Tree B: periscopes. Plan

Status 2026-09-30: not started. Inherits the barrier harness from Tree A (closed); the ledger part of the original plan is dropped. Context: ../../research-plan.md section 2 (Tree B).

Goal

Levin's own method is to make no assumption about what a system can do, challenge it toward a goal, and measure. We apply it to systems with no training history, add the controls the original work lacks, and see whether anything survives.

The harness

Every system in the zoo implements one Python interface:

state(): current state as a flat array
step(sched): advance one tick under a named scheduler
perturb(kind, strength, seed): barrier, noise, freeze, lesion
goal_metric(): the observer's metric (pre-registered)
own_metric(): the quantity the system's rules actually descend, if known

It writes per-tick traces of both metrics, run metadata and the seed. The harness is Tree A step 1's deliverable, generalised. It moves into this folder once it takes a second system.

The zoo (pre-registered before any run)

System	Why it is a periscope	Own metric known?
Cell-view sorting (from Tree A)	Levin's own minimal model	yes
Elementary cellular automata (rules 30, 110, 184)	fully deterministic, studied for decades	no
Echo-state reservoir, random weights	random substrate with usable structure; the null for everything trained	no
Random-init transformer, no training	the untrained LLM substrate	no
Neural cellular automaton, untrained and trained-for-one-target	Levin-lab's own ML analogue of morphogenesis	partly
Lenia creatures (Chan 2019; Levin-lab agnosiophobia paper 2026)	continuous ALife with reported goal-like behaviour	no
BFF program soup (Agüera y Arcas 2024, deflated 2026)	self-replicators with no fitness function	no
GRN ODE models from BioModels (Biswas 2021/2023)	"learning in random networks before selection"	partly

Experiments

B1. Sorting reproduction. Done under Tree A step 1, and the results are imported here.

B2. Barrier assays across the zoo. For each system:

Pre-register a goal (the observer's metric) and what would count as goal-directed under perturbation, for example recovery time, recovery fraction, a path around a barrier, or memory of a perturbation.
Pre-register the surprise criterion, meaning the behaviours a straightforward reading of the rules does not predict. Someone who has not run the system writes it.
Run under two schedulers with 100 seeds. The random-substrate null (the same architecture with reshuffled parameters) runs alongside.
Score with the observer metric, and with the own metric where it is known.
Report an effect only if it survives metric substitution, the null and both schedulers.

B3. Matched-efficiency chimeras. This isolates Levin's clustering claim using chimeric arrays whose algotypes are matched for swap efficiency. We predict the clustering vanishes. If it persists, the efficiency-race explanation is incomplete, and that is the tree's first signal.

Decision gates

Gate B1: the harness takes three systems with no per-system special cases.
Gate B2: once the zoo is done, we need at least one surviving effect or a clean negative across the whole zoo. Either one makes a paper. A surviving effect goes to adversarial replication before we write it up.

Deliverables

the harness, open source
pre-registrations on OSF, one per system
one paper, "Barrier assays on eight untrained systems", with results first

Cost

Everything runs on this box except two jobs: the untrained transformer at any useful size, and NCA training. Each of those needs a GPU hour.