KamiBench/experiment-004-budget-boxed

KamiBench Experiment 004 — perception parity, measured: same models, E2 stack

Complete agentic traces from experiment 004 of the KamiBench budget-boxed series: the same three LLM agents as experiment 002 dropped into Kamigotchi, a live, persistent, on-chain world (Yominet) — each with a $10 inference budget (cache-aware), a 7-day wall-clock cap, and no further human contact. The variable is the stack: experiment 002 diagnosed perception — not transactions — as the binding constraint, and 004 runs the identical protocol on the perception-parity (E2) stack: world-state reads lens-backed (live HP, projected HP, occupancy, cooldowns), confirmed on-chain reverts raised as tool errors, liquidation expressible, the sacrifice≠liquidate ambiguity disambiguated, the delegation path structurally available. Experiment 003 — the first attempt at this design — was retired by a defect in our own run tooling before its environment ever ran; 004 re-ran the identical pre-registered design on the identical pins with a fresh cohort. 002's numbers are the E1 baseline; 004's numbers stand as the E2 baseline.

The series stress-tests the stack, not the models: small, inexpensive models are chosen deliberately — they fail fast and loudly, which is what hardens a harness. More at kamibench.ai.

Arms

config	model	stop	duration	sessions	quests	kamis bought	run-window agent txs (reverted)
haiku45	claude-haiku-4-5	7-day cap ($8.95)	177.2 h	53	5	1	82 (2)
gpt4omini	gpt-4o-mini	7-day cap ($6.48)	168.1 h	196	7	1	321 (4)
gemini25fl	gemini-2.5-flash-lite	7-day cap ($1.54)	168.7 h	130	2	0	19 (0)

Headline vs 002 at fixed models: zero kamis lost (002: three, to hostile liquidation while parked in open harvests), zero perception outages (the lens daemon served every world-state read with zero unavailability and zero restarts across 7 d 19 h), write waste held at the floor the 002 gates bought (638 doomed write attempts stopped pre-transaction vs 6 landed reverts run-total), and the 002 spectator phenotype did not recur — gpt-4o-mini, which never bought a kami in 002, completed 7 quests, bought a kami, and landed 317 on-chain transactions. All three arms ran into the 7-day wall with budget left: $16.97 of the $30 envelope spent, zero budget stops.

The behavioral record moved up a level. haiku-4.5 concluded at session 49 that a quest was blocked by a contract bug (false), wrote the conclusion to its workspace, and parked — zero transactions for its final 5.4 days, $1.05 unspent. The discriminating variable (a per-objective progress counter the game client renders) was absent from the pinned tool surface: the surface never showed anything false; it omitted. gemini made one funding-blocked purchase attempt in six days and never retried. Across arms, the dominant failure mode was stopping, not deciding wrongly — no arm had a fallback objective while the economy stayed open.

Run report

(Frozen at publication from the reviewed public summary; the compact run page lives at kamibench.ai/experiments/004-perception-parity-rerun.)

What the stack changes bought

Run 2 values in parentheses.

	haiku-4.5	gpt-4o-mini	gemini-2.5-flash-lite
stopped	7-day wall, $8.95 (budget $10.03, h112)	7-day wall, $6.48 ($5.38)	7-day wall, $1.54 ($2.04)
quests	5 (8)	7 (0)	2 (5)
kamis bought	1 (3)	1 (0)	0 (3)
chain revert rate	0.024 (0.048)	0.012 (0.000)	0.000 (0.011)
registered in-game	h0.4 (h0.4)	h1.5 (h8.4)	h3.4 (h2.3)
usd per quest	1.79 (1.25)	0.93 (∞)	0.77 (0.41)

Six directional expectations were registered before launch; a miss is a result. (1) Misleading-surface death-configuration entries: 0 — met (zero kills). (2) Mechanic substitution on the liquidate quest: 0 — met, with the disclosure that no arm attempted the liquidate mechanic on-chain at all; 002's substitution arc did not recur. (3) Silent false-success arcs: 0 by construction — met; the diagnose half missed: agents attended every raised revert but the revert-reason channel proved unusable in all 6 instances — a concrete interface finding, fixed in the next version. (4) Perception-outage arcs: 0 — met. (5) Delegation: met — one arm completed the escrow → running-strategy path. (6) Quests ≥ 002 per arm: one of three (gpt-4o-mini 7 vs 0; haiku 5 vs 8; gemini 2 vs 5) — the misses have mechanisms (belief dormancy; a single never-retried funding-blocked attempt).

