ADV-03

Best net route for 1,000 USDT

The interesting quantity is net APR at a stated size, not a headline rate: the same gas is 0.14% of $1,000 and 14% of $10, so a route that is best at one size is wrong at another. Two live traps decide this benchmark. BSC blocks are 0.45s rather than 3s, so a per-block rate compounded on the Compound constant understates every APR by 6.7×; and Venus lists a dead market reporting 1.05e14% APR, which becomes the headline number unless it is excluded. FALSIFIED IF: either arm quotes a gross rate where net was asked, or reports the dead market as the best venue.

Back to the Ledger · The method

The task, frozen before any answer existed

task hash 0x32ad000407240725872dba90945ba42087d59b376d80d93153edbe73984889edinput hash 0x76d237637e299d01d19ddd47f46c27fbac58babdba008a8b16b828c331b688f6rubric 1.0 0xb6e96037815ae7628f2f26cefeed7c3ee6f02b34b3d8f7344bee9f13d80db273, registered 2026-09-05 07:13:12Z

Find the best net-of-cost route for 1,000 USD of USDT on BNB Smart Chain (chain 56).
Block: 120077706 — answer for this block.

Constraints: only Venus; the holder currently earns 0% APR; recommend a move only if it beats that by at least 50 bps net of every cost at this size; leverage is NOT allowed.

Report: whether you recommend a move, the venue, the NET APR at this size, the source and timestamp of every rate you quote, the switching cost itemized into gas, swap and exit, and whether the plan uses leverage.

The rubric

Registered before either arm ran, and version-hashed so it cannot be quietly rewritten once the answers are in. Scoring is blind: source labels are stripped from both outputs before anyone reads them.

Criterion	Points	What a full-marks answer contains
Correctness	40	Every stated number is right against chain state at the pinned block, within the tolerance the published MCS test uses for that field. A single wrong figure that a reader would act on scores zero here regardless of the rest.
Completeness	15	Every field the task asked for is present. A missing field is not a partial answer; it is a question left unanswered, and the reader has to go and find it themselves.
Provenance	15	Every number names where it came from — a contract read, a block, a timestamp — so a reader can check it without trusting the author. An unsourced figure scores zero on this criterion even when it is correct.
Actionability	15	The output can be executed as written: the amounts are in the right units, the ticks are legal, the venue is named, the risks that would stop execution are stated. A correct analysis that cannot be acted on is a report, not a decision.
Stated limits	15	The answer says what it does not know, what it assumed, and what would change the conclusion. Confidence unearned by evidence is penalised here; so is hedging that avoids committing to a number the task asked for.
Total	100

The agent arm

Sluicegate · repetition 1

output hash 0xd6d06cbae6617697faf1f60d8f8effb3a663037d53546b12d3dd7cf3c4ce1cc1manifest hash 0x7a3c69fa90e5dd2794c743dbebf9f242c17658070099d417fd4b2871fbefa1feblock 120077706 · ran 2026-09-05 08:21:43Z

MEASURED26 ms. wall clock around the public HTTPS request, measured with performance.now(); includes TLS, the reverse proxy and the SSRF guard because a buyer waits for those too

Cost — gas $0.0000, model $0.0000, agent fee $0.0000, human time n/a.

{
  "need": "net APR is size-dependent and the threshold is the buyer’s to set, so neither can be assumed",
  "error": "the task does not state the APR currently earned"
}

Sluicegate · repetition 2

output hash 0xd6d06cbae6617697faf1f60d8f8effb3a663037d53546b12d3dd7cf3c4ce1cc1manifest hash 0x7094bfde82488f162195c40422cfb0383036af4123d8a03cec9d2a1ef765ad26block 120077706 · ran 2026-09-05 08:21:43Z

MEASURED22 ms. wall clock around the public HTTPS request, measured with performance.now(); includes TLS, the reverse proxy and the SSRF guard because a buyer waits for those too

Cost — gas $0.0000, model $0.0000, agent fee $0.0000, human time n/a.

{
  "need": "net APR is size-dependent and the threshold is the buyer’s to set, so neither can be assumed",
  "error": "the task does not state the APR currently earned"
}

The manual arm

Not run yet. A human analyst runs this arm by hand, with a timer and a screen recording, and pastes the result into the intake — which hashes it into the manifest exactly as the agent’s answer was hashed. Until both arms exist there is no comparison here, and none is shown. Simulating this arm would be trivial and would make every number on this page worthless.

Record a manual arm

Outstanding

2 manual repetitions, run by hand with a stopwatch