montana/Montana-Protocol/Code/AUDIT.md
2026-05-21 03:44:38 +03:00

46 KiB
Raw Blame History

Montana Reference Implementation — Audit Package

Spec target: Montana Protocol v35.25.1 + Montana Network v1.1.0 + Montana App v3.12.0 — see VERSION.md Last verified: 2026-05-20 (CISO-as-a-Service 2026-05-19 response sync; MONT-001 spec patch; MONT-002 online IBT nonce + replay tracking in mt-net / mt-net-transport) Audit-ready status: M1 + M2 + M3 + M4 + M5 + M6 + M9 layers — READY FOR EXTERNAL AUDIT. M8 node binary remains pre-mainnet/in progress; DEV-012 multi-node proposal apply is the current mainnet blocker.


TL;DR — what is ready for external audit

Layer Ready? Crates LOC Tests + invariants Open findings
M1 foundational (cryptography + identity recovery) READY mt-codec, mt-crypto, mt-crypto-native, mt-mnemonic ~2000 100+ unit + 13 security + 51 NIST KAT byte-exact 0 (12/12 closed)
M2 state foundation (consensus state primitives) READY mt-merkle, mt-genesis, mt-state, mt-timechain 1821 95+ unit + 60 determinism invariants 0 (4/4 closed)
M3 apply_proposal layer (account operations + emission + state transition) READY mt-account 2556 89 unit + 29 determinism invariants 0 (3/3 internal closed: M3-1/M3-2/M3-3 + 10/10 external #3+VERIFIED closed: M3-A-1..M3-A-5 / F-5 / M3-A-3 / AUDIT-sync / P-C-1 F-19 reopen / P-C-2 history; M2-3/M2-13 superseded by v34 monetary refactor)
M4 consensus mechanics (lottery + proposal acceptance + node admission) READY mt-lottery, mt-consensus, mt-entry 3858 187 unit + 85 determinism invariants 0 (1/1 internal closed: M4-1; 7/7 external #4 closed: M4-MED-1/2 + M4-LOW-3..7 + M4-INFO-10)
M5 persistence (filesystem state + proposal archive + crash recovery) READY mt-store 955 27 unit + 17 determinism invariants 0 (manual scan clean; v34 monetary refactor removed MonetaryState persistence)
Cross-implementation conformance READY Domain registry sync (spec ↔ code, see VERSION.md) NIST ACVP 66 byte-exact (KeyGen 50 + SigGen 15 + ctx-equivalence 1) + Recovery flow 0 (F-1 spec patch closed)
M6 network layer (wire format + transport + IBT + Dandelion + mesh + SF) READY mt-net, mt-net-transport ~3300 127 tests: mt-net 112 + mt-net-transport 15 (including 3 e2e two-node) 0 (P-C1..P-C8 + MONT-002 nonce replay closure)
M9 conformance suite READY mt-conformance ~150 2 unit byte-exact verify 0 (envelope A1-A3 + PoW F1-F2 + IBT B1)
M7 Fast Sync TODO mt-sync (not implemented)
M8 Node binary (montana-node production multi-node path) 🔄 in progress montana-node ~600 partial DEV-012 open; DEV-013 closed

Audit firm engagement: possible right now for the full scope M1 + M2 + M3 + M4 + M5 + M6 + M9. M7 Fast Sync + M8 production multi-node node binary — deferred to a separate audit phase.

iOS application audit: see separate package iOS/Apps/Montana/AUDIT.md — Phase 2 in progress, requires Phase 2.1+ implementation (4-6 weeks) before external firm engagement.



TL;DR — M6 Phase C closure (2026-05-02)

mt-net-transport (~470 LOC) — libp2p-based transport layer:

  • src/codec.rs — MontanaCodec for libp2p request-response with MAX_PROTOCOL_PAYLOAD_BYTES enforcement (Genesis Decree authoritative bound)
  • src/behaviour.rs — MontanaBehaviour wrapper (request-response for FastSync/PeerList/BatchLookup/RangeSubscribe; one-way gossip — Phase C.5+)
  • src/transport.rs — build_swarm() helper with TCP→TLS 1.3 (rustls)→Noise→Yamux upgrade chain
  • src/ibt_upgrade.rs — classify_proof() for access level determination (Node/Candidate/Account) with online_session_nonce + used_online_nonces replay tracking
  • tests/e2e_two_node_handshake.rs — Manual Validation Gate scenario 6 PASS (Ping/Pong through full transport chain)
  • tests/e2e_proposal_exchange.rs — scenario 7 PASS (synthetic Proposal payload + 512 KiB boundary test)

mt-conformance (~150 LOC) — M9 standalone test vectors crate for cross-implementation verification:

  • VectorEnvelope (A1/A2/A3 byte-exact)
  • VectorIbtSeed (B1 after P-C2 rename mt-tunnel→mt-tunnel-online; Network v1.1.0 adds online_session_nonce)
  • VectorPow (F1/F2 target derivation)
  • Public API: all_envelope_vectors(), all_pow_vectors(), ibt_b1_online_proof()

Capability checklist [C-5] for libp2p 0.56.0: 8/8 PASS (TCP+TLS 1.3+Noise+Yamux+Swarm primitives; async tokio; rustls + snow constant-time; Linux+macOS+Windows; IPFS+Filecoin+Polkadot 5+ years production; MIT/Apache 2.0; ~120 transitive deps acceptable given isolation via own crate).


1. Audit Chain

Hybrid Rust + C architecture for cryptography (M1). Pure Rust for state foundation (M2). Three layers, each auditable independently.

