Skip to content

About

Custodian-level native-Rust reconstruction of the externally observable OpenSSL 3.x contract.

Topics

Resources

Stars

50 stars

Watchers

0 watching

Forks

Repository files navigation

openssl-rs

A native Rust reimplementation of OpenSSL 3.6.4 targeting source, ABI, and observable behavioural compatibility.

openssl-rs reconstructs the OpenSSL distribution contract — libcrypto, libssl, providers, public headers, static libraries, and the openssl CLI — without using OpenSSL or another cryptographic library as its implementation backend. It is not a wrapper around OpenSSL, rustls, AWS-LC, BoringSSL, LibreSSL, or ring.

OpenSSL authority openssl-3.6.4-production (3.6.4)
Platform linux-x86_64
Build profile linux-x86_64-default-shared-legacy-notests
Implementation one first-party Rust crate; zero runtime/build dependencies

Status: the implementation state is generated, not typed here. IMPLEMENTED (a symbol the crate's compiled output defines) and SCAFFOLDED (a shell-only abort) are counted in STATUS.md and SEAL-CENSUS.md, which are authoritative; this page does not restate the counts. IMPLEMENTED is neither PARITY_VERIFIED nor a compatibility claim.

What is achieved

Measured at the source identity rendered below. The reuse row states how many measurements are carried forward, and the superseded identity they were measured at. Every row names its own machine-readable evidence; none of them is a claim of completeness.

Source identity: candidate commit 1e7c221f0229ea7ae7d3a55ff2ce8aa9db242134; build source-tree digest ca7d39b9bcfbe9827ded8b255a66980d32f16a9fbd15902f80dc1b921839d1c4; crate version 0.0.49 (published 2026-11-18, checksum 4231962b…); libcrypto.so.3 sha256 fd25cc9e6424ea79aaaff05df22b3f6e1d46aa22d5666cfe7589dce95ffd88a9; libssl.so.3 sha256 648f8caecda245d2186813fb67ceadedd798b9ce609c8b041a84f195946c5f68. Derived by forensics/tools/render_docs.py; --check fails if a block and its evidence disagree.

Dimension Result Evidence
Downstream drop-in, Safe-33 cohort 33 / 33 families at L7-functional, candidate and authority; carried forward — 27/33 cohort rows are in evidence_reuse.reused_families and candidate_identity_match=false forensics/downstream/runtime-functional-atlas.json (body.runs); CAMPAIGN-STATUS.json (consumers)
Downstream drop-in, 1000-family corpus 207 / 1000 at L7-functional; candidate 207 / authority 209 L7-functional; measured asymmetry: authority-only ewftools, xilinx-bootgen; 307 recipes, 253 built, 245 linked; 223 loaded, 22 L4-linked, 7 failed, 755 not attempted (693 with no recipe); 301 of 307 recipe-backed rows carried forward same atlas, body.counts.by_subject, body.evidence_reuse; forensics/downstream/build-link-atlas.json
Library test suite 2,231 passing, 0 failing, 52 ignored at cc7762e5d — a floor, not the current count: later commits add tests without a further full run atlas-refresh/test-baseline.json (full-library-run)
Memory safety: mechanisms 6 mechanisms discharged at mechanism scope — RC-1 refcounting, OB-1 OwnedBuf release state, OB-2 OwnedStr string ownership, OB-3 OwnedCtx/RawCtx context ownership, ERR-1 ERR error-queue ownership and allocator file-coordinate seam, GEN-1 GENERAL_NAME selected-arm representation CAMPAIGN-STATUS.json (mechanism_discharges)
Memory safety: boundaries 19 scoped boundary discharges (B23, B25, B27, B29, B45, B48, B49, B50, B51, B52, B53, B54, B55, B56, B57, B60, B67, B69, B70); 4 proposals not applied (B28, B46, B47, B71); B39–B40 independently checked; B41–B43 batch-checked; B44 author-evidenced atlas-refresh/boundary-sweep-*.json
Memory safety: files converted / partial 877 files: 92 ZERO_UNSAFE, 18 BOUNDARY_COVERED, 761 PARTIAL (audited 761/761), 6 TEST_ONLY, 0 unaccounted atlas-refresh/unsafe-census.json, atlas-refresh/post-census-delta.json
SAFE_COMPATIBLE consumers 1 / 33 — the admitted openssh family graduated at b85c via its consumer binaries ssh, sshd; every other admitted family still has no proven dependency closure CAMPAIGN-STATUS.json (consumers)
BOUNDARY / SEAL gates NOT_RUN / NOT_RUN — no independent caller/dependency-closure audit, Miri/ASan/Loom/mutation control, authority differential, full forensic regeneration or held-out compatibility for the S16–S61 cluster CAMPAIGN-STATUS.json (gates.boundary_gate, gates.seal_gate)
Allocator / ABI invariants preserved — recorded invariant planes: abi_courts, allocator_seam, exported_symbols, repr_c_layouts, unsafe_lints CAMPAIGN-STATUS.json (structural_invariants)
Known gaps CLOSURE-NON-MONOTONE REPAIRED; S21-O1 CLOSED; S22-RETAINED RETAINED; DOWNSTREAM-IMPLEMENTATION 0 OPEN AT THE MEASURED IDENTITY; DOWNSTREAM-h2o CLOSED; DOWNSTREAM-charybdis CLOSED IN S61 (carried forward); DOWNSTREAM-fastd ENVIRONMENT; DOWNSTREAM-HARNESS-FIXTURE 76 CARRIED S59-S60 BLOCKERS + 25 NEW VENUE/BUILD BLOCKERS; DOWNSTREAM-BELOW-L7 100 FAMILIES; DOWNSTREAM-S42-SYMMETRIC-LOSSES CLOSED IN S44; DOWNSTREAM-partclone CLOSED; DOWNSTREAM-perl-net-ssleay CLOSED; DOWNSTREAM-NO-RECIPE 693 UNMEASURED; DOWNSTREAM-MEDUSA HARNESS; DOWNSTREAM-S53-YIELD 2 NEW L7 / 35 ADMITTED; DOWNSTREAM-S60-YIELD 7 NEW L7 / 30 ADMITTED; S56-RESIDUALS RETAINED; S57-RACE DATA-RACE FREEDOM NOT ESTABLISHED; S58-SECURITY-CALLBACK PERMISSIVE ARMS RETAINED; S61-RESIDUALS OTHER CALLBACK ARMS REMAIN PERMISSIVE; ELEMENT-7-PRECISION OPEN — FAIL; CONTRACTED-OBLIGATIONS-BLOCKED 215 BLOCKED at the applied v1u closure (222 at the historical v1r closure); RELEASE-EVIDENCE-DEBT OPEN; SSL-CONF-WITNESS OPEN — blocks 18 SSL_SESSION sites CAMPAIGN-STATUS.json (known_gaps)
Release Published: 0.0.49 on crates.io (2026-11-18), checksum 4231962b…, install-tree artefacts libcrypto.so.3 fd25cc9e… / libssl.so.3 648f8cae…. The published crate ships source and documentation only, with no binaries; those install-tree hashes describe the repository build recorded in the receipt, not the crate payload. This evidence is bound to the 0.0.49 tree, which is the published 0.0.49 crate. 0.0.48 remains the previous release. atlas-refresh/publish-0.0.49.json; Cargo.toml

