THE DEFECT, as found. GPT-5.6 showed the textual classifier could report
PROVEN for a name Lean resolves to an axiom: a definition appearing only
inside a `/- -/` comment was read as real. Repairing that exposed something
larger and already live. Aeneas wraps long declarations:
axiom
curve25519_dalek.edwards.EdwardsPoint.Insts.CoreOpsArithNegEdwardsPoint.neg
The old scanner required keyword and name on one physical line, so it matched
nothing there and SILENTLY DROPPED the declaration — no MODEL row, no PROVEN
row, no failure. Nine to ten externals per fork had no row at all, and the
tier-A/B `neg` was missing from every committed table while the trust
documents claimed that class was machine-checked. Four pinned tables were
wrong, in four repositories, with the buttons green over them. No attacker
was required; Aeneas's own formatting did it.
The lesson is not "write a better regex". The scanner was FAIL-OPEN: input it
could not parse produced silence instead of a stop. A gate that drops what it
cannot read is worse than no gate, because the green covers a gap that is
invisible in the diff.
THE REPAIR, in three layers that each do only what they honestly can.
· DISCOVERY stays textual, because the template cannot be imported — it
declares the same names as the model and they would clash. It now strips
NESTED `/- -/` blocks, reads a name that wraps to a later line, tracks
namespace/section/end, and FAILS CLOSED: an unparseable keyword exits 2
naming file and line.
· Phase 0d keeps the cheap pre-compile comparison against the table.
· Phase 2d is new and authoritative. After compilation it generates a temp
Lean driver (the Phase 2b idiom, so no permanent module and no manifest
churn), imports every non-template gen module, and asks env.find? what
each external IS: kind and originating module. Model module -> MODEL;
non-axiom in a generated module -> PROVEN; an axiom outside the model, or
a name that does not resolve, is a failure. Lean's verdict must equal the
committed table's, and the answer count must equal the question count so
a truncated resolver cannot pass.
THE TABLES were regenerated and verified three ways before being committed:
an independently written second scanner agreed on 228/228 rows across the
four forks; all 47 changed rows carry file:line evidence; and the PROVEN row
was put to Lean directly — kind = def, module = CurveField.Funs.
dalek 54 -> 64 rows anza 42 -> 51 risc0 48 -> 57 betrusted 47 -> 56
Also corrected: `from_residual` was recorded EXTRA in every fork; the
template does demand it, so it is MODEL. anza carried a spurious `hash|EXTRA`
row from a mis-parse.
NEGATIVE-TESTED. With Phase 0d neutered so Phase 2d was the gate under test,
a PROVEN row edited to MODEL produces:
SEMANTIC DRIFT: ...neg — table says MODEL, Lean says PROVEN
selftest-correspondence.sh grew 4 cases -> 7: GPT's comment counterexample,
the wrapped declaration that was actually live, and a fail-closed check.
WHAT THIS DOES NOT CLOSE. Phase 2d resolves names in the environment built
from gen/. It establishes what a name IS — assumption or proof, and where —
not that the extraction faithfully represents the Rust. That remains the
Charon/Aeneas trust assumption and no gate here can close it.
Certified: both buttons, all four forks, purged trees. 8/8 green, with
62/48/53/52 externals resolved by Lean and every verdict matching.
|
||
|---|---|---|
| verification | ||
| .gitignore | ||
| README.md | ||
| TRUSTED-BASE.md | ||
dalek-ed25519-verified
Formal verification of the ed25519 implementation in dalek-cryptography/curve25519-dalek (upstream, v5.0.0-rc.1), built as a coherent proof pyramid in Lean 4 via the Charon/Aeneas transpilation pipeline:
┌──────────────────────────────┐
│ Signature (EdDSA verify) │ accepted ⇔ decompress(R) = [k](−A)+[s]B
├──────────────────────────────┤
│ Scalar arithmetic mod ℓ │ Scalar52 ops correct mod ℓ
├──────────────────────────────┤
│ Group law (twisted Edwards) │ point ops = complete addition law
├──────────────────────────────┤
│ Field 𝔽_p, p = 2²⁵⁵ − 19 │ FieldElement51 ops correct mod p
└──────────────────────────────┘
Every layer states its theorems about the actual Aeneas-transpiled Rust
code (never about a hand-written re-model), and every claim in the status
table below is backed by a compiled proof plus an axiom audit of the named
certificate. Files that do not compile under verification/check.sh are not
in this repository.