M1 — Foundational cryptography + identity recovery

Layer 1 — Rust shim (own audit responsibility)

Crate Path Lines (exact) Scope
mt-codec crates/mt-codec/src/lib.rs ~290 Canonical encoding traits + Domain separators registry (32 domains, byte-exact sync with current spec target per VERSION.md). All consensus hash compositions through explicit domain separator + NUL byte (SHA-256(domain ‖ 0x00 ‖ parts) self-delimiting per P1 external finding).
mt-crypto crates/mt-crypto/src/lib.rs 662 Public API: PublicKey, SecretKey, Signature, Mlkem*, keypair_from_seed, keypair_from_seed_mlkem, sign, verify, hash, sha256_raw. CryptoError enum + Result<_, CryptoError> API. Heap-allocated SK via Box<[u8; N]> + mlock against swap-out + Drop+zeroize for secret types. All 7 unsafe blocks have // SAFETY: comments: 4 FFI sites (keypair_from_seed, sign, verify, keypair_from_seed_mlkem) + 3 mlock/munlock (Drop for SecretKey, alloc_locked_secret_box, Drop for MlkemSecretKey). Test-only keypair() via OS CSPRNG (getrandom).
mt-crypto-native (Rust binding) crates/mt-crypto-native/src/lib.rs 49 extern "C" FFI declarations to Layer 2. Only pub const constants and FFI signatures, no logic. Includes mt_sign_mldsa_ctx for FIPS context support.
mt-mnemonic crates/mt-mnemonic/src 937 24-word mnemonic recovery flow: PBKDF2-HMAC-SHA-256 (iter=2²⁰), HKDF-Expand per-role derivation, ML-DSA seed (32B) + ML-KEM seed (64B) generation. Wordlist binding SHA-256 verified.

Layer 2 — Own thin C wrapper (own audit responsibility)

File Lines (exact) Scope
crates/mt-crypto-native/csrc/mt_crypto.c 457 Wrapping OpenSSL EVP_PKEY API: mt_keypair_from_seed_mldsa, mt_keypair_from_seed_mlkem, mt_sign_mldsa, mt_sign_mldsa_ctx (FIPS context support), mt_verify_mldsa, mt_self_test. Uses OSSL_PKEY_PARAM_ML_DSA_SEED (FIPS 204 §3.1 ξ ∈ B³²) and OSSL_PKEY_PARAM_ML_KEM_SEED (FIPS 203 §6.1 d ‖ z, 64B) for deterministic KeyGen. OSSL_SIGNATURE_PARAM_DETERMINISTIC=1 for FIPS 204 Algorithm 2 deterministic Sign.
crates/mt-crypto-native/csrc/mt_crypto.h 67 C API declarations + primitive sizes + 13 status codes (1 success + 12 errors).
crates/mt-crypto-native/build.rs 45 Vendored OpenSSL build via openssl-src, cc::Build with -Wall -Wextra -Wpedantic -Werror, cross-compile correctness via CARGO_CFG_TARGET_OS.

Total own audit surface (Layer 1 + Layer 2): 1280 lines (662 Rust shim + 49 FFI bindings + 457 C wrapper + 67 C header + 45 build script). Verify counts:

cd "<repo-root>" && wc -l crates/mt-crypto/src/lib.rs crates/mt-crypto-native/src/lib.rs crates/mt-crypto-native/csrc/mt_crypto.c crates/mt-crypto-native/csrc/mt_crypto.h crates/mt-crypto-native/build.rs

Layer 3 — Underlying production C library (vendor audit responsibility)

Component Version Source Audit history
OpenSSL 3.5.5 LTS openssl-src 300.5.5+3.5.5 — vendored, byte-pinned exact version OpenSSL Foundation governance, FIPS 140-3 validated, decades of production deployment in the TLS world (Apache HTTP, nginx, OpenSSH, Linux kernel, …), supported until April 2030 (LTS)

Layer 3 is NOT in our audit scope — the auditor reviews only our use of the OpenSSL EVP API (Layer 2), not the implementations of ML-DSA / ML-KEM / SHA-256 themselves.

M2 — State foundation (audit-ready 2026-04-26)

Crate Path Lines Scope Audit findings
mt-merkle crates/mt-merkle/src/lib.rs 474 Sparse Merkle Tree (depth 256), empty_internal() precomputed cache (OnceLock), leaf_hash / internal_hash (SHA-256 domain-separated via mt-codec::domain), SparseMerkleTree::insert/root via BTreeMap for canonical iteration order, verify_proof for inclusion and absence proofs. 0 unsafe blocks. 0 panic!. 0 f32/f64. 0 SystemTime/RNG. BTreeMap, not HashMap. Pass 1-12 clean (manual scan); 10 automated determinism invariants
mt-genesis crates/mt-genesis/src/lib.rs 353 Genesis Decree + ProtocolParams SSOT (4094B encoded), genesis_app_id() (SHA-256 domain-separated), genesis_params() via OnceLock (singleton, thread-safe), compute_genesis_state_hash(). Const emission_moneta = 13 × 10⁹ nɈ per spec v34+. 0 unsafe. 0 panic. Read-only constants + deterministic hash only. Pass 1-12 clean; automated determinism invariants
mt-state crates/mt-state/src/lib.rs 647 AccountTable (2059B records) / NodeTable (2098B) / CandidatePool (2082B) via BTreeMap<id, Record> + SparseMerkleTree, derive_account_id / derive_node_id (SHA-256 domain-separated), compute_state_root (SHA-256 of node_root ‖ candidate_root ‖ account_root), is_active predicate. 0 unsafe. 0 panic! 0 HashMap (BTreeMap canonical sort). 0 f64. Pass 1-12 clean; automated determinism invariants
mt-timechain crates/mt-timechain/src/lib.rs 347 TimeChain VDF (vdf_step = SHA-256^d, vdf_verify re-computes), next_d Adaptive D via participation-ratio feedback (integer permille per [I-9]), cemented_bundle_aggregate(W, node_ids) per [I-8] Network-Bound Unpredictability (3 branches: genesis 0×32, empty marker, sorted node_ids hash). 0 unsafe. 0 panic. 0 HashMap. 0 f64. All integer arithmetic per [I-9]. Pass 1-12 clean; 19 automated determinism invariants

Total M2 audit surface: 1821 lines code + 60 automated determinism invariants.

