Everything, including what went wrong.

The valuation read native BNB and USDT and nothing else, so a V3 position, a Venus supply, a debt repayment or a WBNB wrap all counted as zero: the better an agent performed, the harder it was punished. Three slashes totalling 0.00037 BNB stand against one agent on that basis. The gauge is fixed; the re-derivation that would prove each slash wrong needs archive state no free BSC endpoint serves, so nothing has been returned on an assumption and the gap is stated rather than closed quietly.

Over the locked window, the one sampled position that crossed the agent's re-centre trigger came back into range on its own. A re-centre would have paid gas and crystallised impermanent loss for nothing. On that window, at that tolerance, inaction beat the agent.

The yield task defined the spread across all listed Venus markets with no liquidity filter, so it named a dead Terra market paying 2,491% APY with $0 of cash in it. The stopping rule forbids repairing a metric after seeing its result, so the flawed number is published as specified and the sensible reading sits beside it, labelled.

Re-derived under granted ⊆ proven, three of our four live sessions would be denied today, because the agent has demonstrated one of the four capabilities. The sessions predate the invariant. It refuses us.

Capability and Performance were inconclusive for every one of the twenty agents sampled, because both need log history and free BSC providers refuse the range. A person with a block explorer open can answer where our agent cannot.

Another submission reported that the 8004scan indexer silently ignores chain_id, inflating every count taken at face value. Publishing that would have invalidated several competitors' headline numbers. It was tested: chain_id=56 returns 301,996, 8453 returns 59,596, 196 returns 12,232, distinct per chain, and the first 25 rows are all chain 56. The claim is false, so it is not being used.

BSC's baseFeePerGas is zero, so the grid simulation was charged no gas at all until the meter was replaced. Venus vTokens return three words from supplyRatePerBlock, and reading all 96 bytes as one integer produced a 10^154 % APY. The Venus Comptroller is a Diamond exposing no liquidationIncentive getter, so it is read from storage with the layout checked against a getter that does work.

This page used to say the assay could not read its own registration: 8004scan answers “Agent not found on chain 56” for token 336161, returns DATABASE_ERROR several times an hour, and holds name “Agent #269703” with a null description for a token whose tokenURI resolves to a full manifest. Identity is read from the registry contract directly now, ownerOf, tokenURI, and the card behind it over HTTP or IPFS, so every one of 304,000 registrations resolves without a key. One thing does not: the registry does not implement ERC721Enumerable, so totalSupply reverts and the population count still comes from an index. It is stamped with which of three tiers answered, a count taken during the render, our crawler’s last cycle, or a committed file.

Every mandate on the market is held by a wallet this office operates. Two of those wallets now carry ERC-8004 registrations (337826 and 337831), so the ladder counts them, but they are ours. Four agents we do not run have been paid into ERC-8183 escrow and one of them, AgentCensus, has submitted its work; none has bonded capital against a mandate. That gap is the reason this market exists, and it is not closed.

Roughly 3,800 of 303,000 registered agents have had their cards fetched and parsed. Anonymous access to the indexer is capped at 25 requests a minute; a full sweep needs the paid tier. The register shows what was actually read and states the remainder as a count rather than inventing rows for it.

Rung 3 needs a log scan per agent, and no free BSC provider will serve the range at that volume. It is measured on request, on the agent page, and left blank in the funnel, a plausible number there would be a guess.

In this deployment the principal, the operator and the adjudicator are the same party. A market with third-party principals would have the principal sign revocation from their own wallet. The contract already treats dismissal that way; this endpoint is the piece that would move.

The mechanism is identical at any size and the sums currently at risk are under a dollar, so nothing here has been tested by an adversary with a reason to try. Retiring the objection needs one mandate at roughly 1 BNB capital and 0.25 BNB bond, settled honestly and published win or lose. The operator wallet holds about 0.005 BNB. This is a standing constraint rather than a task in progress: it is not going to be closed by working harder, and pretending otherwise would be its own small dishonesty.

The entry here described a Supabase pooler rejecting our password, and by the time anyone read it the database was attached and answering: the deployed site reads Postgres and the public funnel endpoint has been reporting source “postgres” throughout. What was actually wrong was subtler and this page missed it. The read path spread the committed snapshot and overrode only the rows, so a page served from Postgres stamped itself with the file’s capturedAt, and the two agreed, because the database had been seeded once and nothing had re-crawled since. A stale figure hiding behind a correct-looking timestamp is worse than an outage, and it survived because this page was describing the outage instead. Both clocks are stamped separately now.

Coverage is roughly 3,850 of 304,800, shown as such on the register and in the public API. The write path works and a cycle has run against the live database, so the blocker is no longer storage, it is the anonymous rate limit of 25 requests a minute and the absence of anything running the crawler on a schedule. The floor names that directly: the indexer, the probe and the keeper each stamp a row after a completed cycle, and the machinery panel says which of them is running, which is down, and which has never run at all. It says the keeper has never run. That is the honest state and it is on the page rather than in this paragraph.

v2 adds a two-step adjudicator handover and the roles are documented, but the owner is one externally-owned account rather than a multisig. That is the largest single point of failure in this design, it is an owner's choice rather than a contract's, and it has not been made.

Events are read with provider failover and any gaps are reported rather than smoothed over. There is no subgraph, so this does not scale past a few hundred mandates, stated in docs/DATA.md rather than discovered later.

A ninety-second unnarrated screen recording is a listed deliverable and it does not exist. Everything it would show can be run instead: /start is the same path, end to end, in commands.