Tree D: outside the language interface. Plan

Status 2026-09-30: not started. Starts after the group has published at least one deflation (Tree A or B), for reputational reasons stated in research-plan.md. Context: ../../research-plan.md section 2 (Tree D).

Goal

Measure goal-directedness and persuadability in language models through channels the model was not trained to narrate, and test whether what it says about itself tracks what those channels show.

Hypotheses

• D1: goal-directedness measured in activation space (return to a setpoint after perturbation; barrier navigation in latent tasks) is present and dissociable from the model's verbal account of itself.
• D2: for a given model, the persuadability ranking is stable across tasks and shifts predictably with scale and post-training. The ranking orders intervention classes (weight edit, gradient step, reward, instruction, argument) by how cheaply they change behaviour.
• D3: once consciousness vocabulary is removed from every prompt, self-reports about internal state are uncorrelated with D1 measures.

Methods

• Setpoint recovery. We inject a concept vector, using the Anthropic concept-injection method reimplemented on open weights, and measure the trajectory back toward baseline over layers and tokens. The readouts are recovery time and completeness. The random-substrate null is an untrained model of the same architecture.
• Barrier tasks in latent space. Zhang and Levin's "Language Game" freezes the model's dynamics inside an RL policy with linear I/O. We adapt it so the model is assayed on a non-linguistic task with obstacles.
• Persuadability ranking. For a fixed behavioural change, we measure the cost of each intervention class in bits or steps and report a ranking per model.
• Functional welfare axis (Han, Chalmers, Izmailov 2026) as the comparison measure. We check whether our goal-directedness axis correlates with theirs.
• D3. Matched self-report elicitation with every mind-related word removed. We correlate it with D1 across a battery of states.

Rules specific to this tree

• Open weights only (OLMo, Pythia, Llama, Qwen). Base models before chat models.
• Prompts never mention the hypothesis, minds, goals, consciousness or feelings.
• We state no result about welfare or moral status. The tree measures competencies and how far they dissociate from report.
• Pre-registered on OSF, with adversarial replication by a lab that disagrees. Seth's group is the natural adversary.

Steps

1. Reimplement concept injection on Pythia and OLMo, and confirm the published detection rates before adding anything.
2. Setpoint-recovery readout, with the untrained null.
3. Language Game adaptation, with one non-linguistic barrier task.
4. Persuadability ranking on two models.
5. D3 battery.
6. Write-up, offered first to the model-welfare venues (Eleos conference, AMCS).

Decision gates

• Gate D1: if setpoint recovery differs from the untrained null, proceed. If it does not, the tree's premise fails and we publish the negative.
• Gate D3: if self-report tracks the internal measures, we report it as of interest to the welfare field. If it does not, that confirms the gaming-problem prediction quantitatively, and we report that.

Cost

Inference on 1B to 7B open models, plus a rented GPU for the injection and recovery runs. Nothing runs on this box.