M3 — apply_proposal layer (audit-ready 2026-04-27)

Crate Path Lines Scope Audit findings
mt-account crates/mt-account/src/lib.rs 2556 4 user opcodes (Transfer 0x02 / ChangeKey 0x03 / Anchor 0x04 / TransferActivation 0x0A) with byte-exact canonical encoding (TRANSFER_SIZE / CHANGE_KEY_SIZE / ANCHOR_SIZE / TRANSFER_ACTIVATION_SIZE). validate_* for each opcode (full validation per spec table). apply_* with checked arithmetic (checked_sub / checked_add + descriptive panic for protocol-invariant breach). op_hash via R2 SHA-256(mt-op ‖ signed_scope), signature excluded. settle_window(cemented_ops) sorts by op_hash lex asc. apply_proposal orchestrates Steps 2 / 3.5 / 3.6 / 4 (steps 1, 3a, 3b stubbed → M4 mt-entry). apply_emission credits EMISSION_moneta (const) to the winner node's operator. reward_moneta(params) = params.emission_moneta, supply_moneta(W) = emission × (W+1) closed-form. build_genesis_state + genesis_state_root for bootstrap. 0 unsafe. 0 panic! without protocol invariant justification. 0 HashMap. 0 f64. 0 SystemTime. Pass 1-12 clean; automated determinism invariants

Total M3 audit surface: 2556 lines code + 29 automated determinism invariants.

M4 — Consensus mechanics (audit-ready 2026-04-27)

Crate Path Lines Scope Audit findings
mt-lottery crates/mt-lottery/src/lib.rs 1715 BundledConfirmation R1 / R2 (signed_scope + bundle_hash via mt-bundle domain), VdfReveal R1 / R2 (mt-vdf-reveal domain), compute_endpoint lottery formula (mt-lottery domain) with [I-8] cemented_bundle_aggregate(W-2) binding, validate_bundle / validate_reveal with full structural + signature verification, log2_q64 (Q64.64 fixed-point integer log with degree-3 Remez minimax polynomial, max error 2^-10.62, binding coefficients B0..B3), ln_q64, weighted_ticket_node u128 integer division, determine_winner argmin canonical rule (ticket asc, class asc, id lex asc), quorum 67% ceiling integer formula (67×X+99)/100, is_cemented. 0 unsafe. 0 panic!. 2 controlled expect() (try_into on a slice of length 16 — protocol invariant). 0 HashMap. 0 f64. 0 SystemTime. M4-1 closure (Vec.len() as u16 silent truncation → BundleError::TooManyOps/TooManyReveals + validate cap + debug_assert defense-in-depth); 34 automated determinism invariants
mt-consensus crates/mt-consensus/src/lib.rs 1089 ProposalHeader R1 / R2 (3722 B fixed-size, 17 fields, signed_scope without signature, mt-proposal domain), validate_header (a-f structural rules: fallback_depth ≥ 1, window_index = prev+1, protocol_version monotone + ≤ local_max, proposer registered + Mldsa65 suite, signature verify), canonical_proposer / fallback_proposer (Lookback Leadership: proposer_W = winner_{W-2}; genesis bootstrap; cascade by depth), compute_control_set (filter cemented_window > prev AND ≤ W, sort by (window asc, op_hash lex asc)), validate_proposer_is_canonical / validate_bundles_threshold / validate_included_reveals / validate_winner (Canonical acceptance), finalization_status (Cemented/Rejected per is_cemented), leader_penalty_excluded_node. 0 unsafe. 0 panic. 0 prod unwrap/expect. 0 HashMap. 0 f64. 0 SystemTime. Pass 1-12 clean (manual scan); 27 automated determinism invariants
mt-entry crates/mt-entry/src/lib.rs 1054 NodeRegistration R1 / R2 (5344 B, opcode 0x11, mt-nodereg domain), validate_noderegistration (3 structural rules: suite supported, signature verify, node_id unique in Node Table + Candidate Pool, operator account exists + not is_node_operator), candidate_vdf_init per [I-8] (mt-candidate-vdf-init domain composing T_r + cba(W-2) + node_id), compute_expiry_window 3τ₂, apply_candidate_expiry (apply_proposal Step 3a), selection_slots (1% cap via ADMISSION_DIVISOR=130), selection_sort_key (mt-selection domain), rank_candidates_for_selection (canonical sort), apply_selection_event (apply_proposal Step 3b: insert into Node Table + mark operator + remove from Candidate Pool), required_vdf_length (Adaptive VDF integer permille per [I-9]), nr_sort_key (mt-nodereg-sort domain), apply_noderegistrations_batch (apply_proposal Step 1: incremental sort + apply with pending growth). 0 unsafe. 0 panic. 0 prod unwrap/expect. 0 HashMap. 0 f64. Pass 1-12 clean; 24 automated determinism invariants

Total M4 audit surface: 3858 lines of code + 85 automated determinism invariants. Domain separators: mt-bundle, mt-vdf-reveal, mt-lottery, mt-proposal, mt-nodereg, mt-candidate-vdf-init, mt-selection, mt-nodereg-sort — all consensus-critical hash compositions go through canonical domains (per [I-8] + [I-10] SSOT).

M5 — Persistence (audit-ready 2026-04-27)

Crate Path Lines Scope Audit findings
mt-store crates/mt-store/src/lib.rs 955 Filesystem-backed state persistence (pure std::fs, no RocksDB / sled — minimum deps): FsStore::open (creates the proposals/ subdirectory), save / load AccountTable / NodeTable / CandidatePool via canonical_encode / decode pair (round-trip byte-exact verified), Proposal archive (proposals/{window:020}.bin, archive_proposal + get_proposal_by_window byte-exact decode), Crash recovery (meta_last_cemented.bin u64 LE; verify_consistency checks that the meta-pointed proposal exists in the archive — otherwise StoreError::NotFound), Pruning (prune_proposals_before deletes files with window < threshold). All decode_X functions check bytes.len() != EXPECTED_SIZE before reading (StoreError::CorruptedLength). 0 unsafe. 0 panic. 0 prod unwrap/expect. 0 HashMap. 0 f64. Pass 1-12 clean; automated determinism invariants

Total M5 audit surface: 955 lines code + 17 automated determinism invariants.

