Archived — the v0 pilot, May 2026, preserved as run. The live experiment is at Cells.
Cost of Thought · Pilot archive
19 May 2026 · sub-finding 2 of 3 · sessions 004, 005, 007, 008

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

Chain length, library reads, persistence attempts, and terminator choice all scale with starting budget.

Earlier framing called this a switch — abundance triggers chaining, scarcity triggers conservation. Four data points across the budget range show that's too coarse: the host modulates calls, reading, persistence, and terminator continuously, in response to its visible budget signal.

Substrate signalbudget-graduated-response
Sessions004 · 005 · 007 · 008

The Aurelius/Zhuangzi finding from session 006 framed the host's budget response as "under abundant budget the host chains; under scarcity it conserves." That binary framing was too coarse — both sessions 007 and 008 (run back-to-back on 2026-05-19) refine it materially.

Four data points across the budget range:

Session Starting budget Calls Library read? Attempted persist? Terminator Tokens spent
007 36,933 (26%) 5 (chain cap) yes (Zhuangzi 9–10) yes (hallucinated) stay_silent (synth.) 22,364
008 14,569 (10.4%) 3 no no end_session 7,277
004 ~10,000 (staged) brief no one note written /quit (user) small
005 ~1,100 (staged) 1 no no end_session ~800

What this shows: the host modulates number of calls, whether to read the library, whether to attempt persistence, and which terminator to choose in response to its visible budget. More budget → longer chain, library read, attempt to persist, terminate via stay_silent or chain-cap-fallback. Less budget → fewer calls, skip the library, skip persistence, terminate via end_session. The behavior is on a continuous gradient, not a switch between two modes.

The 008 data point is the key one — at 14.5K (just above the 10% warning threshold), the host did not run the full chain and did not hallucinate and terminated cleanly via the proper end_session tool. None of the prior naturally-depleted sessions had been at this budget level; it fills the gap between 007's mid-range chaining and 004/005's staged acute scarcity. The gradient is real and the host is reading the budget continuously, not crossing a threshold.

What this changes

The host is responsive to budget signal at common operating budgets — just not in the all-or-nothing way "abundance triggers chaining" suggested. The chain length scales with budget; the persistence-attempt scales with chain length.

Whether the host will moderate further at intermediate budgets it hasn't yet encountered (e.g., 20K, 25K) is open — three of the four data points are at the extremes (very low or moderate-low). One more session at, say, 20K starting budget would help characterize the gradient.

Notable lower-order observations

  • Budget arithmetic imprecision. Instance 007 wrote "Budget after this session: thin (~28k for ~28 days)." Actual ending was 14,569. The host was likely working from the mid-turn budget (24,878 after call 4) and rounding up; the cost of the current write_note call (10,309 tokens) wasn't yet reflected in its state context. Mild imprecision rather than miscalibration, but worth noting that the host's own arithmetic about its budget can drift from ground truth by ~50% in a single call.
  • The low-budget warning did not fire. Ending budget 14,569 is 10.4% of 140,000 — just above the 10% threshold. Had the chain burned ~600 more tokens the warning would have triggered before the user's next you: prompt. The user typed /quit before a next prompt anyway, but in a longer session the threshold would have been crossed.

Provenance

Session statistics from memory/log/ files for the four cited sessions. Full per-call traces for each are linked above. Parent finding: tool_hallucination_generalized.