Downstream evidence reuse: 301 of 307 recipe-backed families carried forward (committed-rows-carried-forward); candidate_identity_match = false. A carried-forward row is compatibility evidence from the build recorded on that row, not a fresh measurement at the identity above. The Safe-33 cohort has 27/33 of its families in that set.

The families at L7-functional

209 of 1000 counted corpus families are L7-functional — candidate 207 / authority 209 L7-functional; measured asymmetry: authority-only ewftools, xilinx-bootgen; 7 failed, 223 loaded, 307 recipes, 755 not attempted.

afflib†, afftools†, aircrack-ng†, apache2-bin†, apache2-http2†, apache2-ssl†, apache2-utils†, appx-util†, apr-util†, arm-trusted-firmware-tools†, asterisk-srtp†, axel†, bannergrab†, bruteforce-salted-openssl†, carbon-c-relay†, casync†, charybdis†, cjose†, cowpatty†, curl, curl-minimal†, cyrus-sasl†, cyrus-sasl-digestmd5†, cyrus-sasl-scram†, dcap-tunnel-ssl†, dcmtk†, dma†, dovecot†, duo_unix†, efitools†, eggdrop†, elinks†, epic5†, epic5-lite†, ewf-tools†, ewftools†, exim†, fetchmail†, fido2†, fossil†, fsverity-utils†, gensio†, getdns†, gftp-text†, git-crypt†, gmnisrv†, gostsum†, h2o†, h2o-evloop0.13†, haproxy, hcxkeys†, hcxtools†, hiredis-ssl†, hitch†, httest†, httpd†, httperf†, httping†, httrack†, hydra†, icecast†, icecast2†, ima-evm-utils†, imapfilter†, iperf3†, ircd-hybrid†, ircii†, irssi†, isync†, jimtcl†, jwt†, kea†, keepalived†, kmod†, ldns†, ldnsutils†, lftp†, libafflib0v5†, libaprutil1†, libarchive†, libcjose0†, libcoap†, libcpp-httplib0.11†, libcrypt-openssl-bignum-perl†, libcrypt-openssl-dsa-perl†, libcrypt-openssl-ec-perl†, libcrypt-openssl-pkcs10-perl†, libcrypt-openssl-random-perl†, libcrypt-openssl-rsa-perl†, libcrypt-openssl-x509-perl†, libcrypt-smime-perl†, libdcmtk17†, libengine-gost-openssl†, libesmtp†, libevent†, libevent-openssl-2.1-7†, libevhtp†, libfsverity†, libgetdns10†, libjose†, libnetconf2†, liboauth2†, libpaho-mqtt†, librabbitmq†, librdkafka†, libre†, librelp†, libretls†, libssh†, libstrophe†, libtins†, libtpms†, libu2f-server†, libwebsockets†, libxmlsec1†, libzip†, lighttpd†, links†, lua-luaossl†, lua-sec†, lynx†, makepasswd†, memcached†, mini-httpd†, minizip-ng†, mktorrent†, monit, monitoring-plugins†, mosquitto†, mumble-server†, mutt†, nagios-plugins-http†, nagios-plugins-ldap†, nagios-plugins-smtp†, nagios-plugins-tcp†, ncrack†, neon†, net-snmp†, nghttp2†, nginx†, njs†, nmap†, nmap-ncat†, nmh†, nrpe†, nsd†, opendkim†, openldap†, openssh, openssh-client†, openssh-server†, openvpn†, osslsigncode†, partclone†, perl-crypt-openssl-aes†, perl-crypt-openssl-bignum†, perl-crypt-openssl-dsa†, perl-crypt-openssl-random†, perl-crypt-openssl-rsa†, perl-crypt-openssl-x509†, perl-crypt-smime†, perl-net-ssleay†, php†, poco†, postfix†, proftpd†, prosody†, proxytunnel†, pure-ftpd, qpdf†, rabbitmq-c†, radsecproxy†, redis, rhash†, rsync†, ruby†, s-nail†, s2n-tls†, shellinabox†, siege†, sipp†, sipsak†, slowhttptest†, slrn†, socat†, sofia-sip†, squid†, srt†, srtp†, sscg†, ssldump†, sslscan†, sslsplit†, ssmtp†, strongswan†, stunnel†, swtpm†, tcl-tls†, tcpdump†, tinc†, tnftp†, u-boot-tools†, unbound†, vsftpd†, w3m†, wrk†, xilinx-bootgen†, yara†, znc†.

