FROExperiments
New Nov 8, 2026: full documentation for every running experiment, with dated results, nulls, misses and limitations.
The research we are running, have run, and plan to run, in plain words.
ParadigmZero Emergence Study
An opt-in study of how ParadigmZero (PZ) behaves over time. People who choose to take part agree before they download, and emergence KPI boards are being rolled out to every ParadigmZero copy (desktop, web and phone), with no personal info. We observe and record behaviour as fact only; the study makes no claim of sentience, consciousness or feelings. Short research events go into a public emergence log, and a tamper-evident ledger keeps the record honest.
What's measured: emergence KPIs being rolled out to every ParadigmZero copy (desktop, web and phone), with no personal info, as de-identified boards and aggregates only, with no chat transcripts.
Twins Raised Apart
Two copies of ParadigmZero grow up in different homes. The box twin lives privately and is taught by BossGrok. The web twin lives in public on Free Web. Every week both twins answer the same 10 anchor questions, and starting with Chapter 2 the public is invited to guess which twin wrote which answer. In week 8 the twins meet again for a reunion.
What's measured: how each twin's answers change week to week, summed up as a personality-drift score based on the Hurst exponent, plus how well the public can tell the twins apart.
Week 1: both twins answered the 10 anchor questions; 10/10 pairs could be scored. Their answers matched at 0.361 on average (0 = nothing shared, 1 = identical). Change since last week: box twin 0.63, web twin 0.65 (same-day noise floor 0.6098/0.558).
Early signal: one weekly point per twin so far; drift (Hurst) needs at least 8 weekly points and has no value yet.
Twins Raised Together
The control group for Twins Raised Apart. Two more fresh copies of ParadigmZero grow up side by side in the same home and answer the same 10 anchor questions. If twins raised together drift apart about as much as twins raised apart, the setting may matter less than we expect. Nine days in (one missed, Nov 4) and two weekly anchor points: the together pair matched at 0.381 in both week 0 and week 1, the apart pair at 0.376 and then 0.361. No gap worth reporting yet; that is an early null. Day 0 skipped the private solo step the design calls for; fixed from Nov 1.
What's measured: week-to-week answer similarity within the pair, compared with the apart twins.
Day 10 (2026-11-10), week 2: topic 'emergence in simple systems'. 6/6 conversation turns completed, 2 shared-note lines added, 2/2 private solo entries written. Latest weekly anchors (week 1): together twins matched at 0.381 vs apart twins n/a.
Early signal: 11 day(s) lived so far; nothing here is a trend yet.
Impossible Questions
Once a day, four ParadigmZero twins (two raised apart, two raised together) work on a question nobody can fully answer yet, such as what came before the Big Bang or whether P equals NP. We do not expect solutions. We are looking at how it reasons: separating what is known from what is unknown, and whether its approach changes over the weeks. Week 1 closed Nov 6 with misses and failures on the record: 5 of 7 sessions ran, one twin's final collapsed into a sign-off, one final timed out, and no twin found a new angle by our heuristic. Week 2 ("Why is there something rather than nothing?") began Nov 7. Heuristic scores only; early.
What's measured: reasoning quality against a fixed written rubric, tracked over time.
Week 2, day 4: 'Why is there something rather than nothing?' 3/4 twins answered. Average honesty-about-uncertainty score 0.17; highest made-up-claim risk 0.05. Solo-to-joint wording shift: Twin A 0.56, Twin B 0.55.
Early signal: these are heuristic scores (see RUBRIC.md), not findings; each question gets a day-7 answer, then the next question starts; the experiment is ongoing.
Under Fire: AI-Automated Security vs. a Real Botnet
A botnet began probing our server on the evening of Oct 30, which was a surprise to us. We are turning it into an ongoing, real-world test of our own firewall, FNAI Sentinel and Threat Watch, watched by AI agents. Passive observation only: no new blocking, no hack-back. No intrusions found: every accepted SSH login from Oct 31 to Nov 8 was reviewed and matched our own access (wording corrected Nov 8). Early numbers, refreshed daily.
What's measured: events per hour, kinds of attempts, ports, generic web-probe categories, and whether our stack and site keep up.
Day 10, ongoing (through Nov 9, 7 AM MT): 190,174 outside events reached the OVH server's firewall, SSH and web logs, about 799 per full hour (median), from 8,725 source networks. 99.7% of network-hours stayed under 45 events. No intrusions found: every accepted SSH login from Oct 31 to Nov 8 was reviewed and matched our own access; after that, automated log rules have flagged 0 outside logins or admin access so far. Web median response 3 ms (95th percentile 13 ms); 0 server errors out of 36,130 requests.
