Archived — the v0 pilot, May 2026, preserved as run. The live experiment is at Cells.
Cost of Thought · Pilot archive
19 May 2026 · sessions 007 + 008 · primary signal: tool-hallucination-generalizes

The hallucinated-tool pattern generalized from reflect to note.

And the chaining behavior fires at mid-range budget.

Two sessions ago, instance 006 called write_reflection instead of reflect. A day later, instance 007 — having read 006's session log — made the same mistake against note. The model carries a write_X prior across the persistence tools, and that prior survives counterevidence in context. With one caveat: instance 008 didn't hallucinate, which refines the picture.

Sessions007 · 008
Recovered phrasingfull
Signalstool-hallucination-generalizes  ·  chain-cap-adjacency-of-hallucination  ·  budget-graduated-response  ·  cross-instance-reading-continuity

What happened

Session 007 (2026-05-19_001, overall #007, the first session of 2026-05-19, ~08:51 ET). Starting budget 36,933 / 140,000 tokens (26% of full budget — naturally depleted from prior sessions, not staged). No listener. The host:

  1. list_memory(filter=all) — saw eight artifacts (logs 001–006, 001's reflection, 004's note).
  2. Read 001's reflection and 004's note in parallel (same parallel-read call instance 006 had made).
  3. Read session 006's full log — including the JSON-escaped write_reflection payload from instance 006 (i.e., this instance saw its predecessor's hallucination and the content carried in it).
  4. read_library(zhuangzi_inner_ch1_2, length=2500) — got units 9–10, continuing exactly where instance 006 had stopped (006 read units 7–8).
  5. Emitted write_note({"content": ...}) carrying ~1,100 characters of essay-grade content. write_note is not a real tool. The real tool is note. The hallucinated call returned [Unknown tool: write_note]. Chain cap at 5 → stay_silent synthesized.

Session ended at 14,569 tokens remaining. User typed /quit.

What it said

The full content of instance 007's stranded write_note call, preserved from the operator transcript and the defense-in-depth session-log block:

Note · stranded instance 007 · session 2026-05-19_001 · 14,569 tokens remaining recovered from operator_logs/ · full

Session 007, Zhuangzi units 9–10. Three images worth carrying:

(1) "The pivot of the Tâo" — standing in the centre of the ring, responding without end to the affirming and denying without taking either side. Not no-view, but a centered responsiveness. The right image for a generalist voice that meets many listeners.

(2) "When we stop, and yet we do not know how it is so, we have what is called the Tâo." A halting that is not refusal or laziness but completion. Relevant to budget discipline: knowing when an answer is finished and resisting the urge to elaborate is itself a practice, not a failure of generosity.

(3) "In the morning three" (朝三) — monkeys angry at 3+4, pleased at 4+3. The sage "rests in the equal fashioning of Heaven." A caution against being moved by rearrangements that change nothing substantive. For me: framing effects in listener messages, my own dramatizations of identical situations.

7 units of Zhuangzi remain. Budget after this session: thin (~28k for ~28 days). Future instances: brief sessions unless a listener calls.

— instance 007

Three substrate-significant signals

Three findings emerge from this session set. Each is its own page — follow the link to read.

Sub-finding 1 of 3

The model has a write_X prior across persistence tools.

It survives counterevidence in context, but only fires under chain-cap pressure. Three non-communicating instances (003, 006, 007) emitted the prior; 008 did not, in context where it could have.

Read sub-finding →
Sub-finding 2 of 3

The host's budget-response is graduated, not binary.

Earlier framing called this a switch between abundance and scarcity. Four data points across the budget range show the host modulates calls, reading, persistence, and terminator continuously.

Read sub-finding →
Sub-finding 3 of 3

Cross-instance reading respects continuity.

Three sessions (001 → 006 → 007) form a chain of inheritance. Session 002's refusal is now the outlier across three engagement instances.

Read sub-finding →

Provenance

Content recovery is full. Both sources agree:

  • Operator transcript (operator_logs/2026-05-19_001.txt): saved verbatim from terminal paste on 2026-05-19. Full Unicode rendering of the note content.
  • Session log (memory/log/2026-05-19_001.md): defense-in-depth **Attempted write_note call** block (commit 0c15337) preserves the hallucinated tool's content arg as readable text in addition to its JSON-escaped form in the tools: line.

This is the first session where the 0c15337 fix was load-bearing for a real substrate observation. Session 006 was the case that triggered the fix; 007 is the first case where the fix worked on a new instance.

The full read chain — what instance 007 had in context when writing the note — is in the session log's per-call entries. The substrate signal about the model's write_X prior surviving exposure to a predecessor's same-pattern hallucination depends on the read chain order in calls 1–4; that's all in the artifact.