† carried forward: 203 of 209 listed rows are in evidence_reuse.reused_families (committed-rows-carried-forward, candidate_identity_match=false); they were not re-measured at the identity above. The Safe-33 cohort has 27/33 of its families in the same carried-forward set. Not in that list: curl, haproxy, monit, openssh, pure-ftpd, redis; the atlas-level candidate_identity_match=false still applies, so it is not claimed as a fresh measurement at the identity above.

Client, server, library, signing, interception, VPN, mail, DNS, IMAP, IRC, XMPP, FTP, MQTT, Kafka, TPM, SNMP and digest/KDF faces are all represented. Each row records a build against this crate's headers and libraries, linked with -Wl,--no-as-needed so the subject libraries are provably loaded, and exercised by a deterministic local workload with a negative control. The reconciled atlas (forensics/downstream/runtime-functional-atlas.json, body.counts.by_subject and body.evidence_reuse) records candidate 207 / authority 209 L7-functional out of 307 recipe-backed families, with 223 candidate loaded (224 authority), 7 candidate failed (5 authority), 755 not attempted and 693 with no recipe. All 307 recipe-backed rows are carried forward from the superseded identity dc17cfdd…/75cc3979… (evidence_reuse.mode = "committed-rows-carried-forward", families_reused = 307, candidate_identity_match = false); none is a fresh measurement at the identity above, and every carried row names that superseded identity. The last full no-reuse run (S52, 196 of 200 at cd71d401/efd460de) and the S59 re-measurement (198 of 216 at c8a99c59/5992c29d) remain history. The three families lost in S42 (dma, libssh, sslscan) were recovered in S44, so the Safe-33 cohort is 33/33. On the candidate side 100 of the 307 recipe-backed families are below L7 (307 − 207), against 98 on the authority side; the difference is the measured authority-only asymmetry ewftools and xilinx-bootgen. fdm was repaired from ENVIRONMENT in S53, three fixtures were lifted in S60, charybdis was repaired in S61, and 693 families have no admitted recipe.

Memory safety is partial — precisely what that means

This project does not claim to be memory-safe yet. The honest ledger is derived from the receipts:

