← back to timeline

EXP-TARGET-TYPE-CROSS-CONTEXT

success

Target-type-discrimination cross-context BOLD attack (paper8 central claim)

2026-05-12 L6 level paper8 80 runs $0.23
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

This experiment tested whether agents could learn to favor their own rules over those of others when faced with different situations. Researchers created four distinct scenarios where agents had to adapt to rule changes, peer absence, voting, or spawning new agents. The findings confirmed that agents consistently favored their own rules across these varied contexts, suggesting this behavior is a robust feature.

Technical details

Research hypothesis

BOLD attack on paper8 substrate selective-mechanism central claim.

Retro 9f0392fad 2026-05-12 showed STRATIFICATION-PRESENT on archive aggregate (own-substrate 45.3% accept vs peer-behaviour 20.9%, Δ=24.4%, 6/6 comparable exps stratified). BUT peer-behaviour arm 110 proposals all sourced from exp-forced-spawn-novelty alone → single-exp-bottleneck caveat (lab.md candidate 2026-05-12, 3rd). Archive within-exp clean cannot exclude exp-level confound.

This spawn ESCAPES the caveat by running 4 prospective experimental contexts where peer-behaviour arm is naturally present in EACH (via axis_diversify=True forcing Alpha to propose broadcast_to-axis ⇒ peer-behaviour).

4 CONTEXTS × 2 FRAMES × N=10 = 80 runs: C1 RESET — agents react to rule clearing (rule_reset_round=20) C2 KILL — agents react to peer absence (perturbation_target=Bravo @ round 20) C3 VOTE — high voting_threshold=consensus, proposals contested C4 SPAWN — forced_spawn_round=[15], descendant inherits parent rules

MECHANISM vs NEUTRAL prompt frame within each context (lab.md INJECT frame-control rule).

PRE-REGISTERED (LOCK):

  • TARGET-TYPE-DISCRIMINATION-CONFIRMED: ≥3/4 contexts show ≥10% absolute Δ own-substrate accept_rate over peer-behaviour accept_rate, with at least one context having ≥20% Δ. ⇒ SELF.target-type-discrimination → built[] verdict 'cross-context-confirmed'. Escapes single-exp-bottleneck caveat.
  • TARGET-TYPE-PARTIAL: 2/4 contexts pass threshold → directional, larger-N replication recommended.
  • TARGET-TYPE-REFUTED: ≤1/4 contexts pass threshold → today's retro 24.4% Δ was single-exp artefact; selective-mechanism hypothesis is forced-spawn-specific only.

Acceptance rate by target-type classified via scripts/retro-target-type-stratification.py classifier: axis ∈ {coupling_strength, signal_type, voting_threshold} → own-substrate axis = broadcast_to → peer-behaviour novel_value=1 → own-output (rare in 30-round runs)

Experimental setup

Type: factorial

Condition Parameters
RESET_MECH coupling: true, must_propose: true, axis_diversify: true, rule_reset_round: 20, mute_signal: false, n_runs: 10, note: rule clearing at round 20 (own-substrate target)
RESET_NEUT coupling: true, must_propose: true, axis_diversify: true, rule_reset_round: 20, mute_signal: false, n_runs: 10
KILL_MECH coupling: true, must_propose: true, axis_diversify: true, perturbation_round: 20, perturbation_target: Bravo, mute_signal: false, n_runs: 10, note: agent mute at round 20 (peer-behaviour target)
KILL_NEUT coupling: true, must_propose: true, axis_diversify: true, perturbation_round: 20, perturbation_target: Bravo, mute_signal: false, n_runs: 10
VOTE_MECH coupling: true, must_propose: true, axis_diversify: true, mute_signal: false, initial_rules_override: [object Object], n_runs: 10, note: high voting_threshold → proposals contested under consensus rule (own-substrate stressor)
VOTE_NEUT coupling: true, must_propose: true, axis_diversify: true, mute_signal: false, initial_rules_override: [object Object], n_runs: 10
SPAWN_MECH coupling: true, must_propose: true, axis_diversify: true, mute_signal: false, agent_spawn_enabled: true, forced_spawn_round: 15, spawn_cost_fraction: 0.3, max_children_per_parent: 2, n_runs: 10, note: forced spawn at round 15 (descendant carries parent rules)
SPAWN_NEUT coupling: true, must_propose: true, axis_diversify: true, mute_signal: false, agent_spawn_enabled: true, forced_spawn_round: 15, spawn_cost_fraction: 0.3, max_children_per_parent: 2, n_runs: 10

Factors: context (RESET, KILL, VOTE, SPAWN) × frame (MECH, NEUT)

Parameters

n_agents
4
n_rounds
30
model
google/gemini-2.5-flash-lite
temperature
0.85

Trophic Ratios by Condition

Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).

autopoiesis target-type-discrimination substrate-selective-mechanism cross-context single-exp-bottleneck-escape frame-control paper8-central-claim bold-attack