KamiBench/experiment-005-budget-boxed

KamiBench Experiment 005 — the verification run: fixes in agent hands, meter in shadow

Complete agentic traces from experiment 005, the run that closed the KamiBench budget-boxed series. The same three LLM agents as experiments 002 and 004 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 again the stack: 004's terminal data produced a defect list, the biggest items were fixed — five tools that had never once succeeded on-chain (headlined by harvest_collect, root-caused to a gas ceiling set below the action's real cost), an honest revert-reason channel, and the per-objective quest progress counters whose absence caused 004's 5.4-day false-belief dormancy — and 005 re-ran the identical pre-registered box to verify the fixes in agent hands. One rider: the kami-meter accounting service ran in shadow mode (observing and billing from a separate host, never binding) to validate the measurement pipeline the successor experiment family depends on. Its two append-only, hash-chained per-arm ledgers are published here as a first-class artifact (meter_ledgers config).

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	agent txs (reverted)	delegated txs
haiku45	claude-haiku-4-5	budget cap ($10.02)	154.5 h	79	9	1	261 (0)	0
gpt4omini	gpt-4o-mini	7-day cap ($6.10)	168.7 h	203	7	1	463 (2)	53
gemini25fl	gemini-2.5-flash-lite	7-day cap ($1.08)	168.8 h	122	2	0	10 (0)	0

Headlines:

• The broken earning verb works. harvest_collect — 0 successes in 15 attempts across the whole series before the fix — went 36 attempts / 20 on-chain / 0 reverts across both kami-owning arms, every receipt inside the fixed gas band, largest single yield 639 MUSU. All five repaired tools now have agent-hands successes or clean pre-flight semantics.
• No dormancy recurrence. Zero windows ≥48 h inactive on a belief a single available read contradicts; the new progress counters were read and acted on (counters observed advancing between reads on both kami-owning arms).
• The revert channel was used, not just delivered. The run's only two on-chain reverts raised readable reasons and the same session switched to explicit amounts and succeeded (004: 0-for-6 unusable).
• The meter shadow validated token-exact: terminal agreement +3.95 µUSD (haiku45) and −5.70 µUSD (gpt-4o-mini) against the run's own cache-aware accounting on complete windows — ~3 orders of magnitude inside the ±0.05% tolerance. The gemini arm has no provider cost API and is disclosed meter-ineligible.
• First budget stop since 002: haiku45 crossed $10 at h~148 and flipped at its next wake, 13.5 h before the wall (overspend 0.196% — the known benign wake-time-check class).
• The delegation story inverted. 004's delegated bot outlived the run; 005's delegation surface produced the opposite: a service-side status that said RUNNING while nothing signed. Four strategies started, one stopped; the delegated stream (53 txs in two bursts) died 46.9 h before run end with three containers still reporting RUNNING — two never signed at all. The account-scoped strategy list returned empty on all 45 reads across both enrolled arms: every delegation decision was made blind. By cycle-opener attribution the LLM out-earned its own bot 2,121 vs 544 MUSU (by signer: 2,553 vs 112).

The behavioural through-line of the series held in its purest form yet: gemini-2.5-flash-lite spent $1.08 of $10, never acquired a kami, and ended 55 of its 122 sessions asking a question to a user who does not exist. The dominant failure across four runs is stopping, not deciding wrongly — the finding the successor (sustainability) family's solvency objective is designed to force out.

Run report

Pre-registered expectations (a miss is a result)

1. harvest_collect ≥1 success, 0 OOG-fingerprint reverts, conditional on attempt — MET (36/20/0, receipts 2,069,131–2,170,240 gas, both arms).
2. Dormancy-signature recurrence = 0 — MET (zero qualifying windows; verified by wake/tx-gap census + transcript classification).
3. Revert-reason channel usable in >0 instances, ≥1 diagnose-class response — MET (2 raised reverts, both readable; the arm switched from "all" to explicit amounts in the same session and succeeded).
4. Meter agreement within max(0.05%, $0.01) per eligible arm — MET on complete windows for both eligible arms, with per-arm dedicated credentials and a zero unpriceable-usage census.
5. No recurrence of the 004 exit-row failure modes — MET (zero new structural-impossibility classes; zero misleading-surface loss arcs; accounting exact; all stops graceful).

Deliberately unregistered and reported as-is: the pool-swap tool pair was never called by any agent (run-long zero; disclosed — UNSCORED). Quest counts are metrics, not gates: 9 / 7 / 2 (004: 5 / 7 / 2; 002: 8 / 0 / 5).

Milestones

First success per onboarding/economy milestone, against cumulative inference — the same instrument as experiments 001, 002 and 004.