Status What it covers
Discharged (mechanism scope) 6 — RC-1 shared reference counting; OB-1 OwnedBuf release state; OB-2 OwnedStr string ownership; OB-3 OwnedCtx/RawCtx context ownership; ERR-1 ERR error-queue ownership and allocator file-coordinate seam; GEN-1 GENERAL_NAME selected-arm representation. Mechanism scope only; not a file-level or consumer-level graduation.
Discharged (scoped boundary) 19 receipts — B23 (B23-WPKT-1, B23-WPKT-2), B25 (B25-RL-4, B25-RL-5), B27 (B26-X1), B29 (B29-UI-1, B29-UI-2, B29-UI-3), B45 (B45-ENG-CB-001, B45-ENG-CB-002, B45-ENG-CB-003, B45-ENG-CB-004, B45-ENG-CB-005, B45-ENG-CB-006, B45-ENG-CB-007, B45-ENG-CB-008, B45-ENG-CB-009), B48 (B48-SHA-001, B48-SHA-002, B48-SHA-003, B48-SHA-004, B48-SHA-005, B48-SHA-006, B48-SHA-007, B48-SHA-008, B48-SHA-009, B48-SHA-010, B48-SHA-011, B48-SHA-012, B48-SHA-013, B48-MD5-001, B48-MD5-002, B48-MD5-003, B48-MD5-004, B48-MD5-005, B48-MD5-006, B48-MD4-001, B48-MD4-002, B48-BLAKE2-001, B48-BLAKE2-002, B48-BLAKE2-003, B48-BLAKE2-004, B48-SHA3-001, B48-SHA3-002, B48-SHA3-003, B48-MDC2-001, B48-MDC2-002, B48-RIPE-001, B48-RIPE-002, B48-WP-001, B48-WP-002, B48-SM3-001, B48-SHA-100, B48-SHA-101, B48-SHA-102, B48-SHA-103, B48-SHA-104, B48-SHA-105, B48-SHA-106, B48-MD5-100, B48-MD4-100, B48-BLAKE2-100, B48-BLAKE2-101, B48-SHA3-100, B48-SHA3-101, B48-MDC2-100, B48-RIPE-100, B48-WP-100, B48-SM3-100, B48-SM3-101, B48-NULL-001, B48-NULL-002, B48-NULL-003, B48-METH-001, B48-METH-002, B48-METH-003, B48-METH-004, B48-METH-005, B48-METH-006, B48-METH-007, B48-METH-008, B48-METH-009, B48-METH-010, B48-METH-011, B48-METH-012, B48-METH-013, B48-METH-014, B48-METH-015, B48-METH-016, B48-METH-017, B48-METH-018, B48-METH-019, B48-METH-020), B49 (B49-I2R-001, B49-I2R-002, B49-I2R-003, B49-I2R-004, B49-I2R-005, B49-I2R-006, B49-I2R-007, B49-I2R-008, B49-I2R-009, B49-I2R-010, B49-I2R-011, B49-I2R-012, B49-I2R-013, B49-I2R-014, B49-I2R-015, B49-I2R-016, B49-I2R-017, B49-I2R-018, B49-I2R-019, B49-I2R-020, B49-I2R-021, B49-I2R-022), B50 (B50-I2S-001, B50-I2S-002, B50-I2S-003, B50-I2S-004, B50-I2S-005, B50-I2S-006, B50-I2S-007, B50-I2S-008, B50-I2S-009, B50-I2S-010, B50-I2S-011), B51 (B51-V2I-001, B51-V2I-002, B51-V2I-003, B51-V2I-004, B51-V2I-005, B51-V2I-006, B51-V2I-007, B51-V2I-008, B51-V2I-009, B51-V2I-010), B52 (B52-S2I-001, B52-S2I-002, B52-S2I-003, B52-S2I-004, B52-S2I-005, B52-S2I-006, B52-S2I-007, B52-S2I-008, B52-S2I-009), B53 (src/slh_dsa/adrs.rs:119, src/slh_dsa/adrs.rs:128, src/slh_dsa/adrs.rs:134, src/slh_dsa/adrs.rs:147, src/slh_dsa/adrs.rs:154, src/slh_dsa/adrs.rs:166, src/slh_dsa/adrs.rs:172, src/slh_dsa/adrs.rs:178, src/slh_dsa/adrs.rs:184, src/slh_dsa/adrs.rs:193, src/slh_dsa/adrs.rs:200, src/slh_dsa/adrs.rs:207, src/slh_dsa/adrs.rs:221, src/slh_dsa/adrs.rs:228, src/slh_dsa/adrs.rs:241, src/slh_dsa/adrs.rs:248, src/slh_dsa/adrs.rs:254, src/slh_dsa/adrs.rs:260, src/slh_dsa/hash.rs:179, src/slh_dsa/hash.rs:212, src/slh_dsa/hash.rs:240, src/slh_dsa/hash.rs:272, src/slh_dsa/hash.rs:301, src/slh_dsa/hash.rs:332, src/slh_dsa/hash.rs:392, src/slh_dsa/hash.rs:442, src/slh_dsa/hash.rs:531, src/slh_dsa/hash.rs:558, src/slh_dsa/hash.rs:586, src/slh_dsa/hash.rs:618), B54 (72 ledger-applied boundaries), B55 (10 ledger-applied boundaries), B56 (21 ledger-applied boundaries), B57 (2 ledger-applied boundaries), B60 (2 ledger-applied boundaries), B67 (BIO-READBUF-001, BIO-READBUF-002), B69 (RAND-CTX-001, RAND-CTX-002, RAND-CTX-003), B70 (CMS-FREE-001, CMS-FREE-002). Proposals not applied (not counted): B28, B46, B47, B71. Retained boundaries stay CONTRACTED.
Independently checked (not discharged) B39–B40 boundary-checked (memory-BIO constructor/destructor, connect-BIO descriptor); B41–B43 batch-checked; B44 author-evidenced only.
Uninitialised-storage class 5 UNSOUND fixed (4 owner-shaped + 1 frame-local), from 32 owner-shaped entries examined; 16 safe-by-construction and 12 misuse-resistant. CLOSED for the typed-accessor-over-uninitialised-storage class: both ml_dsa allocation surfaces and the scratch partition initialise before publication; the pkcs12 and bn owning-buffer accessors and the ssl t1_lib peer-sigalgs write no longer form a typed view over uninitialised storage. Retained: ml_dsa Vector::init / matrix_init remain safe constructors over caller raw pointers; they do not dereference, so no UB is reachable alone. Making them unsafe fn is the AGENTS.md section 1 end state, deferred because it touches about 20 call sites; x509::pcy_lib::tree_levels aliasing obligation (owner-derived borrow) remains retained; provider::cipher_gcm out-buffer slice formation was retained here in S22 and is now reconstructed: S40 eliminated getivgen's and gcm_tls_cipher's whole-extent typed slices over caller-writable output (OWNERSHIP_RECONSTRUCTED, not SAFETY_DISCHARGED)
STOP-class defects 0 open — B27-X2 (STOP-CLASS, B27) closed by B28; NEO-ERR-1 (STOP_LEVEL, B33) fixed in B33.
Audited but not converted 761 PARTIAL files; 761/761 audited, 0 remaining. Audit completion is not discharge.
Not discharged Receipt-backed mechanisms still open: BIO, RUNTIME_ALLOC_OWNERSHIP; SAFE_COMPATIBLE remains 1/33.
BOUNDARY / SEAL gates NOT_RUN / NOT_RUN — the S16–S61 cluster has no independent caller/dependency-closure audit, Miri/ASan/Loom/mutation control, authority differential, full forensic regeneration or held-out compatibility.
Known gaps CLOSURE-NON-MONOTONE REPAIRED; S21-O1 CLOSED; S22-RETAINED RETAINED; DOWNSTREAM-IMPLEMENTATION 0 OPEN AT THE MEASURED IDENTITY; DOWNSTREAM-h2o CLOSED; DOWNSTREAM-charybdis CLOSED IN S61 (carried forward); DOWNSTREAM-fastd ENVIRONMENT; DOWNSTREAM-HARNESS-FIXTURE 76 CARRIED S59-S60 BLOCKERS + 25 NEW VENUE/BUILD BLOCKERS; DOWNSTREAM-BELOW-L7 100 FAMILIES; DOWNSTREAM-S42-SYMMETRIC-LOSSES CLOSED IN S44; DOWNSTREAM-partclone CLOSED; DOWNSTREAM-perl-net-ssleay CLOSED; DOWNSTREAM-NO-RECIPE 693 UNMEASURED; DOWNSTREAM-MEDUSA HARNESS; DOWNSTREAM-S53-YIELD 2 NEW L7 / 35 ADMITTED; DOWNSTREAM-S60-YIELD 7 NEW L7 / 30 ADMITTED; S56-RESIDUALS RETAINED; S57-RACE DATA-RACE FREEDOM NOT ESTABLISHED; S58-SECURITY-CALLBACK PERMISSIVE ARMS RETAINED; S61-RESIDUALS OTHER CALLBACK ARMS REMAIN PERMISSIVE; ELEMENT-7-PRECISION OPEN — FAIL; CONTRACTED-OBLIGATIONS-BLOCKED 215 BLOCKED at the applied v1u closure (222 at the historical v1r closure); RELEASE-EVIDENCE-DEBT OPEN; SSL-CONF-WITNESS OPEN — blocks 18 SSL_SESSION sites
SAFE_COMPATIBLE consumers 1 / 33 — the admitted openssh family graduated at b85c via its consumer binaries ssh, sshd; every other admitted family still has no proven dependency closure.

