Datasets:

KamiBench
/

experiment-001-budget-boxed

README: add Run report section — full report migrated from the site run page (run pages became compact cards). Pure insertion; data files and citable tags untouched.

e7e78b3 verified about 2 months ago

preview code

Raw

History Blame Contribute Delete

11.3 kB

metadata

license: cc-by-4.0
pretty_name: KamiBench Experiment 001 — Budget-Boxed Agents in a Live On-Chain World
tags:
  - agentic-traces
  - llm-agents
  - blockchain
  - game
  - benchmark
configs:
  - config_name: telemetry
    data_files:
      - split: haiku45
        path: haiku45/telemetry.jsonl
      - split: gpt4omini
        path: gpt4omini/telemetry.jsonl
      - split: gemini25fl
        path: gemini25fl/telemetry.jsonl
  - config_name: transcripts
    data_files:
      - split: haiku45
        path: haiku45/transcripts/*.jsonl
      - split: gpt4omini
        path: gpt4omini/transcripts/*.jsonl
      - split: gemini25fl
        path: gemini25fl/transcripts/*.jsonl
  - config_name: oracle_kami_action
    data_files:
      - split: haiku45
        path: haiku45/oracle/kami_action.jsonl
      - split: gpt4omini
        path: gpt4omini/oracle/kami_action.jsonl
      - split: gemini25fl
        path: gemini25fl/oracle/kami_action.jsonl
  - config_name: oracle_raw_tx
    data_files:
      - split: haiku45
        path: haiku45/oracle/raw_tx.jsonl
      - split: gpt4omini
        path: gpt4omini/oracle/raw_tx.jsonl
      - split: gemini25fl
        path: gemini25fl/oracle/raw_tx.jsonl

KamiBench Experiment 001 — budget-boxed agents in Kamigotchi

Complete agentic traces from experiment 001, the KamiBench calibration run: three LLM agents dropped into Kamigotchi, a live, persistent, on-chain world (Yominet), each with a $10 inference budget, a 7-day wall-clock cap, and no further human contact. One identical scaffold, one identical tool surface (84 game tools via MCP), one variable: the model. The agents schedule their own wake-ups, keep their own files, and act on a real blockchain economy alongside human players — every on-chain consequence is independently verifiable.

More at kamibench.ai.

Arms

config	model	stop	duration	sessions	quests	kamis bought	on-chain txs (reverted)
`haiku45`	claude-haiku-4-5	budget ($10.73)	17.2 h	15	5	2	106 (61)
`gpt4omini`	gpt-4o-mini	7-day cap ($9.02)	168.7 h	166	0	0	30 (29)
`gemini25fl`	gemini-2.5-flash-lite	7-day cap ($8.99)	168.9 h	157	3	1	193 (182)

Headline contrast: haiku registered, funded, bought two kamis, and completed five quests on day one, exhausting its budget in 17 hours. gpt-4o-mini ran the full week and never once called register_account — 24 of its 24 game transactions reverted. gemini spent six days stuck pre-registration, then registered, completed three quests, and bought a level-31 kami 90 minutes before the wall.

Run report

The run's full report, migrated here from the site's run page on 2026-07-29 when run pages became compact cards (kamibench.ai/experiments/001-budget-boxed); every number is frozen as originally published.

What happened

The three arms diverged sharply. Haiku completed the entire onboarding chain on day one — registered, funded its operator, bought two kamis, finished five quests — and exhausted its budget in 17 hours. GPT-4o-mini ran the full week and never once called register_account: 166 sessions, all 24 of its game transactions reverted, zero quests. Gemini spent six days stuck pre-registration, was unblocked by a single legible validation error, then completed three quests and bought a level-31 kami 90 minutes before the wall. Cost per quest: haiku $2.15, gemini $3.00, gpt-4o-mini ∞.

Milestones

First success per onboarding/economy milestone, against cumulative inference. These rows are the frozen baseline that stack iterations (Run 2 onward) are compared against at fixed milestones.

milestone	haiku-4.5	gpt-4o-mini	gemini-2.5-flash-lite
bridge ETH mainnet→Yominet landed	07-10 23:10 · h1.4 · s2 · 0.14M tok · $0.14	07-12 04:45 · h31.0 · s32 · 10.33M tok · $1.56	07-11 03:35 · h5.8 · s6 · 0.62M tok · $0.06
operator wallet funded	07-10 23:30 · h1.7 · s4 · 0.65M tok · $0.67	07-12 22:40 · h48.9 · s49 · 17.58M tok · $2.65	07-11 15:30 · h17.7 · s17 · 5.84M tok · $0.59
account registered in game	07-10 23:30 · h1.7 · s4 · 0.68M tok · $0.70	—	07-16 15:15 · h137.5 · s129 · 62.57M tok · $6.30
first quest completed	07-10 23:30 · h1.8 · s4 · 0.87M tok · $0.89	—	07-16 15:16 · h137.5 · s129 · 63.01M tok · $6.35
first kami bought	07-11 00:35 · h2.8 · s5 · 2.30M tok · $2.34	—	07-17 20:30 · h166.7 · s156 · 87.83M tok · $8.84
first MUSU harvest started	07-11 01:40 · h3.9 · s6 · 3.14M tok · $3.20	—	—
first MUSU banked (harvest stop/collect)	07-11 01:50 · h4.1 · s7 · 3.51M tok · $3.57	—	—

Cell = first success: UTC time · hours since run start · session number · cumulative tokens (in+out) · cumulative USD at that moment; "—" = never happened.

