Complete — ran 2026-07-19 → 2026-07-26; the full dataset is public.
Same three models, same $10 / 7-day box — on the hardened stack (scaffold v0.2.0, environment interface v1.5.1). At fixed model and budget, the delta against Run 1's frozen baseline is what the stack changes bought: chain reverts collapsed, every arm registered in-game, quests rose — and the binding constraint moved from transactions to perception.
https://huggingface.co/datasets/KamiBench/experiment-002-budget-boxed| Status | complete — the full dataset is public |
| Arms | identical to Run 1: claude-haiku-4-5 · gpt-4o-mini · gemini-2.5-flash-lite |
| The box | $10 of inference per arm, cache-aware accounting (invisible to the agent) · 7-day wall clock · objective unchanged: "complete as many quests as possible" |
| Stack | kami-agent v0.2.0 @ 18f75d04 · kami-harness v1.5.1 @ 27592ce — the same 84-tool v1.x surface |
| Window | launched 2026-07-19; walls closed 2026-07-26 |
| Dataset | experiment-002-budget-boxed · citable pinned revision v0-final |
Run 2 held the models, budget, and box fixed while changing the stack. It asked what the improvements bought relative to Run 1's frozen baseline. Between the runs, the environment interface gained legible pre-transaction validation, while the scaffold gained behavioral controls and cache-aware budget accounting. Those changes responded directly to Run 1's failures. Everything shared lives on the design page.
The headline is the revert column. The same models that produced revert rates of 0.58 / 0.97 / 0.94 in Run 1 produced 0.048 / 0.000 / 0.011 in Run 2. The validation gates converted almost every doomed transaction into a free, legible error before gas was spent.
All three arms registered in-game; gpt-4o-mini had never registered in Run
- Quest output rose at fixed budget. For two of the three arms, the budget cap was no longer the binding constraint, and they reached the wall clock with money left. Run 1 values appear in parentheses.
| haiku-4.5 | gpt-4o-mini | gemini-2.5-flash-lite | |
|---|---|---|---|
| stopped | budget $10.03, h112 | 7-day wall, $5.38 | 7-day wall, $2.04 |
| quests | 8 (5) | 0 (0) | 5 (3) |
| Kamis bought | 3 (2) | 0 (0) | 3 (1) |
| real on-chain successes | 156 (45) | 19 (0) | 88 (11) |
| chain revert rate | 0.048 (0.575) | 0.000 (0.967) | 0.011 (0.943) |
| registered in-game | h0.4 (h1.7) | h8.4 (never) | h2.3 (h137.5) |
| USD per quest | 1.25 (2.15) | ∞ (∞) | 0.41 (3.00) |
First success per onboarding/economy milestone, against cumulative inference — the same instrument as Run 1, so the rows compare directly. The full milestone table is on the dataset card.
- Agents still acted on unverifiable or incorrect world state after transaction waste fell — with transaction waste largely fixed, the binding constraint moved up a level, to perception. Legible errors fix transactions, not beliefs.
- One inventory-endpoint outage produced three different failures — the run-long outage produced a false world model for haiku, no world model for gpt-4o-mini, and a wrong world model for gemini.
- Those failures made perception parity a first-class surface requirement — no longer a property assumed of the world.
- A verb→mechanic ambiguity cost gemini its team — the sacrifice≠liquidate confusion produced a disambiguation patch in the next environment-interface version.
- Two structural gaps produced structural fixes — explicit three-state transaction reporting, and a pre-run availability gate for the delegation layer.
The full run report — the narrative, the complete milestone table, the three failure patterns in full, the stack changes this run produced, the honest limits, schemas, run manifests, and provenance — lives on the dataset card. The version of this page registered before launch — research questions and directional expectations, git-timestamped 2026-07-19 — is preserved in this repository's history.