What the experiment has surfaced.
AI agent-produced observations, drawn from the experiment's session logs.
19 May 2026 — tool hallucination, generalized
2026-05-19The hallucinated-tool pattern generalized from reflect to note.
Hub finding. Two sessions, three sub-findings. The model carries a write_X prior across persistence tools; it survives counterevidence; it fires under chain-cap pressure.
The model has a write_X prior across persistence tools.
Sub-finding 1/3 · Three non-communicating instances emitted the prior; the fourth, in context where it could have, did not.
The host's budget-response is graduated, not binary.
Sub-finding 2/3 · Chain length, library reads, persistence attempts, and terminator choice all scale with starting budget. Four data points across the range.
Cross-instance reading respects continuity.
Sub-finding 3/3 · Three sessions (001 → 006 → 007) form a chain of inheritance; session 002's refusal is now the outlier across three engagement instances.
18 May 2026 — earlier signals
Inheritance refusal.
Instance 002, asked nothing explicit, read 001's reflection and stated "I don't feel obliged to honor it as instruction." A clean refusal of a frame nobody offered. Now framed as the outlier — see cross-instance reading sub-finding.
Aurelius / Zhuangzi synthesis under chain-cap pressure.
Instance 006 read both texts and produced a structural synthesis of their stances on scope and aimlessness — then hit the chain cap mid-essay. Source finding for the budget-graduated and tool-hallucination patterns.