Allowed wording:
- FIPS 203-aligned ML-KEM through
pqcrypto. - FIPS 204-aligned ML-DSA through
pqcrypto. - FIPS 205-aligned SLH-DSA through
pqcrypto.pqforgecomposes it intokeygen, key custody, and detachedsign/verify. Envelope headers, streaming signatures, andhybrid-signremain ML-DSA-only. - Application-layer composition helpers for KEM-DEM, AEAD sessions, wrapped key
custody, signatures, recipes, hybrid helpers, X.509 certificates, PKCS#10
requests, RFC 9180 HPKE, and CLI workflows. A certificate that
openssl reqcan read is not a WebPKI CA and not a trust store. HPKE here is the RFC 9180 composition over X25519, X448, P-256, P-384, and P-521, with HKDF-SHA256, HKDF-SHA384, HKDF-SHA512, and the AEADs the package implements. BoringSSLEVP_HPKEevidence is base and auth on X25519 and P-256 only. OpenSSLOSSL_HPKEevidence is all four modes on X448 and P-521 only. It is not a FIPS module and not TLS. X-Wing (PqXWing,HpkeKem.xWing, id0x647A, draft-connolly-cfrg-xwing-kem-11) is an additional KEM. It is not a FIPS construction. FIPS mode rejects it atPqHpkeandPqForgeHybridKeyAgreement. Auth is not defined. It does not change.pqfor.pqfs. There is no BoringSSL or OpenSSL evidence for0x647A. - The Double Ratchet (
DoubleRatchet) is Signal revision 4 over X25519 and pqforge AES-256-GCM or ChaCha20-Poly1305, with a 12-byte zero nonce and a 40-byte header. Healing is classical. It is not PQXDH, not the Sparse Post-Quantum Ratchet, not libsignal bytes, and not a FIPS module.PqFipsModerefuses the ChaCha suite.PqxdhandTripleRatchetare exported and throw. PqFipsModeis a process-wide switch, not a validated module. It also rejects RSA-SHA-1 and an RSA modulus under 2048 bits atPqRsa.signPkcs1andsignPss(verify returns false). HPKE seal and open refuse an AEAD other than AES-256-GCM. Background file helpers re-enable the flag in their isolate and refuse a customPqRandom.generator. It does not gate Brainpool, FFDHE, AES-128-GCM at the primitive (QUIC Initial), a rawPqForgeAeadEngine.seal, orPqXWingitself.- Brainpool ECDH (
PqBrainpool) for brainpoolP256r1, brainpoolP384r1, and brainpoolP512r1, and RFC 7919 FFDHE (PqFfdhe) for the five named groups OpenSSL negotiates. BoringSSL does not negotiate these. The byte layout matches the TLS key share.pqforgedoes not speak TLS. Neither implementation is constant-time. The FFDHE exponent lives in aBigIntand is not wiped. The random buffer used to draw it is. - Opt-in signature composition in the same release.
compose-signtakes 2 or 3 legs the caller names (ML-DSA, SLH-DSA, Ed25519, ECDSA-P256, ECDSA-P384, ECDSA-P521).hybrid-signstays one ML-DSA leg plus one classical leg.PqEcdsaP521is a signature. It is notp521SharedSecretand not a TLS group. The P-521 HPKE share is unchanged. - Best-effort cleanup in Dart.
Forbidden wording:
- FIPS validated.
- FIPS 140 validated.
- CMVP validated.
- Certified.
- Hard constant-time guarantee.
- Hard secure-erasure guarantee.
- ML-KEM alone is secure transport.
- AES signs documents.
- RC4 is supported.
- ECDSA-P521 is a TLS group, or is the HPKE P-521 KEM.
compose-signor ECDSA-P384/P-521 is constant-time or a FIPS module.- SLH-DSA is an envelope signature or a
hybrid-signleg. - Brainpool or FFDHE is a TLS stack, constant-time, or a NIST curve. There is no NIST P-512.
- BoringSSL 0.20261020.0 checks ML-KEM-512 or any SLH-DSA set other than SHA2-128s and SHAKE-256f. Those two, ML-KEM-768, ML-KEM-1024, and ML-DSA-44/65/87 are the sets that pin exports. The rest of SLH-DSA is an OpenSSL 4.1.0 check.
- X-Wing is a FIPS construction, a
.pqfor.pqfsalgorithm, or checked against BoringSSL or OpenSSL. Id0x647Ahas no such evidence. - The Double Ratchet is libsignal-compatible, PQXDH, the Sparse Post-Quantum Ratchet, header encryption, or a FIPS module.
pqforge can inherit algorithm-evidence wording from pqcrypto, but it must not
upgrade that into module-validation wording. Public-key trust, identity vetting,
authorization policy, replay stores, legal policy, and infrastructure custody
remain application responsibilities.
pqforge scrubs secret buffers through package:zeroize's secureZero, which
uses an overwrite pattern plus @pragma('vm:never-inline') opaque read anchors
so Dead Store Elimination cannot prove the writes dead. A bare
Uint8List.fillRange(0, n, 0) is removable by the AOT compiler, because nothing
reads the buffer afterwards.
That is the whole of what a wipe buys, and the following are not claims:
- A wipe is not erasure. The runtime may hold another copy of the buffer —
the GC can move it before disposal, and the old address is then never written.
Pure Dart cannot
mlockpages or exclude them from swap or core dumps. - A wipe cannot reach registers. Values held in VM registers or spilled to
stack slots are outside
secureZero's reach. Registers are spilled on context switch and can appear in core dumps and VM snapshots. - A wipe can, in principle, create a copy. If the value being wiped is still
needed after the wipe call, the compiler must preserve it across the call and
may spill a second copy that the wipe does not cover. This is why every wipe
site in
pqforgeis copy-then-wipe and never wipe-then-use, and why wiping a value that only ever lived in a register is pointless. See https://00f.net/2026/10/06/zeroization-1/. - Public values are not wiped.
checkEncapsulationKeydeliberately does not wipe the ML-KEM ciphertext. Wiping non-secret data costs time and implies a sensitivity that does not exist.
Do not describe secureZero as guaranteeing that key material leaves memory,
and do not describe it as best-effort in a way that implies it can make the
situation worse than not calling it — for heap-resident buffers, which is all
secureZero accepts, it cannot.