Twelve minutes into the 25 the Rate Limiter Challenge allows, you press Submit ranked (8 cr) and the panel answers 78 / 100. That is a sum, and two lines under the number show the parts:
- "Base 63 + 5 speed bonus"
- "+10 for meeting the load SLOs"
Only the 63 is an opinion
The base comes from AI reviewers marking your design against the Challenge's rubric. A ranked submission gets two, one told to read for rigor and one for pragmatism, and the base is their average; Dual-analyst review, further down, shows them apart ("Rigor analyst: 58/100 — Pragmatism analyst: 68/100. They mostly agreed."). A Quick review uses one reviewer on the same scale and rubric. Both bonuses apply to either tier.
The 5 is arithmetic: 15, times the share of the time limit left unused, times the base as a fraction of 100, rounded. Here that is 15 × 0.52 × 0.63. The full +15 would take a perfect design submitted instantly, and a design the reviewers barely credit earns almost nothing however fast it arrives. So time is cheap: at this limit and a base of 80, five more minutes cost about 2.4 points, and rushing a weaker design usually loses more than it gains. Only the limit itself is harsh, because an attempt that reaches it closes without an evaluation.
The 10 comes from no reviewer. Challenges with an SLO badge state a load and its targets; for Rate Limiter, 20,000 requests a second, a tail latency of 150 ms and 99.9% availability. When you submit, the server simulates your submitted Grid at that load and adds +5 for each of those two targets it meets, with a row per target showing the measured value and a ✓ or ✗. The Simulate panel shows that load beforehand: see Stress testing a design before it exists. The total stops at 100. The Estimated running cost line nearby is not part of the sum, but the reviewers see that cost too, and they are told to charge a design that meets its targets only through far more capacity than it needs as over-provisioned.
When the rows say ✓ and the sum disagrees
Draw a limiter whose counters are not shared with anything, and both rows can show a tick beneath this line: "No SLO bonus: the Challenge asks for a rate limiter with a shared counter in front of the service; the limiter keeps its counters to itself, so each instance limits alone."
The targets ask whether the Grid meets its latency and availability numbers; the Challenge's stated requirements ask whether it does the job. One failed requirement withholds the whole +10, with each reason listed, and a base plus speed above 70 comes down to 70: "Score capped at 70: a requirement of this Challenge failed in the simulation." The bonus also needs traffic to enter your design and reach a component past the entry point; if none does, nothing is added.
Where the points went
A base of 63 means 37 points lost. Defects found lists them, costliest first: each mistake with what it cost (−12 pts) and buttons that jump to the nodes involved. Below it, every rubric criterion gets a mark out of 5 with its own feedback.
The criterion you cannot add afterwards
Now look for Trade-off articulation in that rubric. If you skipped the defense it is not there, and it is too late: questions can only be generated while the attempt is live. The rest of a classic grade works from what the Grid shows, so your reasoning is inferred from the boxes. Defend your design is the one place you write it down.
The section appears as soon as the Grid has a node, on classic attempts only; guided attempts have their stages instead (Guided challenges: four stages, one check each). Generate follow-ups (4 cr) writes two or three questions, each tied to a different element of your Grid: a node the simulation flags as a bottleneck, a heavily connected component, a particular connection. They ask why you chose it, what happens under load and what it cost you, and the generator is told never to reveal the answer. What you do next decides what the credits buy:
- Leave every answer blank and the criterion never joins the rubric. Answer only some and it joins, and a blank or vague answer scores low.
- Answer well and you still cannot rescue the design: the reviewers are told not to let the writing inflate the other criteria. The criterion has no fixed value in points, and it is not a third bonus.
- Change the Grid afterwards and your answers stay. The reviewers are told not to deduct when a figure quoted in a question no longer matches, so ask early, but they still judge your reasoning: reread it against the final Grid before you submit.
- Ask again and it costs again. The button now reads Regenerate (4 cr), and an answer survives only if a new question lands on the same element.
A reference is not an answer key
Reference comparison appears on Challenges that ship a reference model, as Rate Limiter does: "Your design shares 4 of the reference's 5 roles and 3 of its 4 connections. The reference is one acceptable design, not a checklist." The coverage is worked out after the grade and adds no points, and the reviewers, who read the reference privately, are told that a different design meeting the requirements is not wrong. A role counts as shared when a component that can fill it is on your Grid, wired or not, so read the connections too.
See a reference solution opens the walkthrough in the panel, for no credits and only after you have submitted that Challenge. A copied Grid teaches nothing about the next exercise, so take two sections instead: check whether your Grid fell for one of the Traps to avoid, and put the Reference trade-offs beside your defense answers to see which you never argued.
What leaves the panel
- The leaderboard takes ranked scores only: your best per Challenge ranks on that Challenge's leaderboard, and the global leaderboard sums those bests, so a weaker retake costs you nothing. A chip such as "Better than 64% of 25 graded attempts" appears once five earlier ranked attempts exist, and it counts only ranked ones.
- Your nickname is assigned at your first evaluation. Change it in your profile, or press Go incognito to stay ranked but unnamed.
- A streak counts consecutive days, by UTC, with at least one graded attempt on either tier, so a day whose attempt timed out does not count.
- Your record goes out from the same profile section: Share creates a link, and Send to a specific person emails a private one that you can Revoke on its own. It says plainly what it is: AI-graded practice feedback, not a certification.
Twelve credits, one result
Open Rate Limiter with Start classic, since the guided route has no defense. Once the first pieces are on the Grid, generate the follow-ups and answer each in two sentences, then finish the design and submit ranked: 4 + 8 credits. Pro and above include a monthly allowance, and AI credits: what they buy and how they are spent says where the rest come from. Do the sum by hand, jump to the node behind your costliest defect, and open the reference last to find the trade-off you never argued. You are done when you can name the one change the result asks of you.