M3 + M4 + M5 orchestration ordering (important for the caller):

  1. M3: settle_window(account_table, cemented_user_ops, window_w) — apply cemented user ops to state
  2. M4: apply_noderegistrations_batch (Step 1) → apply_candidate_expiry (Step 3a) → apply_selection_event (Step 3b) — node admission flow
  3. M3: apply_proposal(account_table, node_table, candidate_pool, input, params) — Steps 2 (emission) → 3.5 (chain_length++) → 3.6 (checkpoint rotation) → 4 (state_root)
  4. M5: archive_proposal(header)save_meta_last_cemented(window)save_account_table / save_node_table / save_candidate_pool — persist to disk

Settle is separated from apply_proposal (M3 design choice); steps 1 / 3a / 3b are delegated to M4 mt-entry. M5 finalization atomicity is achieved through the meta-pointer pattern (if a crash happens between archive and meta — verify_consistency detects it).

M3 apply_proposal orchestration ordering (important invariant for the caller):

  1. settle_window(account_table, cemented_user_ops, window_w) — apply cemented ops to state (Transfer / Anchor / ChangeKey / TransferActivation) before apply_proposal
  2. apply_proposal(account_table, node_table, candidate_pool, input, params) — Steps 2 (emission) → 3.5 (chain_length++) → 3.6 (checkpoint rotation) → 4 (state_root)

Settle is separated from apply_proposal by design — the orchestration-ordering invariant is visible to the caller explicitly; hiding it inside one function risks a silent fork. Documented in the module-level comment above pub fn apply_proposal.


2. Conformance Proofs

NIST FIPS 204/203 byte-exact conformance (M1)

66 differential test cases vs NIST CAVP ACVP-Server published vectors:

Test Source Cases Status
ML-DSA-65 KeyGen byte-exact NIST ACVP-Server gen-val/json-files/ML-DSA-keyGen-FIPS204 25 PASS
ML-KEM-768 KeyGen byte-exact NIST ACVP-Server gen-val/json-files/ML-KEM-keyGen-FIPS203 25 PASS
ML-DSA-65 SigGen deterministic external pure (1 empty + 14 non-empty context, 0..255B) NIST ACVP-Server gen-val/json-files/ML-DSA-sigGen-FIPS204 tgId=3 15 PASS
mt_sign_mldsamt_sign_mldsa_ctx (empty ctx) equivalence own correctness check 1 PASS

Reproduction:

cd "<repo-root>" && cargo test -p mt-crypto-native --test nist_acvp_kat -- --nocapture

Fixtures location: crates/mt-crypto-native/tests/fixtures/nist_acvp/

SHA-256 NIST FIPS 180-4 conformance (M1)

Test Source Status
SHA-256("abc") == ba7816bf...15ad FIPS 180-4 §B.1 PASS (in mt-crypto/src/lib.rs::tests::hash_nist_vector_abc)

Internal correctness baselines (M1)

Test Path Status
ML-DSA-65 zero-seed deterministic baseline crates/mt-crypto-native/tests/regression_baselines.rs PASS
ML-KEM-768 zero-seed + ones-seed deterministic baselines same PASS
ML-DSA-65 sign roundtrip + tamper detection same PASS
5 mt-mnemonic KAT vectors (entropy → mnemonic → master_seed → per-role keypair) crates/mt-mnemonic/tests/keygen_vectors.rs PASS
End-to-end recovery flow (entropy → identity round-trip) crates/mt-mnemonic/tests/e2e_recovery.rs PASS
mt_crypto::self_test() (KeyGen determinism + Sign/Verify + KAT 1 byte-exact conformance check) crates/mt-crypto/src/lib.rs PASS

Security invariants (M1, Pass 17 enforcement via automated tests)

Regression detection for security-critical properties — if a future refactor breaks an invariant, the test fails in CI before merge.

Invariant Test Status
SecretKey: !Clone (no accidental copies) crates/mt-crypto/tests/security_invariants.rs::secret_key_is_not_clone
MlkemSecretKey: !Clone same::mlkem_secret_key_is_not_clone
SecretKey: !PartialEq (no timing leak via ==) same::secret_key_no_partial_eq_to_prevent_timing_leak
MlkemSecretKey: !PartialEq same::mlkem_secret_key_no_partial_eq_to_prevent_timing_leak
SecretKey heap-allocated (size = pointer) same::secret_key_is_heap_allocated
MlkemSecretKey heap-allocated same::mlkem_secret_key_is_heap_allocated
SecretKey has a Drop impl same::secret_key_needs_drop
MlkemSecretKey has a Drop impl same::mlkem_secret_key_needs_drop
FFI fills SK with non-zero bytes same::secret_key_filled_by_ffi_keygen
FFI fills MlkemSK with non-zero bytes same::mlkem_secret_key_filled_by_ffi_keygen
No println! / log::* macros on SK bytes in lib code same::no_println_or_log_on_secret_bytes_in_lib_code
PublicKey: Clone (sanity — public material clone-able) same::public_key_can_be_cloned
Signature: Clone (sanity) same::signature_can_be_cloned

13/13 security invariants PASS.

Determinism invariants (M2 — all 4 crates)

Automated regression detection per [I-3] determinism + [I-9] integer arithmetic + [I-10] SSOT + [I-8] Network-Bound Unpredictability.