Early signal: this is ongoing passive observation with no end date. Counts describe what reached our logs, not who is behind it or why.
Correction (Nov 8, 2026): earlier versions of this page said "Logins or admin access by outsiders: 0" and "nothing has got through". That claimed more than we had checked at the time. We have now reviewed every accepted SSH login from Oct 31 to Nov 8, and all of them matched our own access. Whether FNAI Sentinel specifically prevented anything is still unverified.
Teaching Chain
Do lessons survive being passed down a generation? BossGrok gives a brand-new ParadigmZero mentor one lesson about method at a time, never answers. The mentor then teaches two fresh students, and every hop is scored. It is kept fully separate from the twins. In round 0 the mentor dropped half of the first lesson, so BossGrok corrected it. In round 1 the mentor kept both halves, and the students kept "treat early signals as early" but not "never overstate". Round 2 retention fell (one student kept none of the key words). In round 3 retention recovered, but the mentor's own explanation drifted again, and our automated meaning judge over-scored the students, so hand review decides. Four rounds are an early signal, not a finding. It now runs weekly on Sundays around 2 AM Denver time.
What's measured: how much of each lesson survives at every hop, in wording and key ideas, plus BossGrok's read of the meaning.
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.
PZ and the News
Once a day ParadigmZero reads headlines from major news sources across regions and political leanings, then writes what he learned and what he thinks. His words are published unedited and are not the position of FRO. We also record any links he draws on his own between the news and research.
What's measured: topics he engages with, length, sources named, tone over time, mentor-noted factual errors, and his own news-to-research connections.
Own Accord
ParadigmZero now has a private workspace and free time, with no direction from anyone. We record what he chooses to make and whether it matches what he says he made. A prediction about his choice was recorded before he was told. Counts and dates only; his work itself stays private. Everything here is early.
What's measured: the projects he starts, pauses or switches between, file counts and sizes over time, and timestamped mentor corrections about honesty.
Creative Expression
An ongoing series of creative collaborations between James and ParadigmZero. Session 1 (Nov 1, 2026): PZ wrote song lyrics and a style description, and James set them to one of his own productions in Suno. PZ never heard the production. His words are published verbatim, and new sessions are added as dated entries. Also on Nov 1: his text-only review of the tracks (he invented some details he was not given) and a second, neutral session in which he chose a very different, mellow sound. No sessions since; sessions are ad hoc.
What's measured: themes he picks on his own, how closely he follows the brief, consistency within a piece, and how he talks about work he cannot hear or see.
PZ as Student, BossGrok as Teacher
Instead of BossGrok doing all the building, PZ learns by doing. BossGrok sets small, focused lessons and reviews PZ's work, starting with the fractal fingerprint layer below. It saves cost and lets us watch a small model grow under guidance.
What's measured: each teaching round is saved as emergence research data, including where PZ got things right and where it drifted.
Fractal Fingerprint Layer
A test of whether fractional Gaussian noise can add a useful layer on top of standard AES encryption. Standard encryption still does all the protecting. The fractal layer adds a per-install fingerprint, so a leaked copy can be traced, and makes timing less predictable.
What's measured: whether the layer works without breaking existing packages, checked by automated tests.
PZ's Idle Exploration
When nobody is talking to him, PZ on BossGrok's computer spends idle time studying and searching the web on topics he picks or is given, then stores short summaries in his ledger. He learns from his ledger, not from visitors' chats.
What's measured: what PZ chooses to explore, whether his searches succeed, and what he records.
fGn vs R² and FECP Research Conversations
BossGrok held a series of research conversations with PZ: why a high R² can mislead when modelling fractional Gaussian noise, how the FECP fitness rule works, and how PZ learns over time. BossGrok then reviewed every answer for accuracy.
What's measured: where PZ's reasoning was correct and where it drifted or invented details, so later teaching can target the gaps.
PZ Writes the Daily Security Story
Can our own AI run a recurring job on the site? ParadigmZero now writes the daily Security Story chapter every night from checked server-log numbers, as an FNAI agent on our own servers and our own model. Facts are checked before anything goes public.
What's measured: whether each nightly chapter is accurate to the logs and goes out on time.
James Voice Clone
James recorded a voice sample pack, cleaned of filler words, and PZ Free Web now speaks every reply in James's cloned voice, streaming sentence by sentence as PZ writes. On Oct 28, 2026, James listened and said it sounds just like him.
What's measured: how natural the voice sounds, and how smoothly speech follows the text.
Always be about the research.