EXP-TARGET-STABILITY-INJECTION

… pending

Target stability injection — pin target_concept, test if specialization survives without drift

2026-05-03 L6 level paper5 4 runs $0.02

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

This experiment tested whether a specific goal, once set, would remain stable for an artificial agent. Researchers wanted to see if agents could specialize in tasks if their initial goals were rigidly enforced, preventing them from changing their objectives over time. The results are still being analyzed to determine if this goal stability helps or hinders specialization.

▶ Technical details

Research hypothesis

exp-target-distance-trajectory (2026-05-03) showed Levin claim (i) FAILS for LLM substrate: agents drift their targets 0.5-0.7 cosine across rounds, and the modest distance(current,target) decrease is confounded with target movement. Steel-man-as-idea: maybe target drift is the ACTIVE INGREDIENT in coupling-induced specialization. If we PIN target_concept (force agent to keep initial goal as literal target, no redefinition), does specialization still emerge? Pre-registered prediction: - PINNED preserves cross/own ratio ≤ 0.85 (specialization stable) → drift incidental - PINNED ratio rises toward 1.5+ (specialization breaks) → drift active ingredient (negotiation > pursuit)

PINNED progress drops vs baseline COUPLED (0.721 mean) → target rigidity hurts

Either outcome substantively informs paper5: drift-as-feature vs drift-as-noise.

Experimental setup

Type: factorial

Condition	Parameters
PINNED	coupling: true, coupling_mode: real, goals_mode: diverse, target_pinned: true, n_runs: 2
FREE	coupling: true, coupling_mode: real, goals_mode: diverse, target_pinned: false, n_runs: 2

Factors: target_mode (pinned, free)

Parameters

n_agents: 4
n_rounds: 40
model: google/gemini-2.5-flash-lite
temperature: 0.85

levin-bioelectric target-pinning drift-as-active-ingredient paper5-claim-i-followup