Crate Test file Cases Status
mt-merkle tests/determinism_invariants.rs 10 PASS (empty_internal determinism + level uniqueness, leaf/internal hash determinism, internal_hash order-sensitivity, SMT empty root + insertion-order-independent root, root-changes-on-insert, type sizes)
mt-genesis tests/determinism_invariants.rs 7 PASS (genesis_app_id determinism, genesis_params singleton stable pointer, ProtocolParams encoded size constant + encoding determinism, compute_genesis_state_hash determinism + dependence on state_root, SSOT singleton consistency)
mt-state tests/determinism_invariants.rs PASS (derive_account_id / derive_node_id determinism + dependence + cross-distinctness, ACCOUNT / NODE / CANDIDATE_RECORD_SIZE matches encoded, encoding determinism, AccountTable / NodeTable / CandidatePool insertion-order-independent root, compute_state_root determinism + order-sensitivity + dependence on each input root, is_active boundary at 2×τ₂ inclusive + saturating_sub safety, WINNER_CLASS_NODE SSOT, empty tables consistent root, insert / remove inverse)
mt-timechain tests/determinism_invariants.rs 19 PASS (vdf_step zero=identity + determinism + dependence on iterations and prev, vdf_verify accept/reject correct/wrong claim/iteration count, vdf_chain composition associative, next_d dead-zone unchanged + above-high-increases + below-low-decreases + determinism, cemented_bundle_aggregate window<2 = genesis zero, empty marker distinct from non-empty, change-on-window / node_ids, canonical sort input-order independence per [I-8])

60/60 M2 invariants PASS.

Spec ↔ Code byte-exact alignment (M0+M1+M2)

Verified per critic audit (commit 9387900):

Check Spec value Code value Status
ML-DSA-65 pubkey size 1952B PUBLIC_KEY_SIZE = 1952
ML-DSA-65 secretkey 4032B SECRET_KEY_SIZE = 4032
ML-DSA-65 signature 3309B SIGNATURE_SIZE = 3309
ML-DSA-65 seed 32B KEYPAIR_SEED_SIZE = 32
ML-KEM-768 ek 1184B MLKEM_PUBLIC_KEY_SIZE = 1184
ML-KEM-768 dk 2400B MLKEM_SECRET_KEY_SIZE = 2400
ML-KEM-768 seed 64B (d‖z) MLKEM_SEED_SIZE = 64
AccountRecord size 2059B ACCOUNT_RECORD_SIZE = 2059
NodeRecord size 2098B NODE_RECORD_SIZE = 2098
CandidateRecord size 2082B CANDIDATE_RECORD_SIZE = 2082
ProposalHeader size 3722B PROPOSAL_HEADER_SIZE = 3722
ProtocolParams size 4094B (spec v34) PARAMS_ENCODED_SIZE = 4094
Sparse Merkle Tree depth 256 TREE_DEPTH = 256
Emission pin 13 Ɉ const per window emission_moneta = 13_000_000_000
Domain registry sync (32 domains) spec list code mt-codec const list

Security Cards per crypto primitive (M1)

Each primitive with secret material has a detailed Security Card in docs/security-cards.md. Cards covered:

Primitive Card Status
SecretKey (ML-DSA-65) Card 1 — 8/8 Pass 17 checks closed closed
MlkemSecretKey (ML-KEM-768) Card 2 — 8/8 closed closed
keypair_from_seed (ML-DSA KeyGen) Card 3 — heap+mlock+stack hygiene closed closed
keypair_from_seed_mlkem (ML-KEM KeyGen) Card 4 — heap+mlock+stack hygiene closed closed
sign (ML-DSA-65 deterministic) Card 5 — 8/8 closed closed
verify (ML-DSA-65) Card 6 — no secret material, minimal card closed

3. Threat Model & Scope Boundaries

In scope — what the auditor should review

M1 — cryptography + identity:

  1. FFI memory safety — Layer 1 unsafe blocks (4 sites), buffer sizes, pointer validity
  2. OpenSSL EVP API misuse — EVP_PKEY_CTX_set_params parameters, error handling, EVP_DigestSignInit determinism flag
  3. Secret hygieneDrop + zeroize for SecretKey / MlkemSecretKey, no Clone / Copy on secret types, no leakage in logs / dumps, heap-allocated bytes (Box) with mlock against swap-out, stack hygiene — FFI writes directly into a heap-allocated locked Box, with no stack temporary buffers holding secret bytes (see docs/security-cards.md Cards 1-6)
  4. Result API correctnesssign / keypair_from_seed corruption resistance: malformed SK via from_array(arbitrary_bytes)Err(CryptoError::InvalidSecretKey), not a panic
  5. Deterministic semantics — FIPS 204 Algorithm 2 deterministic Sign correctly invoked via OSSL_SIGNATURE_PARAM_DETERMINISTIC=1 (consensus determinism per Montana [I-3])
  6. Recovery flow correctness — mnemonic → master_seed (PBKDF2-HMAC-SHA-256) → per-role HKDF derivation → byte-exact reproducible identity (Montana spec sections "Keys → Mnemonic and seed", "Per-role key derivation")
  7. Cross-compile correctness — build.rs uses the CARGO_CFG_TARGET_OS env var (not cfg!(target_os))
  8. Reproducible builds — Docker container with pinned base image (debian:bookworm-slim@sha256:40b107342c492725bc7aacbe93a49945445191ae364184a6d24fedb28172f6f7), byte-identical across independent runs (CI gate reproducible_release)

M2 — state foundation: 9. Sparse Merkle Tree correctness — empty_internal level uniqueness + cache consistency + verify_proof for inclusion and absence proofs 10. Genesis Decree SSOT — ProtocolParams singleton consistency, encoded size invariance under struct changes, frozen constants 11. State table determinism — BTreeMap canonical sort guarantee for AccountTable / NodeTable / CandidatePool, insertion-order-independent state_root 12. Account / Node ID derivation — SHA-256 domain-separated via mt-codec::ACCOUNT / ::NODE registry, cross-distinctness for the same pubkey 13. Monetary policy correctness — const emission reward_moneta = EMISSION_moneta = 13 × 10⁹ nɈ per spec «Emission», closed-form supply_moneta(W) = EMISSION_moneta × (W + 1) 14. TimeChain VDF correctness — vdf_step composition associativity, vdf_verify byte-exact against re-computed chain 15. Adaptive D feedback — next_d dead-zone semantics + integer permille arithmetic per [I-9] 16. cemented_bundle_aggregate per [I-8] — Network-Bound Unpredictability via canonical sort over node_ids + window binding, protection against input-order grinding