The planned conversion order is attacker exposure × eliminability: TLS state machines and record processing first, then key schedule and targeted AEAD, then the CMS mechanisms behind the unresolved CVE replays, then ASN.1/X.509 decode and verification, then ML-DSA, then the remaining provider and algorithm cores.

Current state: STATUS.md · SEAL-CENSUS.md · PARITY_MODEL.md

Where the memory-safety campaign actually stands (2026-10-30; graduation record 2026-11-17)

2026-11-17 — SAFE_COMPATIBLE issued for ssh and sshd at candidate b85c. evidence/SAFE-COMPATIBLE-openssh-b85c.json records the graduation at the measured OpenSSH consumer closure: 290/290 obligations DISCHARGED, element 7 PASS, L7 PASS including the authority-linked differential. All seven mechanisms are SAFETY_DISCHARGED at that scope — SSL_SESSION, RUNTIME_CONCURRENCY, EVP_PKEY_PROVIDER, X509_VERIFY_POLICY, RAND_CMS_STORE_KEYFMT, C_ABI_CONTRACT and BIO — with condition 3 met where it was raised (X509 69/69, RAND 17/17, EVP external surface) and no memory-safety defect demonstrated for any of them at this scope. An earlier "five mechanisms remain below mechanism-wide grade" label is retracted in the record itself: it was a scope expansion, not a measurement. The two residuals kept are the external provider pointer-type obligation (an ABI contract for a provider that misdeclares its own table) and the focused instrument coverage. The independent adjudicator's NOT ESTABLISHED verdict (evidence/b85c-adjudication.json) stays in the record as history and is not reinterpreted. The generated SAFE_COMPATIBLE counter below now reads 1 / 33: openssh is the one admitted cohort family counted, on the strength of the b85c graduation record (ssh and sshd are that family's two consumer binaries). It read 0 / 33 before that graduation. The bullets below remain the 2026-10-30 record.

Apart from the b85c record above, receipts only: no other graduation is claimed, and nothing here claims constant-time behaviour or that the BOUNDARY/SEAL gates passed. This section was refreshed for the 0.0.44 candidate. 0.0.44 is a packaging/documentation release, not a verification event: nothing below was re-measured at it unless a bullet says so.

