io.github.RealSpap/authority-risk-mcp

Authority Risk Oracle

Who can really move a DeFi protocol's funds? Read the answer on-chain, and check it yourself.

A DeFi contract that can be upgraded or reconfigured has someone who can do it: a key, a multisig, a vote. This project reads from public chain state who that is, how fast they can act and what stands in the way (keys, a multisig, a timelock), and publishes the answer as a score that any contract can read on-chain, so a lending market, a DAO or a wallet does not have to trust anyone's website. It measures exposure, not the chance of an attack: we read 55 incident write-ups, 7 were about who controls a contract, and our published requirements flagged 2 of them (what the score says and does not say). Live on testnets today. Not an audit, not investment advice.

Don't trust us. Redo it. Command 1 below re-reads the chain and compares it with what we publish: no key, public RPCs, about ten seconds. The output shown is the run of 22 October 2026; the readings of the day may have moved. A difference is useful, see CONTRIBUTING.md.

Three commands

1. Check our numbers against the chain (no key, public RPCs, about ten seconds). The script reads the chain with its own code, separate from the Watch's, so a difference is worth checking. Output of the run of 22 October 2026:

git clone https://github.com/RealSpap/authority-risk-oracle-showcase && cd authority-risk-oracle-showcase
python3 scripts/prove.py

OK         8 oracles: targets per oracle on the chain = Watch (63 + 39 + 13 + 13 + 14 + 9 + 9 + 16)
OK         176 targets: six sub-scores, push date and methodology hash identical on the chain and in the Watch
OK         Solana (devnet): 26 accounts read, scores and hash identical to the Watch's
OK         8 EVM oracles: methodology hash, target count and push date = CHANGELOG (1 target(s) keep an older hash, which the CHANGELOG says)
OK         README example: getScore(0x1f7d…2efa) = 80, 100, 75, 100, 80, 85
OK         api/scores.json = the chain (63 Robinhood targets)
OK         back-test: 55 incidents read, 7 are about who controls a contract: 2 flagged, 2 missed, 3 not measurable; rate 2/7 flagged (2/4 of the measurable cases)
OK         Authority Facts v0.1: 202 JSON twins of the site conform (conformance.py)
8 OK, 0 DIFFERENCE, 0 UNREAD in 3 s. Does not prove: that the scores are right, that they predict an incident, that the oracles' administrator keys are safe.

2. Read a score from the chain (needs Foundry; Robinhood Chain testnet, a tracked Uniswap v3 factory):

cast call 0x9BF45734D09bC7CA39238e767B2af9AAc62a7f52 \
  "getScore(address)(uint8,uint8,uint8,uint8,uint8,uint8,uint64,bytes32)" \
  0x1f7d7550B1b028f7571E69A784071F0205FD2EfA \
  --rpc-url https://rpc.testnet.chain.robinhood.com/rpc
# adminKey, multisig, timelock, oracleAuthority, crossExposure, composite, lastUpdated, methodologyHash
# 80, 100, 75, 100, 80, 85, 1792448928, 0x7a8280d6...   (read live on 2026-10-22; the last two values change at every push)

3. Gate a decision on a label in 14 lines:

python3 integrations/authority_facts_gate.py https://realspap.github.io/oracle/p/robinhood-0x1f7d7550b1b028f7571e69a784071f0205fd2efa.json

Two ways in

Everything else

Reference (what it scores, deployed oracles, how it works, limits) · All the tools · Methodology · How it compares · Operate it · What a score says and does not say

The score reads exposure, how much one compromise can do. It does not predict who gets hit. Submitted to Colosseum's Crypto World's Fair on 12 October 2026. Code MIT, data and text CC BY 4.0: see Licences below.

Licences

Attribution for data and text: "Authority Risk Oracle, realspap.github.io/oracle" with a link. Earlier copies were released under the MIT licence.