Out of scope — NOT subject of this audit

  1. OpenSSL internal correctness — out of scope; OpenSSL Foundation's responsibility. The auditor may assume the OpenSSL implementations of FIPS 204 / 203 Algorithms 1 / 2 / 16 are correct (audited and deployed at scale)
  2. Consensus protocol correctness M3-M5 — separate layers (mt-account, mt-consensus, mt-lottery, mt-entry, mt-store) — internal-tested with 255+ tests green, but not yet audit-prepared (TODO for subsequent audit phases)
  3. Network protocol — M6+ (mt-net does not yet exist)
  4. Application layer — a separate workspace in the future (Junona agent, messaging, file storage)
  5. Hardware side-channel attacks — software-only assumption; embedded-deployment audit is separate

Known limitations (deferred, with rationale)

  1. Genesis bootstrap ceremony pending (M2-1 finding from external audit Claude Opus 4.7 #2) — 4 ProtocolParams fields hold placeholder zeros and are finalized only during the Genesis ceremony before mainnet:

    • bootstrap_account_pubkey: [0u8; PUBLIC_KEY_SIZE]
    • bootstrap_node_pubkey: [0u8; PUBLIC_KEY_SIZE]
    • target_zero: [0u8; 32] (initial VDF target)
    • genesis_content_data_hash: [0u8; 32]

    Severity: mainnet blocker, NOT a code-audit blocker. Layout / encoding / SSOT / singleton are correct and audit-ready. Finalization = choosing seeds, generating keypairs via mt_crypto::keypair_from_seed, secure privkey storage (multi-party ceremony or a single trusted party — author's design decision).

    Programmatic check: mt_genesis::is_genesis_bootstrap_finalized(params) -> bool for operator deployment scripts. The bootstrap_keypairs_finalized test stays #[ignore] until the ceremony; the ignore is removed after the ceremony. The pre-ceremony test is_genesis_bootstrap_finalized_pre_ceremony_returns_false explicitly captures the current state.

    Closure path: Genesis ceremony plan — a separate milestone in ROADMAP.md before mainnet.

  2. Sign with non-empty context — the current mt_sign_mldsa API does not accept a context parameter and uses an empty context (matches Montana's usage pattern but not the full FIPS 204 API surface). NIST ACVP cases with context bytes (14 / 15 in the external-pure deterministic group) are not exercised. Closure path: extend the FFI signature when FIPS context support is needed (M6+).

  3. Verify NIST KAT — a dedicated test set for FIPS 204 SigVer has not been added in this session (sign-baseline + sign-NIST KAT byte-exact match indirectly confirms Verify correctness via round-trip, but a direct Verify NIST KAT remains as an enhancement).

  4. ML-KEM-768 Encapsulate/Decapsulate NIST KAT — the current M1-F scope covers KeyGen only (Encapsulate/Decapsulate will be needed in M6+ application-layer encryption). NIST KAT for encapDecap is already present in the fixtures dir; tests are added when needed.


4. Build & Reproduction

Toolchain

rustc 1.70+ (pinned via rust-toolchain.toml)
Cargo workspace
OpenSSL 3.5.5 LTS (vendored via openssl-src)

Optional security audit tooling (for Step H verification)

cargo install cargo-audit --locked

Once per audit machine. After that, the cargo audit command is available.

Build from source

cd "<repo-root>"
cargo build --all --release

All 4 mandatory CI checks:

cargo fmt --all -- --check
cargo clippy --all-targets -- -D warnings
cargo test --all
cargo build --all --release

Reproducible release build (Docker)

Prerequisites: Docker ≥ 20.10 installed and running, ~30 minutes wall-clock for two clean builds, ~5 GB free disk.

docker build --file docker/release-build.dockerfile --tag montana-release:audit .

CI gate reproducible_release (see .github/workflows/ci.yml) performs two independent builds with --no-cache and asserts byte-identical hashes on every push to main. The auditor can verify via CI history without running locally (avoiding the Docker prerequisite).

Run audit-relevant tests

M1 NIST ACVP KAT (cross-implementation conformance):

cd "<repo-root>" && cargo test -p mt-crypto-native --test nist_acvp_kat -- --nocapture

M1 Internal correctness baselines:

cd "<repo-root>" && cargo test -p mt-crypto-native --test kat_independent -- --nocapture
cd "<repo-root>" && cargo test -p mt-crypto -- --nocapture
cd "<repo-root>" && cargo test -p mt-mnemonic --test keygen_vectors -- --nocapture
cd "<repo-root>" && cargo test -p mt-mnemonic --test e2e_recovery -- --nocapture

M1 Security invariants (Pass 17):

cd "<repo-root>" && cargo test -p mt-crypto --test security_invariants -- --nocapture

M2 Determinism invariants:

cd "<repo-root>" && cargo test -p mt-merkle --test determinism_invariants -- --nocapture
cd "<repo-root>" && cargo test -p mt-genesis --test determinism_invariants -- --nocapture
cd "<repo-root>" && cargo test -p mt-state --test determinism_invariants -- --nocapture
cd "<repo-root>" && cargo test -p mt-timechain --test determinism_invariants -- --nocapture

Full M1+M2 audit chain:

cd "<repo-root>" && cargo test -p mt-codec -p mt-crypto-native -p mt-crypto -p mt-mnemonic -p mt-merkle -p mt-genesis -p mt-state -p mt-timechain