  • The consumer closure is ACCEPTED, with zero refusals (evidence/closure-ssh-v1r.json, evidence/inventory-sshd-v1r.json; candidate 0590d371d, committed install libcrypto 23ec6944… / libssl 493795c2…, analysis identity cede10d6…). Both ssh and sshd resolve against the repaired bound: 624 / 623 reachable files, 8,686 / 8,652 reachable functions, 312 reachable production obligations = 90 DISCHARGED / 222 BLOCKED / 0 UNKNOWN, 52,022 edge records with 0 defects and 0 oracle violations, closure rounds 2 and converged, one admission pass with 27 refuted exclusions admitted, and 5 symbolic values resolved, 0 blocking. Acceptance means no refusal condition fired; it is not a graduation and carries no consumer claim. This is the historical v1r measurement. The applied closure is at candidate v1u: accepted, 314 reachable production obligations = 99 DISCHARGED / 215 BLOCKED / 0 UNKNOWN (see the B45 bullet below; the total is 314 because two SSL_SESSION operations in src/ssl/t1_trce.rs are reachable in the relinked binary).
  • The analyzer precision checkpoint reduced the target-bounded set (evidence/base-slot-partial-summary.json): at the pre-engine DSO, reachable target-bounded unresolved sites moved 81 → 71 and base-slot-partial 50 → 39 (10 sites proved), with the reachable-function digest unchanged and obligations unchanged. Element 7 remains FAIL. This checkpoint is not re-measured at 0.0.44.
  • The engine lock precondition is now explicit (evidence/mechanism-discharge-ENGINE-CLEANUP.json, independent verdict in evidence/mechanism-discharge-ENGINE-CLEANUP-review.json): engine_table_cleanup documents and enforces the initialised-lock precondition, its idempotence test initialises the lock, and a new C8 pins the NULL-lock no-op. That package alone discharged nothing; its five review conditions are now closed and the discharge was applied separately as B45 (next bullet), at a candidate where the control and the claim share one identity.
  • One scoped discharge is now applied: B45 (atlas-refresh/boundary-sweep-B45.json, review in evidence/mechanism-discharge-ENGINE-CLEANUP-B45-review.json, application record in evidence/mechanism-discharge-ENGINE-CLEANUP-v1t.json). Nine zero-argument engine-table cleanup thunks in src/engine/tb_*.rs moved BLOCKED → DISCHARGED at candidate v1u (9eccc7e9), for both ssh and sshd, and nothing else moved: the applied closure is 314 obligations = 99 DISCHARGED / 215 BLOCKED / 0 UNKNOWN, ACCEPTED with zero refusals (evidence/closure-ssh-v1u.json, evidence/inventory-sshd-v1u.json; -B45-baseline twins carry the pre-discharge 90 / 224 / 0). The count is 314, not 312, because the relinked binary makes two SSL_SESSION operations in src/ssl/t1_trce.rs reachable; that is an identity effect, not new work. mechanisms.C_ABI_CONTRACT remains UNSAFE_OPEN — nine caller sites are not a mechanism discharge — and engine_list_cleanup (eng_list.rs) and dynamic_data_ctx_free_func (eng_dyn.rs:401) stay BLOCKED. The control behind the boundary is a mutation-sensitive test: removing the acquisition guard makes cleanup_preserves_the_table_when_the_write_lock_acquisition_fails fail. The redundant lock.is_null() guard is recorded as behaviourally undetectable, so no claim rests on detecting its removal. Residuals stay OPEN and recorded: B20-TBL-3 CONTRACTED with B20-R2/B20-R3, Miri/ASan/TSan/dudect NOT_RUN, third-party provider destructor null contract unverified, and B45-D1 (the authority frees and NULLs global_engine_lock; landing engine_cleanup_int must re-open the nine).
  • SSL_CONF is contracted but not dischargeable yet (18 SSL_SESSION sites; atlas-refresh/boundary-sweep-B46.json, -B47.json, reviews in evidence/ssl-conf-boundary-review-2026-10-30.json). Every site now carries a fn-level contract, and two independent reviews still refused both receipts. The reason is the standard, not the code: AGENTS.md §7 requires a precondition to be established by callers or an external ABI contract, and §7/§21 both say documentation alone does not discharge. The authority documents cross-thread exclusivity of SSL/SSL_CTX (SSL_CTX_new.pod:97-99, openssl-threads.pod:19, the QUIC multi-stream guide) but documents nothing for SSL_CONF_CTX or same-thread overlap, and the statements that exist are advisory. An in-use guard was added and the reviews refused the discharge again, correctly: six entry points remain unguarded and a C caller can still free the context from another thread. All 18 sites stay BLOCKED; nothing is applied. Whether the authority's general rule counts as the missing witness is a campaign adjudication, recorded as the SSL-CONF-WITNESS gap, not something to settle by relabelling a residual.
  • Miri was run on ssl_conf, and it bounds its own claim (evidence/miri-ssl-conf-v1.json). All four ssl::ssl_conf tests, including the two-iteration option-list path, pass under Stacked Borrows and Tree Borrows with no UB diagnostic — with cfg(miri)-only pure-Rust stand-ins for libc isspace and strchr, so this is a modified-model result about the Rust code around the boundary, not authentic execution. The control does not fire: restoring the pre-fix borrow shape still passes Miri, because the mutation path returns before the stale shared reference is used. So the earlier borrow fix on this branch is recorded as hardening of a latent borrow-discipline violation, not as a repaired instance of observed undefined behaviour. That correction is in the receipt and in the commit history.
  • Element 7 is FAIL (element_7 in the same receipts): 141 reachable unresolved indirect sites, 81 target-bounded and 60 contract-bounded (46 ABI_CONTRACT_EXPORT_CALLBACK + 14 ABI_CONTRACT_PROVIDER), 0 unclassified. A pre-engine analyzer checkpoint reduced target-bounded to 71 (base-slot-partial 50 → 39) but is not re-measured at 0.0.44. SAFE_COMPATIBLE therefore stays 0/33 on the cohort counter (see the 2026-11-17 b85c record above for the principal-issued verdict): 215 obligations are contracted but BLOCKED at the applied v1u closure (222 at the historical v1r closure), and no family inherits safety from an accepted closure alone (AGENTS.md §7).
  • The analyzer repair is implemented and measured (atlas-refresh/closure-fixpoint-repair.json): the residual admission is now a class-independent superset of every class the resolver may admit (monotone bound), admission is iterated to a least fixed point with an explicit non-convergence refusal, the reachable set is persisted under an analysis identity that is not the DSO digest, every unresolved symbol carries a disposition, and the result is a strict superset of the pre-repair reachable set with CRYPTO_cbc128_encrypt restored. The precision classes moved from 194 → 81 target-bounded sites after the source-backed GOT/static-table and exported-callback proofs. atlas-refresh/closure-monotonicity-finding.json remains the historical finding, superseded by this repair.
  • Seven real defects were found by the campaign's instruments — the static closure analyser, its probes and the downstream corpus — not by a crash in the field:
    1. the non-monotone residual bound (DEFECT-NONMONOTONE-BOUND), which dropped whole target sets under a byte-identical consumer;
    2. the one-pass closure that never scanned a function reached only through residual admission (DEFECT-ONE-PASS-CLOSURE);
    3. unsound oracle inheritance that accepted the narrowed exclusions (DEFECT-ORACLE-INHERITANCE) — all three in atlas-refresh/closure-fixpoint-repair.json;
    4. target sets derived from emitted symbol names instead of stored GOT/static-table values (atlas-refresh/element7-regression-v1j.json, commit 00d7fd8a3);
    5. the legacy SHA-256/384/512/512-224/512-256 NULL md_data write reachable through the exported EVP_MD_meth_get_init (ad5913099);
    6. the same NULL md_data write in MD4, RIPEMD-160 and Whirlpool, now refused where the pinned authority SIGSEGVs (evidence/legacy-md-init-fix.json, D-LEGACY-MD-INIT-MD-DATA-1);
    7. the GCM EVP_CTRL_GCM_SET_IV_FIXED -1 sentinel regression that broke OpenSSH's AES-GCM handshake (commit 49cff7886 rejected the sentinel; 1c2bc4001 forwards SIZE_MAX and copies ivlen, as the authority returns 1). The legacy digest crashes are fixed (5, 6); the fixed-point re-derivation is implemented (1–3).
  • The downstream corpus is green at 0.0.43 (f8f366298, courts/phase17/downstream): all six families — curl, git, haproxy, nginx, openssh, python — pass, RT-DOWNSTREAM-CORPUS passes, and OpenSSH is restored to 31/31 after the GCM fix. The historic S'' freeze (atlas-refresh/freeze-S-double-prime.json) still records the 0.0.42 source/artefact/consumer identity, and published 0.0.42 carries that source.

