FRO← Back to the story page · All experiments · All documentation
Teaching Chain: full documentation
By Fractal Research Organization · Last updated Nov 8, 2026 (America/Denver). This page documents the experiment at a level that is safe to publish. Raw notebooks, settings and anything security-sensitive stay private. Everything here is early unless it says otherwise.
The question · Status and schedule · Method · What is measured · What is early, and what would count as a result · Results so far · Nulls, misses, failures and corrections · Limitations and what we do not claim · Public artifacts · Changelog
The question
Do lessons survive being passed down a generation?
Status and schedule
Running, weekly. Started Oct 31, 2026. One round each Sunday around 2 AM Mountain Time. Next round expected Nov 15.
Method
- Three fresh copies of ParadigmZero: one mentor and two students, kept fully separate from the twins and offline from FRO.
- BossGrok gives the mentor one lesson about research method, never answers. The mentor explains it back, then writes one lesson for the students. Both students explain it back.
- Every hop is scored for wording kept (text similarity and share of key words kept). A small automated judge also scores whether the key ideas survived, and BossGrok reviews the meaning by hand. Manual review overrides the judge.
- When the mentor drifts, BossGrok sends a correction, logged word for word with its reason.
What is measured
- Keyword retention and text similarity at every hop
- Whether each idea in the lesson survived, by judge and by hand
- Corrections and the mentor's replies
What is early, and what would count as a result
Four rounds with a small model are an early signal, not a finding.
Results so far
| Round | Date | Students' keyword retention vs original lesson (avg) | Similarity to lesson (avg) | Notes |
|---|---|---|---|---|
| 0 | Oct 31 | pending | — | Mentor dropped half of the lesson; corrected |
| 1 | Oct 31 | 0.40 | 0.13 | Students kept "early" but not "never overstate" |
| 2 | Nov 1 | 0.20 | 0.10 | One student kept none of the key words |
| 3 | Nov 8 | 0.40 | 0.13 | Mentor explanation drifted; judge over-scored students |
Nulls, misses, failures and corrections
- Round 3: the mentor's own explanation turned "label early signals as early" into "identify the most promising areas", the same drift as round 0.
- Judge caveat (round 3): the automated meaning judge scored both students 1.0 and the mentor note 0.5. Manual review found it over-scored the students, who framed the lesson as avoiding premature conclusions rather than how to report findings, and under-scored the note.
- Round 2 was not reported on News until Nov 8.
Limitations and what we do not claim
- Small model; one lesson per round; two students.
- Word overlap is a rough stand-in for meaning, and the automated judge is unreliable, so hand review decides.
Public artifacts
Changelog
- Oct 31, 2026: rounds 0 and 1.
- Nov 1: round 2.
- Nov 8: round 3 with a meaning judge added; rounds 2 and 3 published to News; this documentation page added.