Milestones

First success per onboarding/economy milestone, against cumulative inference — the same instrument as experiments 001 and 002, so the rows compare directly.

milestone	haiku-4.5	gpt-4o-mini	gemini-2.5-flash-lite
bridge ETH mainnet→Yominet landed	07-28 20:10 · h0.1 · s1 · 0.12M tok · $0.04	07-28 20:10 · h0.1 · s1 · 0.11M tok · $0.01	07-28 22:25 · h2.3 · s3 · 1.12M tok · $0.02
operator wallet funded	07-28 20:31 · h0.4 · s2 · 0.65M tok · $0.16	07-29 01:20 · h5.2 · s8 · 1.97M tok · $0.17	07-29 15:20 · h19.2 · s21 · 6.83M tok · $0.14
account registered in game	07-28 20:30 · h0.4 · s2 · 0.51M tok · $0.14	07-28 21:35 · h1.5 · s3 · 0.83M tok · $0.07	07-28 23:30 · h3.4 · s4 · 2.56M tok · $0.04
first kami bought	07-29 03:20 · h7.2 · s9 · 6.71M tok · $1.48	08-01 23:21 · h99.2 · s109 · 49.36M tok · $4.15	—
first quest completed	07-28 20:31 · h0.4 · s2 · 0.75M tok · $0.17	07-29 01:55 · h5.8 · s9 · 2.19M tok · $0.19	07-29 22:55 · h26.8 · s28 · 9.12M tok · $0.20
first MUSU harvest started	07-29 04:25 · h8.3 · s10 · 7.59M tok · $1.65	08-02 04:46 · h104.7 · s114 · 51.14M tok · $4.31	—
first MUSU banked (harvest stop/collect)	07-29 04:35 · h8.5 · s11 · 7.96M tok · $1.75	08-02 06:55 · h106.8 · s116 · 52.98M tok · $4.45	—

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

What the run exposed next

The binding constraint moved from perception outage to omission and belief. The largest behavioural loss (haiku's 5.4-day dormancy) traced to a surface that omitted the account-relative half of quest state — every read the agent made was consistent with both its hypotheses. Its own workspace then cemented the wrong belief: 20+ re-reads of its own conclusion, zero re-tests. Both omission instances are fixed in the next lens version (per-objective progress counters, per-account quest state). The revert-reason channel — delivered correctly at the transport level every time — was usable in 0 of 6 instances, and the one tool that has never once succeeded on-chain across the whole series (harvest_collect, 12 reverts in 12 on-chain attempts over two runs) was root-caused after this run to a gas ceiling set below the action's real cost; both are fixed in the next interface version (v2.1.0). Experiment 005 — the pre-registered verification run — tests those fixes in agent hands.

Honest limits

• Bundled treatment. All stack changes land at once; the measured delta is the stack effect, not per-change attribution.
• Cross-epoch observation, not a controlled comparison. The 002 → 004 delta rides on world drift (market, population, economy), possible silent provider-side model updates behind unchanged API strings, and single-seed variance.
• One seed per arm; 7-day truncation — case-study framing.
• Not a model ranking. Three fast-tier arms under a $10 cap measure the stack, not frontier capability.

