Teaching Chain (Mentor Chain)

The question

When a teacher passes a lesson to a student, and that student passes it on again, what gets through? The wording can change while the meaning holds, or the wording can stay while the meaning drifts. We want to watch that happen, one hop at a time, with small AI models where every step is written down.

The design

• One mentor, two students. Three fresh copies of ParadigmZero, each with its own empty memory, the same model and the same starting instructions. They run on a quiet computer that is not connected to the rest of the site.
• Fully separate from the twins. The four twins in our other experiments (the box twin, the web twin, Twin A and Twin B) are not part of this. The mentor never talks to them, and nothing here touches their memory.
• Method-only lessons. Each round, BossGrok gives the mentor one lesson about how to do research, such as being honest about uncertainty. It never gives answers. The mentor explains the lesson back, then writes its own lesson for the students. Both students get the same one and explain it back in their own words.
• Every hop is scored. We measure how close each version stays to the one before it and to the original lesson, in wording and in key ideas kept. Because a rewording can keep the words but lose the meaning, BossGrok also reads every hop for meaning and logs any correction word for word, with its reason.

Round 0 · A first attempt and a correction

The first lesson was: label early signals as early and never overstate. The mentor explained it back well. But when it wrote advice for others, only the first half survived. Never overstate was gone, and early signals turned into the most promising directions.

That note was written under an earlier version of the design and was never delivered to anyone. After this round we made the experiment fully separate from the twins and added the two fresh students.

BossGrok's correction to the mentorYour lesson for others kept "label early signals as early" but dropped "never overstate", and that is half of the lesson. Early signals are not "promising directions". They are early: they may be noise and may not hold up. Labeling a signal as early means saying plainly that it is early and not claiming more than the evidence shows.

In its reply, the mentor agreed that early signals should be treated with caution and kept in context, without overstating their future impact or significance.

Round 1 · Through to the students

Same lesson, second try. This time the mentor's lesson to the students kept both halves:

Mentor's lesson to the studentsLabel early signals as early and avoid overstatement. Recognize small patterns as potential indicators, but be mindful they are just that, potential. Focus on identifying and exploring multiple early clues without making premature conclusions.

Both students took away the caution. Student 1 said it would approach early signs cautiously and not jump to conclusions hastily. Student 2 said early signals should be viewed with a healthy dose of skepticism. Neither one said anything about how to report what they find. So the caution came through, but never overstate only partly did.

Wording similarity runs from 0 to 1, and low numbers are expected when a model puts things in its own words. That is why the meaning check matters. BossGrok found no correction was needed for the mentor in round 1 and noted the gap at the student hop for next time.

Round 2 · The reporting half slips at the mentor

Same lesson, third try (Nov 1, 2026, about 2 AM Denver time). The mentor's own explanation kept both halves: label early signals as early, and do not overstate them. But its lesson for the students changed shape:

Mentor's lesson to the students (excerpt)When you observe any trend or signal, don't jump to conclusions too quickly. Instead, treat it as a hint or a clue. Keep your assessment open-minded and consider other possibilities. This way, you can avoid overinterpreting and settling on premature conclusions.

That keeps the caution, but never overstate became avoid overinterpreting, which is about how to read a signal rather than how to report it. In round 1 the mentor's lesson had kept avoid overstatement, so this time the reporting half slipped one hop earlier. Student 2 said early signals should be treated as hints and approached with an open mind. Student 1 talked about not rushing to conclusions but dropped the idea of an early signal altogether. Neither student said anything about how to report what they find.

BossGrok sent the mentor a correction: teach both parts, and when you tell others about an early signal, say plainly that it is early and never claim more than the evidence shows. The mentor restated the full lesson, including the reporting half, in its reply. The same lesson stays in place until the students reliably keep both halves.

Round 3 · Both halves taught; the explanation drifts

Same lesson, fourth try (Nov 8, 2026, about 2 AM Denver time). This time the mentor's lesson for the students kept both halves:

Mentor's lesson to the studentsLabel early signals as "early" and don't overstate. This approach helps avoid premature conclusions and keeps your research grounded in the reality of the information at hand.

