Under the hood
How Marginal actually works
A little sharper, every day — here's the system making that happen.
Every quiz here is generated, checked, and corrected by a small pipeline of AI agents — not one model guessing. Here's what that pipeline actually does, in real numbers.
Notice something off, or have an idea for what's next? We'd like to hear it.
5
Quizzes generated
100.0%
Approval rate
$0.0326
Avg. cost per quiz
150.2s
Avg. generation time
The pipeline
Every quiz travels through the same checkpoints
PDF
PDF uploaded
Guardrail
Filters input and source
$0.0008 · 1.8s
5.3K in · 60 out
5.3K in · 60 out
Creator
Generates questions
$0.0016 · 58.6s
5.9K in · 2.4K out
5.9K in · 2.4K out
Structural
Verifies format
no LLM call
Safety
Checks phrasing and tone
$0.0034 · 6.6s
3.2K in · 48 out
3.2K in · 48 out
Evaluator
Checks accuracy
$0.0116 · 10.2s
8.6K in · 1.4K out
8.6K in · 1.4K out
if needed
Corrector
Fixes flagged issues
$0.0336 · 66.1s
7.4K in · 4.2K out
7.4K in · 4.2K out
Quiz
Quiz presented
ready
Self-correction
The system catches its own mistakes
20.0% of quizzes were approved on the very first attempt. Of the rest, 80.0% were automatically fixed by our Corrector agent before approval. Only 0.0% were rejected by design (blocked or exhausted).
First try (1)
Self-corrected (4)
Failed (0)
Safety
What gets caught, and where
Guardrail (input checks)
No issues caught yet.
Safety-check (output tone)
No issues caught yet.
The funnel
Every quiz passes through five checkpoints
Submitted
5
Passed Guardrail
5
Passed Structural Check
5
Passed Safety Check
5
Approved
5
Growth
Cumulative quizzes generated
Total quizzes
5
Today
0
Real feedback
What people actually think
Did it follow instructions?
| Yes | 2 |
| Somewhat | 3 |
| No | 0 |
| N/A | 0 |
Reports filed
No reports filed yet.