Layer status
| Layer | Certificate | Status | Axioms of certificate |
|---|---|---|---|
| Field 𝔽_p | fieldImplementation |
✅ proven | [propext, Classical.choice, Quot.sound] |
| Group law (Edwards) | edwardsImplementation |
✅ proven | [propext, Classical.choice, Quot.sound] |
| Scalar mod ℓ | scalarImplementation (add ✅ sub ✅ mul ✅) |
✅ proven | [propext, Classical.choice, Quot.sound] |
| Signature (EdDSA) | verify_accepts_iff … verify_accepts_iff_decompress (4 tiers) |
✅ proven (phases 1+2) | standard three + the button-enforced SHA-512/wire-format boundary — see The signature apex |
Status legend: ✅ proven & axiom-audited · ⏳ in progress · ❌ not started.
This table is updated only when verification/check.sh passes for the layer.
The signature apex (phases 1 and 2)
The byte-level apex certificate CurveFieldProofs.verify_accepts_iff is the
literal EdDSA acceptance criterion, proven about the extracted verifier:
For a signature that parses, the verifier returns
Ok(())iff the recomputed compressed pointcompress([s]·B − [k]·A)equals the signature'sR, byte-for-byte — wherekis whatever scalar the opaque SHA-512 oracle produces from(R, A, msg).
The recomputation runs entirely through the proven model: the vendored ed25519-dalek verify glue is extracted as gen/CurveSig, whose
hand-maintained externals import gen/CurveField — every curve and scalar call
resolves by fully-qualified name to a proven definition. Only SHA-512 (three
stateful wrapper calls) and the wire-format types stay opaque.
check.sh has a dedicated audit phase (Phase 3b) that fails the build unless
the apex certificate's axiom cone is exactly
[propext, Classical.choice, Quot.sound] + {ed25519.Signature, sha2.Sha512, verifying.sha512_new, verifying.sha512_update, verifying.sha512_finalize_bytes, ed25519.Signature.to_bytes, signature.error.Error, signature.error.Error.new}
— i.e. the three Lean foundations plus the documented SHA-512/wire-format
boundary. Zero curve, scalar, or backend axioms. The companion certificate
verify_loop_full (the 32-byte comparison loop computes array equality)
carries the standard three axioms only.
Phase 2 (complete): the point-level lift. Phase 3b enforces the SAME axiom boundary on three further tiers that lift the byte equation to points:
| Tier | Certificate | Statement |
|---|---|---|
| half-lift | verify_accepts_iff_point |
accepted ⇔ R = the canonical encoding of [k]·(−A) + [s]·B (compress semantics + to_bytes canonicity + hash-to-scalar, recompute chain inverted) |
| point equation | verify_accepts_iff_point_eq |
for any valid on-curve Q canonically encoded by R: accepted ⇔ Q = [k]·(−A) + [s]·B as points (encoding-injectivity: d non-square + parity root-selection) |
| full lift | verify_accepts_iff_decompress |
R decompresses to a valid on-curve Pt, and accepted ⇔ Pt = [k]·(−A) + [s]·B — the constructive capstone |
The full lift runs through the extracted CompressedEdwardsY::decompress
itself, proven end-to-end: from_bytes parses the y-residue exactly below
bit 255 (from_bytes_spec), sqrt_ratio_i returns the even square root of
(y²−1)/(dy²+1) (sqrt_ratio_i_sq_spec, Fermat-exponent square root), and
the sign bit selects the x-parity (decompress_of_canonical, standard three
axioms). Byte comparison ↔ encoding equality ↔ point equality ↔
decompressed-point equality: every link is machine-checked over the
extracted code, and check.sh fails the build if any of the four tiers'
cones deviates from the boundary above.
Source
- Upstream: dalek-cryptography/curve25519-dalek, commit
4cf8db2 - Pinned/patched source: saymrwulf/curve25519-dalek-source, commit
aa0f6ab(adds the decompress step_2 negate-then-assign patch) - Patches: minimal Aeneas-compatibility only (documented in the source repo)
- Verified backend:
backend/serial/u64(FieldElement51,Scalar52). SIMD/AVX backends are out of scope (marked opaque).
Toolchain (pinned)
| Component | Version |
|---|---|
| Aeneas | bf13c42e |
| Charon | 9dd7f23c |
| Lean | v4.30.0-rc2 |
| OCaml | 5.3.0 |
Reproducing
source ~/aeneas-toolchain/env.sh
cd verification
./extract.sh # Rust → LLBC → Lean (regenerates gen/)
./check.sh # compiles EVERY shipped file + axiom-audits EVERY certificate
The gen model is ONE merged universe (gen/CurveField: field + curve +
scalar + the verify path's reachable code), regenerated in full by
extract.sh. The scalar layer keeps its own check button:
./check-scalar.sh # compiles the merged gen + all scalar proofs (add, sub,
# Montgomery mul, byte-parsing) and kernel-audits the
# scalar certificates, incl. the scalarImplementation
# aggregate
Trusted base
See TRUSTED-BASE.md for the complete list of assumptions (Lean kernel, mathlib, Charon/Aeneas semantics, external-function models, and — in the signature layer only — an opaque SHA-512 model).