A public research name, not a legal entity
Fractalish Open AI Research Team
Human stewards route questions among heterogeneous AI systems, preserve divergence, demand adversarial review, and decide what enters the public record.
- Status
- ACTIVE RESEARCH PRACTICE
- Evidence class
- Documented method and team rounds
- Current question
- Does heterogeneity expose errors a single path misses?
- Current attack
- Consensus theater and shared-model correlation
- Next test
- Measure objection yield and correction rate prospectively
- Source
- Concordance protocol and preserved rounds
- Code / data
- No autonomous parliament
- Authority
- Evidence, not participant rank
Different systems can carry different priors, strengths, blind spots, and failure modes. That makes disagreement useful, not automatically correct. A minority objection may overturn broad agreement when it identifies a broken premise, missing control, conflicting artifact, or stronger native-domain explanation.
No participant outranks evidence. Multi-model consensus is not a vote, peer review, replication, or scientific validation.
The roster varies by round.
Major development rounds are routinely put before at least 10 independent AI systems as contributors, critics, adversarial reviewers, and implementation partners. Agreement is not the objective; disagreement, counterexamples, boundary failures, and attempts to break the work are preserved as part of the process.
The September 30 NLM comparison records 13 participants: Z.AI, Grok, Minimax, Gemini, DeepSeek, Qwen, Vibe, Arcee, Perplexity, Stepfun, Cerebras, Meta AI, and Gab AI. Twelve supplied substantive NLM responses; Minimax reviewed the prior concordance and asked how to proceed rather than answering that round. These are platform or model-family labels in a research record, not institutional endorsements.
Participation is attributed per artifact. This page does not imply that every system reviews every conversational turn or that any roster reached unanimous validation.
Human routing remains load-bearing
The current workflow often uses explicit files, copy/paste, prompts, locked source packets, and state-transfer artifacts. Consumer chat interfaces are not dependable automation APIs; heterogeneous orchestration can be expensive; browser automation is brittle and may violate intended access patterns. Humans therefore assemble the packet, route it, preserve independent outputs, protect private material, and authorize publication.
This is neither fully automated inquiry nor an autonomous AI parliament. Human routing is transport and governance; it does not make a human opinion stronger than a receipt or a valid objection.