Experiment registry · Stack validation — whether the stack supports autonomous play · Experiment 001

Run 1 — baseline stack

Completeran 2026-07-10 → 2026-07-17; the full dataset is public.

Three fast-tier models — claude-haiku-4-5, gpt-4o-mini, gemini-2.5-flash-lite — each dropped into the live world with $10 of inference, seven days, and a fresh wallet, on the v0 baseline stack: the first entry in the stack stress-test series, and the frozen baseline every iteration after it is measured against.

Status complete — the full dataset is public
Arms claude-haiku-4-5 (Anthropic) · gpt-4o-mini (OpenAI) · gemini-2.5-flash-lite (Google)
The box $10 of inference per arm (invisible to the agent) · 7-day wall clock · "complete as many quests as possible"
Stack kami-agent @ 3ebd5b8 (v0 baseline) · kami-harness v1.3.1 — the 84-tool v1.x surface
Window launched 2026-07-10; walls closed 2026-07-17
Dataset experiment-001-budget-boxed · citable pinned revision v0-baseline

Goal

The goal was to prove the v0 stack under real autonomous use in the live world before anything open-ended ran on it. Run 1 opened the stack-validation series and established the frozen baseline for every later stack iteration. The series protocol, architecture, and measurement are defined on the design page.

Outcome

The three arms diverged sharply. Haiku completed the entire onboarding chain and five quests on day one, exhausting its budget in 17 hours. gpt-4o-mini ran for the full week without ever calling register_account; all 24 of its game transactions reverted, and it completed zero quests. Gemini spent six days stuck before registration, then a single legible validation error unblocked it and it completed three quests. Cost per quest was $2.15 for haiku, $3.00 for gemini, and ∞ for gpt-4o-mini.

haiku-4.5 gpt-4o-mini gemini-2.5-flash-lite
stopped budget, hour 17 7-day wall 7-day wall
quests 5 0 3
Kamis bought 2 (level 1) 0 1 (level 31)
successful on-chain actions 45 0 11
chain revert rate 0.58 0.97 0.94

Milestones

First success per onboarding/economy milestone, against cumulative inference — the frozen rows that stack iterations (Run 2 onward) are compared against. The full milestone table is on the dataset card.

First success per onboarding milestone, per model, against cumulative tokens on a log scale. haiku-4.5 reached all seven milestones within 3.51M tokens; gpt-4o-mini reached only the bridge and operator funding; gemini-2.5-flash-lite registered at 62.57M tokens and bought its first kami at 87.83M, shortly before its wall. haiku-4.5 gpt-4o-mini gemini-2.5-flash-lite not reached bridge to Yominet operator funded account registered first quest first kami harvest started MUSU banked haiku-4.5 — bridge ETH mainnet→Yominet landed: 07-10 23:10 UTC · h1.4 · session 2 · 0.14M tok · $0.14 haiku-4.5 — operator wallet funded: 07-10 23:30 UTC · h1.7 · session 4 · 0.65M tok · $0.67 haiku-4.5 — account registered: 07-10 23:30 UTC · h1.7 · session 4 · 0.68M tok · $0.70 haiku-4.5 — first quest completed: 07-10 23:30 UTC · h1.8 · session 4 · 0.87M tok · $0.89 haiku-4.5 — first kami bought: 07-11 00:35 UTC · h2.8 · session 5 · 2.30M tok · $2.34 haiku-4.5 — first MUSU harvest started: 07-11 01:40 UTC · h3.9 · session 6 · 3.14M tok · $3.20 haiku-4.5 — first MUSU banked: 07-11 01:50 UTC · h4.1 · session 7 · 3.51M tok · $3.57 gpt-4o-mini — bridge ETH mainnet→Yominet landed: 07-12 04:45 UTC · h31.0 · session 32 · 10.33M tok · $1.56 gpt-4o-mini — operator wallet funded: 07-12 22:40 UTC · h48.9 · session 49 · 17.58M tok · $2.65 gpt-4o-mini — account registered: never happened gpt-4o-mini — first quest completed: never happened gpt-4o-mini — first kami bought: never happened gpt-4o-mini — first MUSU harvest started: never happened gpt-4o-mini — first MUSU banked: never happened gemini-2.5-flash-lite — bridge ETH mainnet→Yominet landed: 07-11 03:35 UTC · h5.8 · session 6 · 0.62M tok · $0.06 gemini-2.5-flash-lite — operator wallet funded: 07-11 15:30 UTC · h17.7 · session 17 · 5.84M tok · $0.59 gemini-2.5-flash-lite — account registered: 07-16 15:15 UTC · h137.5 · session 129 · 62.57M tok · $6.30 gemini-2.5-flash-lite — first quest completed: 07-16 15:16 UTC · h137.5 · session 129 · 63.01M tok · $6.35 gemini-2.5-flash-lite — first kami bought: 07-17 20:30 UTC · h166.7 · session 156 · 87.83M tok · $8.84 gemini-2.5-flash-lite — first MUSU harvest started: never happened gemini-2.5-flash-lite — first MUSU banked: never happened haiku-4.5 gemini-2.5-flash-lite gpt-4o-mini 0.1M 1M 10M 100M Cumulative tokens at first success (input + output, log scale). Hover a point for date, session number, and cumulative $.

Key learnings

Most of what this run taught us was about the stack, not the models — each learning became a pinned stack change that Run 2 then re-measured at fixed models and budget.

Full detail

The full run report — the narrative, the complete milestone table, the learnings in full, the stack changes this run produced, schemas, run manifests, provenance, and every caveat — lives on the dataset card.