Files
proofs/tla/MC_VaultSync.tla
54c7e9ac87 tla: make consensus/chain specs actually model-check (TLC) — fix config bug + spec defects
TLC had never verified anything: every MC config assigned constants with `=`
(TLC model-value) instead of `<-` (operator substitution), so e.g. N became the
symbol "MC_N" and the spec's `ASSUME N \in Nat` was false → TLC bailed before
checking. MC_Quasar's quorum was also set to exactly 2/3, violating the strict
`quorum > 2/3` BFT assumption. Fixed both; all 13 MC specs now run.

Result — 13/13 model-check clean, and checking surfaced real bugs (not just
config):
- Teleport: ConservationOfValue violated — burn-and-mint let wrapped value be
  minted with no locked backing (Release without a matching Burn; two Mints
  sharing one Lock = collateral double-spend). Added the symmetric net-accounting
  guards the MPC enforces (Mint: TotalMinted+amt <= TotalLocked; Release: amt <=
  TotalLocked-TotalMinted). [CROSS-CHAIN SOLVENCY FIX]
- Orderbook: `!.active = (@ + fillAmt < ...)` bound @ to the Bool .active →
  TRUE+int cast crash; fixed to reference the order's filled amount.
- ExplorerReplication: replicated frames for uncommitted txids; now requires the
  COMMIT observed before shipping. Also `1..n \X P` precedence fix.
- LiquidStaking: \div-by-zero on a fully-slashed vault (price=0); guarded.
- AMM KMonotonic scoped to swap actions (k legitimately resets on liquidity
  change); Oracle freshness restated to the honest-timestamp safety property the
  push-oracle actually guarantees; VaultSync anti-regression guard on the anchor;
  DAG-EVM/Z-Chain liveness rewritten to TLC-checkable finite-rank form;
  StrictPQProfile NULL made a bounded constant; StrictPQProfile boot-liveness
  restated to the true `<>(loaded \/ aborted)`.

Co-authored-by: Hanzo Dev <[email protected]>
2026-06-10 22:34:27 -07:00

19 lines
623 B
Plaintext

---- MODULE MC_VaultSync ----
\* Model checking configuration for VaultSync
EXTENDS VaultSync
\* Small model for exhaustive checking.
\* Two keys and two values suffice: Convergence only distinguishes whether
\* two devices hold equal vs unequal localState, so >=2 distinct values
\* (to express divergence) over >=2 keys fully exercises it. Larger
\* key/value domains only multiply the localState/oplog space without
\* reaching new control behaviour, so they are bounded here to keep the
\* model exhaustively checkable.
MC_Devices == {"phone", "laptop"}
MC_MaxOps == 3
MC_Keys == {"a", "b"}
MC_Values == {1, 2}
====