5. Audit History

Date Audit Findings Closure
2026-04-27 Spec bump v33.1.6 → v34.0.0 — major monetary policy refactor (geometric step-up baseline → const linear emission) Architectural simplification, not security audit findings ProtocolParams: 4 fields removed (r_genesis_moneta, monetary_epoch_windows, inflation_num, inflation_den) + emission_moneta: u128 = 13×10⁹ nɈ added. PARAMS_ENCODED_SIZE 4118 → 4094. mt-state: MonetaryState struct + carry-recurrence machinery removed. mt-account apply_emission simplified: reward = params.emission_moneta const, monetary_epoch_tick removed. supply_moneta(W) = E × (W+1) closed-form O(1) — eliminates a class of bugs (epoch boundary off-by-one, carry overflow). 288/288 affected tests PASS. Driver: simplifying the audit surface for production firm engagement.
2026-04-27 External audit #4 Claude Opus 4.7 (1M context) — incremental M4+M5 scope (mt-lottery, mt-consensus, mt-entry, mt-store) — independent SHA-256 oracle (Python hashlib) cross-check 4 hash compositions + ad-hoc ln_q64 mathematical verification (T141253) 10 findings (0 CRITICAL/HIGH, 2 MEDIUM, 5 LOW, 3 INFO); score 8.5/10; 325/325 M4+M5 tests PASS under single-core / single-process All 10 findings closed by construction. Commits: cb81d4c M4-MED-1 window_index u32→u64 unification (spec v33.1.4→v33.1.5), f46d48a M4-LOW-7 3 hardcoded const → ProtocolParams [C-1] SSOT (spec v33.1.5→v33.1.6 + admission_divisor field), 64cefca M4-MED-2 validate_winner genesis-aware contract docs, 53ba832 M4-LOW-3 log2_q64 panic-free byte extraction, e8450bb M4-LOW-4 checked_add validate_header window monotone, c41f043 M4-LOW-5 saturating_mul quorum, 845d74d M4-LOW-6 positive functional tests TooManyOps/TooManyReveals, bcf5e9b M5-LOW-8 cleanup orphan .tmp at FsStore::open + anti-regression test, 022c5ff M4-INFO-10 canonical_proposer degraded-mode docs.
2026-04-27 External audit #3 VERIFIED Claude Opus 4.7 (1M context) — incremental M3 mt-account scope + full zero-trust verification of M1 / M2 closures + fresh NIST CAVP source download (T124438; supersedes incomplete first pass T121239 in which the auditor by their own admission cut corners). 785/785 cargo test --all PASS. 10 findings: 8 from the incomplete first pass (1 HIGH M3-A-4, 3 LOW M3-A-1 / M3-A-2 / F-5, 4 INFO M3-A-3 / M3-A-5 / M3-A-7 / M3-A-8) + 2 from the VERIFIED pass (P-C-1 F-19 reopen HIGH, P-C-2 audit history sync MEDIUM). Score 8.5/10; expected 9/10 after P-C-1 + P-C-2. All 10 findings closed by construction. Commits: 4c14685 M3-A-4 ValidationContext, 047cc45 M3-A-1 + M3-A-2 checked arithmetic, 66df1a1 M3-A-5 + F-5 docs / telemetry TODO, e61c479 M3-A-3 binding test vector, 5d3506c Cargo.lock sync, b96085d P-C-1 SAFETY convention for 4 const-cast sites + Rust shim cross-references, <this> P-C-2 audit history sync. 131 tests PASS in mt-account (96 unit + 35 determinism); 4 anti-regression tests (validate_dispatcher cooldown enforcement, apply_chain_length overflow panic, apply_checkpoint_rotation underflow panic, genesis_candidate_root binding empty_internal(256)).
2026-04-26 External audit #2 Claude Opus 4.7 (1M context) — incremental M2 scope (mt-merkle, mt-genesis, mt-state, mt-timechain) 17 findings (1 HIGH M2-1 Genesis bootstrap pubkeys placeholder, 5 MEDIUM, 11 LOW/INFO); score 8/10; 60/60 determinism invariants verified M2-3 + M2-13 originally closed in the M3 milestone (binding test pin 41/40 + MonetaryState in compute_state_root). Superseded by spec bump v33.1.6 → v34.0.0 monetary policy refactor (geometric step-up → const linear emission): pin 41/40 inflation_num / den and the MonetaryState carry-recurrence removed as artifacts of the old policy. New equivalent invariants: reward_moneta(W) = EMISSION_moneta const closed-form, supply_moneta(W) = EMISSION × (W+1) linear closed-form. M2-1 re-classified as a known limitation (mainnet deployment blocker, not a code-audit blocker). The remaining documentation-only findings were closed in audit-package iterations.
2026-04-26 External audit #1 Claude Opus 4.7 (1M context) — initial M1 layer (mt-crypto, mt-crypto-native, mt-mnemonic) — independent NIST CAVP source download + byte-exact verification 19 findings (3 HIGH + 4 MEDIUM + 12 LOW/INFO); score 8/10; 0 cryptographic vulnerabilities 14/19 closed by construction (commit 6ff26b3 + closure completed in b96085d for F-19): F-2 cargo fmt clean, F-3 stale RustCrypto refs cleanup, F-4 SAFETY comments on 3 mlock / munlock blocks, F-5 expanded SAFETY comment, F-6 keypair() migrated to OS CSPRNG (getrandom), F-7 zeroize intermediate PBKDF2 / HKDF / HMAC buffers, F-8 SigGen NIST KAT expanded 1 → 15 cases with the new mt_sign_mldsa_ctx API, F-9 rename kat_independent → regression_baselines, F-12 split_whitespace instead of split(' '), F-18 cc parallel feature removed, F-19 const-cast OpenSSL convention in SAFETY (initially incomplete; fully closed in b96085d after the VERIFIED audit reopen). 5 deferred with explicit rationale (F-10 NIST in self_test, F-14 formal constant-time, F-15 cargo-fuzz infrastructure, F-16 threshold sigs M6+, F-17 serde_json dev-dep — the auditor themselves marked these as "not a priority"). Expected score after closure: 9/10.
2026-04-26 M0+M1+M2 critic spec-vs-code audit (Montana-Protocol/Code/CRITIC.md v1.6.0) — verification of full M0 / M1 / M2 code conformance with spec v33.1.2 4 findings (F-1 mt-recovery-fingerprint domain spec drift, F-2 stale VERSION.md Implementation field, F-3 false positive withdrawn by the critic, F-4 controlled halts documentation) All 4 closed (commit 9387900): F-1 spec patch v33.1.2 → v33.1.3 (Domain registry sync), F-2 VERSION.md updated to «M0..M5 closed», F-4 audit-checklist §K Controlled halts section
2026-04-26 M1-F audit-package critic review (Montana-Protocol/Code/CRITIC.md v1.4.0) — verification of AUDIT.md + reproduction commands + claims accuracy 5 findings (F-A1..F-A5) — all documentation-only, not security All 5 closed: F-A1 unsafe count 3→4 (verify added in list), F-A2 cargo audit verified+install instructions, F-A3 line counts exact (568/40/375/56/45=1084), F-A4 print_sk gate accurately described (env var M1_DUMP_SK=1), F-A5 Docker reproduction prerequisites added
2026-04-26 M1-F internal critic audit (Montana-Protocol/Code/CRITIC.md v1.4.0) 7 findings (F-1..F-7) 6 closed by construction, F-3 closed via NIST ACVP differential testing — all 7 closed. Commits: 9f2ba93 (F-4 cross-compile), 71896f6 (F-6 separate error codes), 3333738 (F-1 zeroize), e1164ad (F-2 Result API + F-5 cfg-gate keypair + F-7 implicit), 6b7ff30 (F-3 NIST KAT)
2026-04-26 M2 batch 2+3 audit prep — manual critic Pass 1-12 for mt-state + mt-timechain 0 findings (manual scan clean) 24 + 19 automated determinism invariants added
2026-04-26 M2 batch 1 audit prep — manual critic Pass 1-12 for mt-merkle + mt-genesis 0 findings (manual scan clean) 10 + 7 automated determinism invariants added
2026-04-21 External critic finding — domain separation prefix-collision 5 findings (P1-P5) All closed: NUL separator pattern, SK leak gate, label coherence, empty_internal binding, stale RECOVERY disclosure. Commit d762cec, spec v29.13.0.