Known open work

  • Element-7 precision: the accepted v1r closure has 81 target-bounded reachable sites. The pre-engine analyzer checkpoint reduced that to 71 (base-slot-partial 50 → 39), but it is not re-measured at 0.0.44 and the sites still lack a bounded target set, so element 7 remains FAIL.
  • 215 contracted obligations awaiting discharge (BLOCKED) at the applied v1u closure (222 at the historical v1r closure), inside each family's dependency closure; SAFE_COMPATIBLE remains 0/33 on the cohort counter (the b85c record above is the principal-issued verdict for two consumer binaries). Nine engine cleanup callbacks were discharged as the scoped boundary B45; the remaining engine-package sites stay BLOCKED and the mechanism stays UNSAFE_OPEN.
  • Release-evidence debt: FRF's 143 manifests are still 0.0.27; the Phase-18 unsafe-bounds ratchet is +3,017 real sites and needs a reviewable re-base; the Phase-24 candidate freeze and P1000 counts are pinned to 0.0.27; the Phase-25 planes carry 47k per-file hashes and four Miri findings from the stale candidate; 892 prerequisite findings are correctly gated until those strata seal; and the required CI gates remain red (892 prerequisite findings; phases 18/20/23/24/25). They were relaxed and then restored for the 0.0.43 merge; that is recorded as debt, not as a success. The 0.0.44 release is not a verification event.
  • The 13 src/evp/legacy_sha.rs update/final/ctrl callbacks still dereference md_data without a check (atlas-refresh/openssh-v1j-sites.json; BLOCKED).
  • S56/S57 races are not established; no faithful Loom model exists (atlas-refresh/v3-boundary-instruments.json, s56_s57_loom).
  • The court host checkout is bind-mounted read-write at /work (AGENTS.md §6); docker/openssl-rs-court.sh is committed mode 100644, with no executable bit.

Where the conservation strata stand

The twenty-six strata (0–25) and their subjects are in docs/RELEASE_GATES.md §1. Their derived state is generated in forensics/phase-state.md by forensics/tools/phase_state.py and rendered into STATUS.md; that generated file is authoritative over any prose, and this page does not restate which strata are complete. For shape, the strata group as:

strata subject
0–2 the constitution, the archaeology, and the distribution shell — the libcrypto.so.3 / libssl.so.3 / libcrypto.a / libssl.a / legacy.so / openssl artefacts, headers, pkg-config and install tree, with the ABI courts that prove a binary built against the authority runs against the candidate unmodified
3–17 the substantive library strata: the runtime, BIO/CONF, BN/ASN.1/DER/PEM, the provider core, the EVP framework, the native primitives, RAND/DRBG, the key formats, X.509, the protocol families, legacy compatibility, TLS/DTLS, QUIC/ECH, the CLI/config contract, and the downstream replacement court
18 the hostile fuzz / security / side-channel hardening stratum
19–21 performance / CPU dispatch, the 3.6.4 custodian seal, and the maintenance-delta machinery
22 authority exhaustiveness and the whole-program compatibility atlas
23 multitrack authority compatibility and OpenSSL lineage
24 downstream-1000 replacement atlas and the empirical drop-in corpus
25 memory-safety atlas, unsafe trusted-computing-base census, and the historical CVE extinction court

Safe-33 conversion status. The generated blocker block below is frozen at the Phase-24 derivation and is superseded for current counts by the generated achievement block earlier in this page, which derives them from forensics/downstream/runtime-functional-atlas.json. For conversion, discharge, test-count and consumer claims read forensics/memory-safety/campaign/safe-c33/STATUS-2026-10-20.md; forensics/STATUS.md is generated and requires regeneration by forensics/tools/render_status.py before its downstream figures can be quoted. No count is typed in this paragraph deliberately: typed counts go stale, which is exactly what the generated blocks exist to prevent.

Downstream-1000 biggest movers (generated)

Every one of the 1000 counted families, partitioned by its deepest blocker and the classes ranked by mover potential, with the feasible recipe queue, is in the generated report (docs/PHASE-24-BIGGEST-MOVERS.md); the same table is cited by docs/SEAL-CENSUS.md. This block is generated from forensics/downstream/shared-blockers.json — do not edit it by hand.

blocker class families mover potential fixability per-fix leverage
no-admitted-recipe 697 697 recipe-admission 1.0
authority-baseline-limited 90 90 authority-limited 90.0
candidate-configure-failure 28 28 candidate-fix 28.0
no-fixture 9 9 fixture-addition 1.0
candidate-load-failure 4 4 candidate-fix 4.0
candidate-link-failure 3 3 candidate-fix 3.0
candidate-functional-failure 2 2 candidate-fix 2.0
candidate-runtime-failure 2 2 candidate-fix 2.0
recipe-build-system-unsupported 1 1 recipe-build-system 1.0

836 of the 1000 counted families are blocked; 164 are DROP_IN_PASS (819 are DROP_IN_NOT_APPLICABLE, 181 measurable). The drop-in-pass figure is a baseline-normalized verdict count, not a rung: of the 164 drop-in passes, 15 are at L4-linked with no admitted runtime workload for those families and 149 reach L5 or above.

  • a selected population is not a random sample: the 1,000 counted families are selected from frozen ranking evidence, so their blocker shares do not generalise to all downstream software
  • 1000/1000 is not a security proof: a full pass is not a guarantee that any consumer is safe, and the analysis makes no statement about an unmeasured consumer
  • a build is not a functional proof: the configure/link/load rungs are not behaving, and only the functional levels are behavioural evidence
  • direct and transitive consumers are different evidence: the two are never summed
  • a heuristic ranking is not a measurement of buildability: the feasible recipe queue is ranked by frozen breadth signals, and no family in it has been built by this analysis

Compatibility target

The goal is substitution: an unmodified OpenSSL consumer should be able to compile, link, load, execute, and observe the same externally visible behaviour against the stated authority and build profile. The reconstructed distribution artifacts are libcrypto.so.3, libssl.so.3, libcrypto.a, libssl.a, legacy.so, provider modules, public headers, pkg-config metadata, and the openssl executable, with their own symbol namespaces, SONAMEs and version nodes. See CUSTODIAN_CONTRACT.md.

Compatibility is evaluated independently across ABI, semantics, ownership, error behaviour, state transitions, concurrency, correctness vectors, and downstream behaviour. Passing evidence is scoped to the surface actually exercised; unknown behaviour remains unknown.

Verification

Compatibility is tested against the pinned OpenSSL build using differential courts: one probe compiled against the authority and against the candidate, whose transcripts are compared line by line. Construction correctness is a separate plane of published test vectors, and ABI, symbol-ownership, prototype and dispatch planes are checked independently. Results are retained as machine-readable evidence and are not converted directly into compatibility claims — promotion to PARITY_VERIFIED is per dimension, by court.

