Eden

This week's results

Every night Eden trains, then answers questions it was never trained on. These are those scores, published every week.

This week (seven days)
Nov 5 to Nov 11, 2026
Version
0.1.0
Nights trained this week
6
Nights trained in all
6 since Nov 5, 2026
Subjects scored
5 of 7

Eden is in training and testing. As it passes our benchmarks, it will power more of Forage Academy.

  1. Science

    Last scored Nov 11, 2026. 300 held-out questions.

    88.3%

    likely between 84.2% and 91.5%

    265 of 300 right

    Untrained starting model
    85.3%
    Frontier model
    91.3%

    On the same 300 questions, Eden was right 11 times where the frontier model missed, and missed 20 the frontier model got. That gap is within chance.

    Weekly trend, one point per seven days

    First week of results. The line starts next week.

    1 night of training this week

  2. Bible

    Last scored Nov 11, 2026. 255 held-out questions.

    93.3%

    likely between 89.6% and 95.8%

    238 of 255 right

    Untrained starting model
    88.2%
    Frontier model
    98%

    On the same 255 questions, Eden was right 3 times where the frontier model missed, and missed 15 the frontier model got. That gap is larger than chance.

    Weekly trend, one point per seven days

    First week of results. The line starts next week.

    1 night of training this week

  3. Math

    Last scored Nov 8, 2026. 40 held-out questions.

    72.5%

    likely between 57.2% and 83.9%

    29 of 40 right

    Untrained starting model
    57.5%
    Frontier model
    77.5%

    On the same 40 questions, Eden was right 6 times where the frontier model missed, and missed 8 the frontier model got. That gap is within chance.

    Weekly trend, one point per seven days

    First week of results. The line starts next week.

    3 nights of training this week

  4. Wellness

    Last scored Nov 8, 2026. 131 held-out questions.

    98.5%

    likely between 94.6% and 99.6%

    129 of 131 right

    Untrained starting model
    94.7%
    Frontier model
    96.9%

    On the same 131 questions, Eden was right 4 times where the frontier model missed, and missed 2 the frontier model got. That gap is within chance.

    Weekly trend, one point per seven days

    First week of results. The line starts next week.

    3 nights of training this week

  5. Geography

    Last scored Nov 11, 2026. 150 held-out questions.

    92.7%

    likely between 87.3% and 95.9%

    139 of 150 right

    Untrained starting model
    88.7%
    Frontier model
    97.3%

    On the same 150 questions, Eden was right 2 times where the frontier model missed, and missed 9 the frontier model got. That gap is within chance.

    Weekly trend, one point per seven days

    First week of results. The line starts next week.

    6 nights of training this week

  6. Language arts

    Not scored yet. The question bank for this subject is still being filled, and Eden is scored only once there are enough questions to hold some back.

  7. American history

    Not scored yet. The question bank for this subject is still being filled, and Eden is scored only once there are enough questions to hold some back.

The bars run from 40% to 100%, the same scale for every subject.

How to read these numbers

Held-out questions
Questions from Forage Academy's banks that are kept out of training, so Eden is answering them for the first time. The big number is the share it got right.
The 95% range
A score from 40 to 300 held-out questions (the count is beside each subject) is an estimate. The range is where the true score very likely sits, so two nights whose ranges overlap may not differ at all.
Untrained starting model
The open model Eden begins from, before any of our training. The distance between it and Eden is what the nightly training added.
Frontier model
The large commercial model Forage Academy uses today, asked the same questions. It is the bar Eden is working to meet.
What this measures
How well Eden knows Forage Academy's own curriculum, closed book. It is not a general benchmark. Some held-out questions may share related facts with training questions, so treat a gain as an upper bound.
What it never contains
Any child's words, name or details. This page is built from counts and scores alone.

The same numbers are available as results.json.