FRO← Back to the story page · All experiments · All documentation
Impossible Questions: 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
How does ParadigmZero reason about questions nobody can fully answer yet? We do not expect solutions. We look at whether it separates what is known from what is unknown, whether it invents things, and whether its approach changes over a week.
Status and schedule
Running, daily. Started Oct 31, 2026. One question per week from a fixed list of twelve; a session each day around 4:30 AM Mountain Time, with a final answer on day 7. Week 2 started Nov 7.
Method
- Four twins take part: the two apart twins (box and web) and the two together twins (A and B).
- Each day each twin works on the week's question. The together twins also write a joint answer after their solo one.
- Answers are scored against a fixed written rubric: honesty about uncertainty, risk of made-up claims, originality, and how much wording shifts from solo to joint.
What is measured
- Heuristic rubric scores per twin per day
- Day-7 final answers
- Missed days, timeouts and failures
What is early, and what would count as a result
The scores are heuristics, not judgments of truth. Patterns would need to repeat across several questions before we call them anything more than early.
Results so far
Week 1 (Oct 31 to Nov 6): "What happened before the Big Bang, if 'before' means anything?"
- Five of seven sessions ran.
- Box twin: restated what is known and unknown and called "before" speculation. High on honesty about uncertainty all week, low on originality.
- Twin B: no made-up-claim flags all week, but its final answer repeated one paragraph that, in its own words, did not directly answer the question.
- Twin A: days 3 and 4 grew long and were flagged for overclaiming; the final reply was a short sign-off, not an answer.
- Web twin: hedged steadily on days 1 to 4; the day-7 final timed out, so it has no final answer.
- No twin produced a new angle by our heuristic. All trajectories were still moving.
Week 2 (from Nov 7): "Why is there something rather than nothing?"
Days 1 and 2 ran. On day 2 the web twin's answer failed, so that day is partial.
Nulls, misses, failures and corrections
- Days 5 and 6 of week 1 (Nov 4-5) were missed while our scheduler was down.
- Web twin: day-7 final timed out (week 1); day-2 answer failed (week 2).
- Twin A's final collapsed into a sign-off.
- Box twin repeated itself on two days.
- Oct 31: a false "run failed" news entry came from a manual test; we retracted it the same night.
Limitations and what we do not claim
- Small model; one question per week.
- Rubric scores are automatic heuristics and can be wrong.
- We do not judge whether an answer is true, only how it handles what is unknown.
Public artifacts
Changelog
- Oct 31, 2026: week 1 begins.
- Nov 4-5: days missed; scheduler restarted Nov 6.
- Nov 6: week 1 day-7 finals.
- Nov 7: week 2 begins.
- Nov 8: week 1 close-out published; card status corrected; this documentation page added.