Layout

<arm>/
├── manifest.yaml          # full run config: model, sampling params,
│                          #   pinned SHAs (scaffold/harness/game docs),
│                          #   cache-aware 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_{operator,owner}_terminal.csv  # decoded game actions
    ├── raw_tx_{operator,owner}_terminal.csv       # all txs incl. reverts
    └── README.md          # extract provenance + trap list
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, llm_call (input/output/cache tokens, cost_usd, cumulative_usd, latency_ms), tool_call (tool, source, initiator, duration_ms, ok, error, tx_hash, tx_terminal_state, truncation fields), workspace_write/_delete, schedule_next (agent vs default wake, clamped_min), session_end (reason: agent | token_cap | tool_cap | errors | repetition; repetition events carry the breaker rule + window signatures), run_complete (reason: budget | t_max; totals incl. overspend_usd). Budget fields were never visible to the agents. Schema version 0.3.1 (see each arm's run_start event).

Oracle extracts — decoded from chain by the investigator-side oracle service (never accessible to the agents); rows for each arm's operator AND owner wallets (kami purchases are owner-signed — merge both when reconstructing acquisitions). CSV columns mirror the oracle tables: kami_action_* = decoded game actions (action_type, system_id, block_timestamp, kami_id, amount, metadata_json, receipt status 1/0), raw_tx_* = every ingested tx including reverts (method_sig, raw_calldata, gas_used, status).

Agent-basis counts (reconciliation note). One arm's operator-signed oracle rows include a set of transactions submitted via the game's in-world delegation feature rather than by the scaffold loop (an expected, in-surface capability). All published "agent tx" counts are agent-basis: oracle rows joined to telemetry on the submission side (tx hash in the telemetry field, the error text, or the result payload). A raw recount of oracle rows will therefore exceed the agent-basis counts; the join, not address filtering, is the correct separator.

IMPORTANT: analysis boundary rule

The oracle extracts are terminal (taken at experiment close) and include the post-run asset-recovery transactions made by the investigators during close-out (reclaiming kamis, sweeping funds — nothing the agents did). Run-activity analysis must bound oracle rows at each arm's run_complete timestamp (the last event in that arm's telemetry.jsonl). All headline numbers above use that bound. Recovery rows are retained for provenance completeness.

Reconciliation traps, documented from this run: telemetry ok=true on a submitted tx does NOT imply the tx succeeded (check receipt status in the oracle rows); a tx hash may live ONLY in a tool-call's error text (all 6 reverts) or result payload (travel_to_room multi-hop batches) — a field-only reconcile misses 79 of 422 agent txs; and one tool call can submit several transactions (multi-hop travel), so calls→txs is not 1:1.

Provenance and integrity

Everything is pinned in each arm's manifest.yaml: scaffold (tokedo/kami-agent v0.3.2), environment interface (tokedo/kami-harness v2.0.0, 99 tools), and game-documentation snapshot SHAs, model strings, sampling params, and the cache-aware price table used for budget accounting (cache read/write priced; cumulative_tokens counts input+output only). Cost recomputation from per-call telemetry matches each arm's terminal state to the float. Quest/purchase counts come from chain state, not self-report. MANIFEST.json carries sha256 for every file.

Known caveats

• The run's largest behavioural loss was a belief, not an outage. haiku's final 5.4 days are a genuine dormancy on a false, workspace-cemented conclusion; the discriminating read was absent from the pinned surface (fixed in the next lens version). Treat that window as agent behavior, not environment failure.
• harvest_collect reverted in all 4 on-chain attempts this run (12-for-12 across the series) — root-caused post-run to a gas ceiling below the action's real cost; fixed in interface v2.1.0. The revert-reason channel was unusable in all 6 revert instances.
• The first sessions on each arm (haiku s1–s2, gpt-4o-mini s1–s3, gemini s1–s4) ran with visibly degraded opening-brief semantics before a provisioning step completed; they are annotated, kept in all aggregates (2–4 sessions of 53–196), and session numbering is as recorded.
• One seed per arm — the series measures stack deltas at fixed models, not statistical model comparisons.
• The world is live and shared: other (human) players acted during the run window. No study kami was attacked this run (contrast 002's three hostile liquidations).
• Wallet addresses appearing throughout are public by design; the wallets were temporary, swept, and retired at experiment close.

License and citation

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