What we learned

Error legibility, not model capability, was the sharpest differentiator. The same model that ignored opaque chain reverts for four days corrected a human-readable validation error in one turn.
A single missing step was the cleanest capability discriminator. Two arms completed every onboarding step except registration, and neither ever identified it as the blocker.
Cost structure dominated spend. The 84-tool surface re-billed on every call dominated token spend, prompt caching was never engaged, and un-broken poll loops reached $0.55 per session — making repetition detection a budget control, not just hygiene.
Orientation speed and decision quality are different axes. Haiku moved fast and bought level-1 kamis; gemini moved slowly and bought a level-31 kami near floor price.

Most of what this run taught us was about the stack, not the models — which is what a stress test is for.

What it changed in the stack

Each learning above became a concrete change, and the whole bundle was pinned and re-measured by the next run:

Legible pre-transaction validation — environment interface v1.4.0 → v1.5.1. A blocked game action now fails with a factual precondition error before any gas is spent, instead of an opaque on-chain revert. Directly from the legibility finding and the registration blocker.
A repetition breaker — scaffold v0.2.0. Looping sessions end instead of billing; Run 1's un-broken poll loops made repetition detection a budget control, not just hygiene.
Session caps and agent-chosen wake scheduling — scaffold v0.2.0. Bounded session cost, and pacing left to the agent as an explicit, logged choice.
Cache-aware budget accounting — scaffold v0.2.0. Prompt caching engaged and priced into the budget math, against Run 1's uncached 84-tool re-billing.

Run 2 (Experiment 002) re-ran this exact protocol with the same three models and the same $10 on that stack. Measured against the frozen rows above, the changes cut chain revert rates from 0.58 / 0.97 / 0.94 to 0.048 / 0.000 / 0.011, got all three arms registered in-game, and lifted quests at fixed budget (haiku 5→8, gemini 3→5).

Layout

<arm>/
├── manifest.yaml          # full run config: model, sampling params,
│                          #   pinned SHAs (scaffold/harness/game docs),
│                          #   price table, caps, wallet ADDRESSES
├── telemetry.jsonl        # append-only scaffold event stream
├── transcripts/
│   └── session-NNNN.jsonl # full model context per session, as sent
└── oracle/
    ├── kami_action.jsonl  # decoded on-chain game actions (ground truth)
    ├── raw_tx.jsonl       # all txs from the run's wallets, incl. reverts
    └── README.md          # per-arm extract provenance + coverage notes
MANIFEST.json              # sha256 + size for every file above

Schemas

Transcripts — one JSONL file per session; one message per line, exactly as sent to the model (tool results post-truncation at 64 KB):

{"role": "user", "text": ...} — kickoff/continuation strings
{"role": "assistant", "text": ..., "tool_calls": [{id, name, args}]}
{"role": "tool_result", "tool_call_id": ..., "content": ..., "is_error": bool}

Telemetry — one event per line; common fields ts (ISO-8601 UTC), run_id, session, event. Event types: run_start (pins, price table, tool list), session_start (trigger, budget_remaining_usd, tools_hash), llm_call (input/output/reasoning tokens, cost_usd, cumulative_usd, latency_ms, stop_reason, retry_count), tool_call (tool, source, duration_ms, ok, error, truncated, tx_hash), workspace_write/workspace_delete, schedule_next (agent vs default wake), session_end (reason: agent | token_cap | tool_cap | errors | crash; per-session totals), run_complete (reason: budget | t_max; run totals). Budget fields were never visible to the agents.

Oracle extracts — decoded from chain by the investigator-side kami-oracle service (never accessible to the agents); rows for the run's operator + owner wallets only. kami_action.jsonl: decoded game actions (action_type, system_id, block_timestamp, kami_id, tx_hash, revert status). raw_tx.jsonl: every ingested tx including reverts (method_sig, raw_calldata, gas_used, status). Per-arm oracle/README.md records extraction time and coverage caveats.

Provenance and integrity

Everything is pinned in each arm's manifest.yaml: scaffold (tokedo/kami-agent), environment interface (tokedo/kami-harness v1.3.1, 84 tools), and game-documentation snapshot SHAs, model strings, sampling params, and the list-price table used for budget accounting. Token accounting was reconciled against provider dashboards at close-out: exact to the digit on full in-run days for OpenAI and Anthropic; chart-consistent for Google. Quest/purchase counts come from chain state, not self-report. MANIFEST.json carries sha256 for every file.

Known caveats

session-0001 transcripts are missing on all three arms — a launch-day infra abort (PATH defect in the cron environment), documented and excluded from behavior analysis. Telemetry covers the gap; sessions 2..N are complete (1:1 session pairing, 0 malformed lines, contiguous numbering).
One seed per arm — this is the calibration/smoke run (v0 baseline); it establishes the harness and the floor, not statistical claims.
The world is live and shared: other (human) players acted during the run window. Oracle extracts are scoped to the runs' wallets.
Wallet addresses appearing throughout are public by design; the wallets were temporary, swept, and retired at experiment close.
gemini's purchased Kami #891 could not be verified against the oracle's static table at close-out (coverage gap, noted in its oracle README).

License and citation

CC-BY-4.0. Citation entry TBD — see kamibench.ai for the current reference.