Notes / Audits and snapshots

Tempo -- Morpho Vault V2 target type (rule R8), 3 targets scored, 2026-09-18

Audits and snapshots2026-09-18Chain: TempoLinked: Sentora pathUSD (Morpho Vault V2), Unnamed pathUSD Vault V2 feeding Sentora, Tempo Earn (Morpho Vault V2, Gauntlet-curated)

Source: chains/tempo/data/scored_targets_2026-09-18-vault-v2.md in the repository, rendered here without changes. Source file

Closes chains/tempo/data/scouted_targets_2026-09-17-run2.md section 8's open point 4 for Morpho Vault V2 ("Morpho Vault V2 and market oracles need a written target type"). Unlike the same day's other R2 extension (4 controller shapes, data/scored_targets_2026-09-18-controller-types.md), this is a genuinely NEW target type (rule R8, METHODOLOGY.md section 4.1) -- Vault V2's owner/curator were already resolvable by the existing classify()/weakest() (rules R2/R3) without any change, since both are plain addresses (a Safe or an EOA on every live target found so far). What did not exist before this pass was: (a) the decision that the root-control set for a vault is {owner, curator}, not just one of them, and not the allocator/sentinel roles either; and (b) a delay dimension based on Vault V2's own per-selector timelock(bytes4)/abdicated(bytes4) mappings rather than a single getMinDelay().

Reference implementation: score_vault_v2() / _vault_v2_timelock() / VAULT_V2_TARGETS / VAULT_V2_FUND_DESTINATION_SELECTORS, all in chains/tempo/scripts/methodology_test.py; wired into score_all() in chains/tempo/scorers.py.

Rule (METHODOLOGY.md section 4.1, R8)#

• Root-control set: {owner, curator}. Both are public state-variable getters on morpho-org/vault-v2 src/VaultV2.sol (confirmed against real source, tag matches the live bytecode's dispatcher for every selector used below). Allocators and sentinels are excluded from the NUMERIC set -- both are real but bounded, non-originating powers (confirmed against real source: allocate/deallocate can only move funds among adapters the curator has ALREADY approved, within caps the curator has ALREADY set; the sentinel role can only revoke a pending submission, decrease a cap, or deallocate -- never add a destination or raise a cap), the same convention R1 already applies to TIP-20's UNPAUSE_ROLE and _classify_timelock applies to EXECUTOR_ROLE/CANCELLER_ROLE.
• Weakest key: weakest({classify(owner), classify(curator)}), rule R3 unchanged.
• adminKeyScore / multisigScore: R4/R5 unchanged, applied to that weakest key.
• timelockScore (R5b, Vault-V2-specific): the minimum timelock(bytes4) over 5 "fund-destination" curator selectors -- addAdapter, removeAdapter, setAdapterRegistry, increaseAbsoluteCap, increaseRelativeCap -- EXCLUDING any selector where abdicated(bytes4) is true. Confirmed against real source: abdicate() sets abdicated[selector] = true permanently, with no reverse function anywhere in the contract, so an abdicated selector is not merely delayed, it is permanently closed -- a STRONGER guarantee than any finite delay. Rather than invent a new score band, a fully-abdicated set is mapped to a 7-day delay, i.e. R5b's own existing top band (timelockScore = 80), reusing the scale instead of adding a special case. Selectors are computed from the real function signatures (selector("addAdapter(address)") etc.), never hardcoded -- selector("addAdapter(address)") independently reproduces 0x60d54d41, the exact value scouted_targets_2026-09-17-run2.md's own verification command already used.
• oracleAuthorityScore: 100 (repo "not applicable" convention -- a vault reads no price itself; the market oracle it lends against, T5, is a separate, not-yet-wired target, see "Not yet promoted" below).

Targets scored#

All addresses, thresholds, owner sets and per-selector timelock/abdicated reads above were read live against https://rpc.tempo.xyz on 2026-09-18 (both via cast call and via this project's own call2() dual-endpoint cross-check inside score_vault_v2()), and match scouted_targets_2026-09-17-run2.md section 3/5's own live tables exactly (same owner/curator addresses and thresholds, same per-selector delay/abdicated values for every one of the 5 selectors on all 3 vaults).

Per-selector reads (live, both endpoints agree)#

Scores#

Cross-check, not a coincidence: T1 and T3 independently reproduce scouted_targets_2026-09-17-run2.md section 4's own "resolved reading (proposal)" column EXACTLY -- "T1 curator 1-of-1 + 3 days = 10 / 15 / 60 / 27" and "T3 curator 3-of-7 + 3 days = 65 / 56 / 60 / 61" -- even though that column's numbers were hand-derived by a DIFFERENT research pass and this pass's score_vault_v2() re-derives them independently from a live classify() call, not by copying the target row. T2 has no example in that column to compare against (the proposal table only worked through T1 and T3); its 10/15/0/9 is this pass's own first reading, disclosed as such rather than presented as a reproduction.

crossExposureScore for these 3 targets is computed the same way as every other Tempo target once they run inside the same score_all() batch as the chain baseline and the 9 TIP-20s (rule R6, shared-root-signer count across the WHOLE tracked set) -- not reproduced here as a standalone number, since it depends on the other 10 targets' signer sets too.

Not yet promoted / disclosed limitations#

• Only T1-T3 (the 3 Morpho Vault V2 instances) are wired into score_all() by this pass. T4 (Morpho Blue core, owner-only fee/IRM-allowlist authority, a DIFFERENT shape with no curator/timelock concept at all) and T5 (the cbBTC/pathUSD market's RedStone oracle feeds, a plain EIP-1967 proxy-admin shape already resolvable by unmodified R2, not R8) are still open, exactly as deploy/README.md's own "Not yet promoted" section already disclosed before this pass -- not silently folded into this rule.
• The allocator and sentinel addresses on T1/T2/T3 are not individually enumerated or scored here (no getter list; would need a SetIsAllocator/SetIsSentinel event-log replay analogous to _role_members_via_logs). Their power is bounded (see the rule statement above) but their identity is not recorded in notes this pass -- a real, disclosed gap, not a silent one.

Not an audit, not investment advice. All notes · Back to the dashboard.