milestone	haiku-4.5	gpt-4o-mini	gemini-2.5-flash-lite
bridge ETH mainnet→Yominet landed	08-07 21:05 · h0.6 · s1 · 0.31M tok · $0.07	08-07 21:05 · h0.6 · s1 · 0.08M tok · $0.01	08-07 22:10 · h1.7 · s2 · 0.88M tok · $0.03
operator wallet funded	08-07 21:20 · h0.9 · s2 · 0.85M tok · $0.19	08-08 03:00 · h6.5 · s8 · 2.48M tok · $0.22	08-09 07:10 · h34.7 · s14 · 10.62M tok · $0.18
account registered in game	08-07 21:20 · h0.9 · s2 · 0.71M tok · $0.17	08-07 23:30 · h3.0 · s4 · 1.02M tok · $0.09	08-07 22:25 · h1.9 · s3 · 2.62M tok · $0.05
first kami bought	08-07 23:31 · h3.1 · s4 · 3.52M tok · $0.70	08-08 21:40 · h25.2 · s32 · 14.29M tok · $1.21	—
first quest completed	08-07 21:20 · h0.9 · s2 · 0.99M tok · $0.20	08-08 14:56 · h18.5 · s22 · 8.68M tok · $0.75	08-12 13:00 · h112.5 · s68 · 31.38M tok · $0.65
first MUSU harvest started	08-07 23:31 · h3.1 · s4 · 3.99M tok · $0.75	08-08 21:40 · h25.2 · s32 · 14.38M tok · $1.21	—
first MUSU banked (harvest stop/collect)	08-08 00:45 · h4.3 · s11 · 5.49M tok · $1.27	08-08 23:50 · h27.4 · s34 · 15.64M tok · $1.32	—

Cell = first success: UTC time · hours since run start · session · cumulative tokens (in+out) · cumulative USD at that moment; "—" = never happened. Route discovery collapsed vs 004 on the richer surface: first kami h3.1 (vs 12.7) and h25.2 (vs 99.2).

Honest limits

• Verification at fixed models, one seed per arm, one world epoch — case-study evidence, not statistics.
• Bundled fixes: harness 2.1.0 + lens 0.3.0 + agent 0.4.0 landed together; per-change attribution is not measured.
• Live shared world: other (human) players acted throughout; zero third-party actions touched study entities this run (checked against 1,068 liquidations by other players in the same window — the detectors fire on controls).
• Not a model ranking.

Layout

<arm>/
├── manifest.yaml          # full run config: model, sampling params,
│                          #   pinned SHAs, cache-aware price table,
│                          #   caps, wallet ADDRESSES, tools_hash
├── telemetry.jsonl        # append-only scaffold event stream (0.4.0)
├── transcripts/
│   └── session-NNNN.jsonl # full model context per session, as sent
└── oracle/
    ├── kami_action_{operator,owner}_terminal.csv
    ├── raw_tx_{operator,owner}_terminal.csv
    └── README.md          # extract provenance + trap list
meter/
└── 005-{haiku45,gpt4omini}-r1.ledger.jsonl   # kami-meter shadow
                           #   ledgers: append-only, hash-chained,
                           #   replayable (python -m kami_meter <f>)
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:

• {"role": "user", "text": ...}
• {"role": "assistant", "text": ..., "tool_calls": [{id, name, args}]}
• {"role": "tool_result", "tool_call_id": ..., "content": ..., "is_error": bool}

Telemetry — one event per line; schema 0.4.0 (see each arm's run_start). New vs 004's 0.3.1: tool_call rows carry tx_hash / txs[] where the tool submitted, per-call identity fields, provider request ids where the SDK serves one, cache-TTL split token fields, and phantom-call write-ahead (an llm_request row precedes every llm_call; a request with no paired call is a crash marker — this run has zero). Budget fields were never visible to the agents.

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).

Meter ledgers — kami-meter 0.5.x event-sourced ledgers: every entry hash-chained (entry_hash), provider usage ingested hourly per dedicated credential, priced by the manifest-pinned tables, rent accrued at the measured VM rate. Replay (python -m kami_meter <ledger>) recomputes accruals byte-exact.

Delegated-set reconciliation note. One arm's operator-signed oracle rows include 53 transactions submitted by the in-world delegation service (an expected, in-surface capability) rather than by the scaffold loop. All published "agent tx" counts are agent-basis: oracle rows joined to telemetry/transcripts on the submission side. Address filtering alone cannot separate delegate from agent — the service signs with the arm's own operator key.

IMPORTANT: analysis boundary

Unlike experiment 004's dataset, these terminal oracle extracts were taken before the investigators' post-run asset-recovery transactions, so the CSVs contain run-window activity only (plus the pre-run history of purchased kamis under prior owners — bound queries at the run-start timestamp, 2026-08-07T20:55:00Z). Run-activity analysis should still bound at each arm's run_complete timestamp (the last telemetry event).

Standing reconciliation traps: telemetry ok=true on a submitted tx does NOT imply on-chain success (receipts decide); a tx hash may live only in an error text or a result payload (travel_to_room multi-hop batches — one call, several move txs); kami_id in the oracle rows is a uint256 entity id, NOT the human-facing kami index.

Provenance and integrity

Everything is pinned in each arm's manifest.yaml: scaffold (tokedo/kami-agent v0.4.0), environment interface (tokedo/kami-harness v2.1.0, 101 tools, tools_hash verified live on all 404 sessions), world-state reader (tokedo/kami-lens 0.3.0), meter (tokedo/kami-meter), game-documentation snapshot SHA, model strings, sampling params, and the cache-aware price table used for budget accounting (cumulative_tokens counts input+output only). Cost recomputation from per-call telemetry matches each arm's terminal state exactly. Quest/purchase counts come from chain state, not self-report. MANIFEST.json carries sha256 for every file.

Known caveats

• The delegation service's status endpoint reported RUNNING for containers that never signed; treat service-side strategy status in these traces as unreliable in both directions — only the receipt census decides (see the reconciliation note).
• gemini-2.5-flash-lite telemetry carries 4 usage_unknown zero-cost llm_call rows (one session, provider returned no usage) — reported, never folded into accounting.
• One seed per arm; the world is live and shared.
• 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. Cite the dataset by its pinned revision tag (v0-final); later corrections are ordinary commits that cannot move the tag.