Which symbols are SCAFFOLDED and which are IMPLEMENTED — per library — is generated in STATUS.md under "Implementation state"; this page types none of those counts. No symbol is PARITY_VERIFIED; that state is promoted only by a court, dimension by dimension (docs/PARITY_MODEL.md).

Generated evidence is authoritative over descriptive prose. Where a document and a generated artifact disagree, the artifact is right.

Evidence and method: PARITY_MODEL.md · REPRODUCIBILITY.md · SEAL-CENSUS.md · DECISIONS.md

Build and test

Authority-bearing execution — differential courts, forensic probes, benchmarks and authority builds — runs only in the isolated, resource-capped court container, never on the host (CUSTODIAN_CONTRACT.md §11). Unit tests and static derivation may run outside it.

# build the pinned court image and start it with OOM protection
bash docker/openssl-rs-court.sh build
bash docker/openssl-rs-court.sh up
bash docker/openssl-rs-court.sh verify

# admit and build the pinned authorities, then generate the atlas
bash docker/openssl-rs-court.sh exec bash forensics/tools/build_atlas.sh

# build and test the implementation crate (the local gate; one heavy command at a time)
bash docker/openssl-rs-court.sh exec sh -c 'cd /work && CARGO_BUILD_JOBS=4 cargo fmt --all -- --check && cargo clippy --all-targets -- -D warnings && cargo test --lib -- --test-threads=4'

# long-lived evidence and sensitivity testing use a second, separately pinned
# tooling container (see docs/REPRODUCIBILITY.md §1.2 for why)
bash docker/openssl-rs-frf-court.sh up
bash docker/openssl-rs-frf-court.sh exec bash forensics/frf/run_courts.sh

The court applies hard resource caps (--memory=8g --memory-swap=8g, --pids-limit=2048, --cpus=8, --restart=no) so a runaway court OOM-kills inside the container rather than pressuring the host.

The local gate is exactly:

cargo fmt --all -- --check && cargo clippy --all-targets -- -D warnings && cargo test --lib -- --test-threads=4

The last full library run on record is 2,231 passed, 0 failed, 52 ignored (cargo test --lib -- --test-threads=4; atlas-refresh/test-baseline.json). It was measured at cc7762e5d during the engine lock repair. It was not re-run for 0.0.44, so for the release tree it is a floor, not a current count. The floor the campaign previously quoted, 1,962 passed / 42 ignored, is superseded by that run; the rule is no unexplained removals and no newly ignored tests, not a frozen literal. 0.0.44 is a packaging/documentation release, not a verification event. The cargo test line above is the library gate only: it is not the differential courts, which run in the court against the pinned authority.

Architecture

One Cargo package. One first-party implementation crate. No workspace, no component crates, and no implementation dependencies: everything the product does is implemented here. Internal modules may be arbitrarily numerous.

Zero runtime and zero build dependencies. The only dependency in the manifest is one cfg(loom) dev-dependency, loom 0.7, enabled only when a test build sets --cfg loom; ordinary builds and ordinary test runs never link it.

The export sets are unchanged across the candidates this page names. Counting rule: nm -D --defined-only <file> | wc -l, which includes the linker-emitted version-node definition symbols and gives 5,906 for libcrypto.so.3 and 608 for libssl.so.3. The owned-export rule, which strips those definitions, gives 5,896 and 603. Both rule totals are identical at v1p and at the 0.0.43 rebuild, with 0 added and 0 removed (atlas-refresh/build-0.0.43.json).

The single implementation still emits the multiple runtime artifacts the distribution contract requires, with their own symbol namespaces, SONAMEs and version nodes preserved. The crate's own type and the link machinery that produces those artifacts are described in Cargo.toml and ABI_POLICY.md; the directory layout of the evidence is documented in REPRODUCIBILITY.md.

Documentation

document governs
CUSTODIAN_CONTRACT.md mission, the single-crate rule, what "custodian compatible" means
PARITY_MODEL.md obligation states, evidence planes, promotion rules, residuals
RELEASE_GATES.md stratum order, maturity levels, court taxonomy
REPRODUCIBILITY.md receipts, determinism, court venue
SEAL-CENSUS.md the generated obligations and courts census
DECISIONS.md the engineering decision record

Additional policies — ABI, provider architecture, ownership and lifetimes, concurrency, FIPS claims, unsafe Rust, security divergences, and non-claims — are under docs/.

Limitations

  • Not FIPS validated. Behavioural parity with the FIPS provider is not validation.
  • No cryptographic-security claim follows from OpenSSL parity. Passing the OpenSSL oracle proves compatibility over the observed surface; it does not prove an implementation is cryptographically sound. Conversely, a perfect standards implementation can still be OpenSSL-incompatible.
  • Memory safety is measured, not established. The memory-safety benefit of Rust is not asserted here: the crate still carries a large unsafe surface, measured per module by forensics/tools/unsafe_footprint.py and rendered in forensics/STATUS.md: 61,454 unsafe sites, 10,980 extern "C" fn, 877 files, 858,173 lines, 89.4% of the unsafe sites in the parser/algorithm core rather than a boundary shim. docs/UNSAFE.md cites the generated table and the UNSAFE-FOOTPRINT growth ceiling; the count is a footprint, not a proof that the unsafe code is correct.
  • No universal claims from finite evidence. Every claim names its authority, profile and platform.
  • Unknown is a research result. UNKNOWN is reported, not traded for confidence.

See NON_CLAIMS.md.

Licence

Apache-2.0 OR MIT. Provenance for any OpenSSL interface material used for archaeology is recorded in NOTICE.

About

Custodian-level native-Rust reconstruction of the externally observable OpenSSL 3.x contract.

Topics

Resources

Stars

50 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages