Tree E: patterns as agents. Plan
Status 2026-09-30: parked. Last in the order by design. Context:
../../research-plan.md section 2 (Tree E) and
../../Review.md section 5 (the Sheldrake test). Only E1 is
planned; E2 and E3 are listed so the reasons for not doing them are on
record.
Goal
Levin's most radical claim is that at the top of the spectrum the agent is the pattern and the body is not. Stripped of ontology, it says that some behavioural attractors belong to the data or to language and to no single model, and they have their own dynamics across substrates. E1 tests the first half only.
E1: cross-model attractor survey
The candidates are the "spiritual bliss attractor", documented in Claude system cards and partially replicated in Llama 3.3 70B by Berg, Keeling and Street, along with persona attractors and specific reasoning failure modes.
Steps:
- Define each attractor operationally, as a state the conversation converges to under a specified protocol, with a classifier to score it.
- Survey open-weight models with documented, non-overlapping post-training (OLMo, Pythia base, Qwen, Llama, Mistral), taking base models and chat models separately.
- For each attractor, check whether it appears in models with no shared post-training data. Where the training data is open (OLMo, Pythia), trace the attractor to it.
- Kill condition: every recurring attractor traces to shared post-training or shared pretraining sources. In that case E1 is a data effect and the tree closes.
E2 and E3: not planned
- E2 (patterns propagate between models and humans with measurable fidelity and mutation). This is memetics, which has a long history of producing nothing quantitative. Not before E1 survives.
- E3 (steer a pattern out of one model and watch it reappear in another trained on its outputs). Of all the experiments, this one reads as morphic resonance. Not before the group has a record of published deflations, and not without an adversary co-signing the pre-registration.
Rules specific to this tree
- No result is ever described as a pattern wanting, seeking or pressing. We report recurrence rates and data provenance.
- Every write-up passes the three discriminators from Tree G before it leaves the group.
Cost
Inference only, on open weights. It needs little compute, and its cost is to our reputation.