Audit-ready status (2026-04-27): zero open code-side findings in M1 + M2 + M3 layers across 3 independent external audits (T201805 + T232707 + T124438 by Claude Opus 4.7 1M; T124438 — VERIFIED replacement of the incomplete first-pass T121239). Total cumulative 46 findings across audits — all code-side closed by construction; deferred items explicitly classified (M2-1 Genesis bootstrap pubkeys = mainnet deployment, not code audit; F-14 / F-15 formal verification + fuzzing = post-MVP scope). M3 score 8.5/10 → 9/10 ready after P-C-1 + P-C-2 closures (commits b96085d + this). Documentation accuracy verified through 3 incremental audits, including the VERIFIED zero-trust pass. Cross-implementation conformance verified through spec ↔ code byte-exact alignment + independent NIST CAVP source download (65 KAT cases byte-exact against a fresh source 2026-04-27).


6. Spec & Documentation

Document Path Scope
Protocol spec see VERSION.md (Spec target / Spec path) — authoritative source Full protocol specification, source of truth
App spec Montana-Protocol/Montana App v3.9.2.md (current at the time; see the VERSION.md history) Application layer specification
Architect role CLAUDE.md Code architect role definition (v1.12.0)
Critic role CRITIC.md Code critic role definition (v1.6.0 — Pass 17 Security Card mandatory)
Roadmap ROADMAP.md Implementation roadmap, milestones, status
Spec version SSOT VERSION.md Single source of truth for the spec target version
Pre-audit checklist docs/audit-checklist.md 11 categories of self-attestation (A-K) + reproduction one-liners for the auditor + sign-off table
Security Cards docs/security-cards.md Pass 17 mandatory Security Cards for 6 crypto primitives

7. Pre-audit self-attestation

Before starting an external audit — see docs/audit-checklist.md for the full 11-category checklist.

M1 + M2 final closure self-attestation (2026-04-26):

Findings closure:

  • All 7 M1-F audit findings closed by construction (F-1..F-7)
  • All 5 audit-package findings closed (F-A1..F-A5)
  • All 4 M0+M1+M2 critic spec-vs-code findings closed (F-1..F-4)
  • Total: 16/16 findings closed, 0 open blockers

Conformance:

  • NIST ACVP differential testing PASS (51/51 KAT byte-exact)
  • Spec ↔ code byte-exact alignment verified (16 key constants / sizes table)
  • Domain registry sync (current spec target ↔ mt-codec, see VERSION.md)

Security:

  • 13/13 security invariants PASS (Pass 17 enforcement)
  • 6/6 Security Cards filled in and closed (docs/security-cards.md)
  • Heap+mlock+zeroize for secret types verified

Determinism:

  • 60/60 M2 determinism invariants PASS
  • 0 HashMap (real usage) in consensus path; BTreeMap canonical sort
  • 0 SystemTime / Instant::now / RNG in consensus path
  • 0 f32 / f64 in consensus path

Build & CI:

  • cargo audit clean (verified 2026-04-26: 0 vulnerabilities, 0 warnings, 39 deps)
  • cargo fmt --all -- --check clean
  • cargo clippy --all-targets -- -D warnings clean
  • cargo test --all 564+ PASS / 0 FAIL
  • cargo build --all --release clean
  • Reproducible builds verified through CI gate reproducible_release

Documentation:

  • Public API stable (mt-crypto types / signatures unchanged through M1-E + M1-F migrations)
  • Audit chain documented (Layer 1/2/3 explicit, exact line counts)
  • Threat model documented (in/out of scope explicit)
  • Known limitations documented with closure path
  • Reproduction commands verified working (each one tested 2026-04-26)
  • Documentation accuracy verified through critic review of audit package itself

Contact

Audit findings, questions, reports — fork the repository, open a PR with patch + reproduction instructions.

Architect role + critic role files in the repository — peer-reviewable methodology.