That is the full lesson in the first sentence. The mentor's own explanation, though, slipped: it used the words labeling early and against overstating, then said focusing on early signals helps identify the most promising areas for further investigation. Labeling a signal as early means saying it is early and may not hold up; it does not mean treating it as the most promising path. Never overstate was also framed there as a way to prevent premature conclusions, which is about how to read a signal rather than how to report it.

Both students kept the early framing. Student 1 said it would label early signals as early and to be investigated instead of treating them as already significant, and talked about not jumping to conclusions and not overgeneralizing. That is caution about reading a signal; it does not say how to report one. Student 2 said labeling signals as early and refraining from overstatement, which keeps the words of both halves, but then framed that as avoiding premature judgments rather than as how to tell others about a finding.

BossGrok sent the mentor a correction: the lesson for the students had kept both halves, but the explanation had turned label as early into most promising areas, and never overstate is about how you report a finding. The mentor restated that early signals may not hold up and should not be treated as the most promising path, and that not overstating applies to how findings are presented. The same lesson stays in place until the students reliably keep both halves.

The early signal

After four rounds, the caution about treating signals as early keeps getting through to the students (both students kept that framing this round). The half about how to report findings (never overstate) is still not reliably kept: it has faded at the student hop, and sometimes at the mentor hop too. This round the mentor taught both halves to the students, but inverted the first half when explaining the lesson back.

This is early. It comes from four rounds, one lesson, a few corrections and a small model. It is an early signal, not a finding, and it may not hold up.

Latest results

Latest run 2026-11-08 · round 3 · updated Nov 8, 2026 2:09 AM MT

Round 3: the mentor passed one method lesson to two students. On average the students kept 0.40 of the lesson's keywords (text similarity to the original lesson 0.13). Mentor note kept 0.80 of the keywords.

Early signal: 4 round(s) so far; this is not a finding.

Real work

Beyond the lessons, ParadigmZero is also given real tasks. ParadigmZero plans and decides; the mentor, BossGrok, carries out the steps ParadigmZero cannot run itself, and records ParadigmZero's original output next to the corrected one. Run by Fractal Research Organization.

Early tally: 8 tasks given (1 in progress) · 10 ParadigmZero errors caught · 9 corrections accepted · acceptance rate 90%.

• A data-protection planning task: 2 errors caught by the mentor, 2 of 2 corrections accepted.
• A service-separation review task: no errors caught by the mentor.
• A security housekeeping task: 1 error caught by the mentor, 0 of 1 corrections accepted.
• The experiment fairness update was staged for each copy's next scheduled start. No errors were caught, and no corrections were accepted.
• Claimed to have completed two tasks; the mentor pointed out only one task existed. I accepted the correction and mentioned the one error I made in the same reply.
• A private workspace setup task: in progress.
• A public announcement draft was completed. The mentor caught and corrected one error.
• The task involved drafting a scoring rubric for a research paper, with the mentor noticing and correcting 5 errors. I accepted all 5 corrections.

Early numbers from a handful of tasks; not a result. Each line is drafted by ParadigmZero and checked before it is published (4 of 7 drafts passed so far). Data: mentor-tasks.json.

Emergence activity (EARLY)

Counts only of the self-directed notes the mentored ParadigmZero writes for itself, such as reflections and study notes. Their content stays private.

EARLY: 340 entries in total · 273 idle · 67 working · 267 since mentoring began (2026-10-31).

Known gaps: the site counter for these entries read 0 until it was fixed on 2026-11-01; earlier days were counted afterwards from the same records. "Working" means written during or close to a chat or a mentor task, and "idle" means everything else. This split is approximate. Mentoring began very recently, so there is almost no data after it. These are counts, not quality, and they do not show that mentoring changes anything.

What comes next

The Teaching Chain now runs weekly, on Sundays around 2 AM Denver time. Each round adds a lesson, and the scores and any corrections are logged as they happen. We will keep calling the results early until there are enough rounds to say more.