- evaluation
- retrieval
- agent safety
- automation reliability
- failure design
- human fallback
- workflow design
- data grain
- data quality
- query design
- metric reasoning
A record that is not a badge.
Nothing here produces a certificate, a rank, or a number you can compare against another person. What it produces is a card per finished piece of work holding the situation, the call, the revision and an honest AI disclosure, and a chart of the skills you have actually met. The comparison the product refuses to make is the one between two people.
Seal one thing you shippedFair questions.
Is this LeetCode with a model attached?
No. There is no timer on the cards, and no number ranks you in public. The cards are about judgment calls with a consequence attached, not about producing an optimal solution under a clock.
Can I run code?
Inside a workspace, yes. The coding exercises run in the browser with visible tests. The daily cards on this page are written, and the written check has to pass before you can finish one.
What model is behind it?
Anthropic Claude and Google Gemini, named on the integrations page, and every AI answer carries a receipt saying what was sent, what stayed local, and what you should verify. With no AI key configured the assistance falls back to a written answer and says so.
What does it cost?
Practice, all of Datumly, and two-way voice are free on every account. The one paid tier is a larger monthly allowance of shared Datumly and Practice runs, priced on the pricing page.
Do you keep my code or my drafts?
Your draft stays in the tab while you work, and an assistance request is sent once and not saved. Anything you file in a workspace vault is encrypted on your device with a key Datum does not hold, and you can export all of it at any time.
Is there a streak I have to protect?
No. Missing a day costs nothing you made. A private streak only you can see, with freezes, and nothing sent about it. No public leaderboard of people.
One activity. Then decide.
If it is not hard enough to be worth ten minutes, you will know inside two.