SuperAPI AI Directory

verus-verification

Coding & development

View repository

Updated 10-23 11:13

About this skill

Regorus Verus Verification

Use this workflow for proof-oriented changes in Regorus, especially Number, BigInt, arithmetic, conversions, and policy-critical value semantics.

The objective is not merely to make Verus pass. The objective is to establish an exact, useful contract for the real executable implementation with the smallest honest trusted boundary.

Core Rules

  1. Specify executable semantics exactly.
    • Model every meaningful result variant and error path.
    • Preserve distinctions such as integer versus float representation.
    • For floating-point operations, specify IEEE-754 behavior rather than ideal real arithmetic.
    • Do not weaken a contract just because the stronger proof is inconvenient.
  1. Prove bodies whenever Verus supports them.

    • Prefer a verified implementation over assume_specification.
    • Remove a trusted assumption once the implementation carries a proved spec.
    • Never describe an external_body function as body-proved.
  2. Preserve executable behavior.

    • Before editing, compare the function with main or the relevant base.
    • Keep executable statements unchanged unless the task explicitly requires a runtime fix.
    • Never use conditional compilation to give Verus and ordinary Rust different executable bodies or behavior. If Verus cannot verify the shared body, retain the narrowest external_body boundary and document the unsupported construct.
    • Put ghost reasoning in proof! blocks. Move proof work to the beginning of the function when it depends only on inputs.
    • Afterward, inspect the focused diff against the base and confirm that only contracts and erased proof code differ, unless a runtime change was intended.
  3. Do not use preconditions to hide valid edge cases.

    • Check minimum signed values, maximum values, zero, and representation boundaries explicitly.
    • Negating i32::MIN as i32 overflows, but its magnitude $2^31$ fits in u32. Widen before negation, for example (-(e as i64)) as u32.
    • If the API can compute a valid result, prove and compute it instead of excluding the input or returning an invented error.
  4. Reuse existing semantic models.

    • Search src/verify/ before adding an uninterpreted spec function.
    • Prefer established models such as pow2, NumberView, to_f64_lossy_ensures, and BigInt view/spec traits.
    • If the same mathematical value can have representation-dependent runtime behavior, quantify over the concrete modeled value rather than pretending the view alone determines the result.
    • When a view deliberately merges concrete variants, use a relational postcondition for representation-sensitive operations. For example, NumberView::Integer merges Int, UInt, and BigInt, whose lossy float conversions and boundary behavior need not be a function of the view alone.
    • Propagate that relation through callers with existential result witnesses. Do not recover hidden representation by existentially inventing a concrete Number whose view matches; that leaks internals and may choose a witness unrelated to the executable receiver.
  5. Minimize and explain trust.

    • Use external_body only at the smallest unsupported boundary.
    • Give an exact postcondition, not merely positivity or successful return, whenever downstream proofs depend on exact behavior.
    • Add a short comment naming the concrete verifier limitation, for example: overloaded <<=/>>= is unsupported, or overloaded ! on external BigInt crashes this Verus version.
    • Avoid broad external wrappers around otherwise verifiable callers.

Workflow

1. Establish the Runtime Baseline

Start with the function, its helper contracts, its callers, and any existing trusted specification.

git show main:path/to/file.rs
rg -n 'function_name|assume_specification|relevant_helper' src tests

Record one falsifiable hypothesis:

  • what the exact behavior should be;
  • which helper contracts it depends on;
  • the cheapest verification or runtime check that could disprove it.

Do not map the whole subsystem before making a small grounded edit.

2. Write the Contract Before the Proof

The contract should answer:

  • Which inputs return Ok, Err, Some, or None?
  • What exact mathematical value is represented?
  • Is the result an integer or float variant?
  • Which rounding, overflow, saturation, or lossy-conversion rule applies?
  • Are multiple concrete representations possible for the same view?

For arithmetic returning Result, avoid vague contracts such as only result is Ok when the exact value is knowable.

Write contracts around semantic inputs first, then state the result with result matches ..., result is None, or result is Err. This is usually clearer than matching every input/result tuple and separately excluding each impossible result variant.

If an abstract view erases representation but the API result depends on it, allow the honest overlap in the postcondition and explain the boundary. For example, at +/-2^53, primitive and BigInt-backed Number values with the same view can legitimately differ between Some and None in an exact-float API.

For BigInt operators, provide exact operator models and prove the caller against them. For division producing a float, model the exact lossy conversions used by the executable code.

3. Remove Redundant Trust

Search for existing assumptions:

rg -n 'assume_specification.*function_name|uninterp spec fn' src/verify src

When moving a spec onto a body-verified function:

  • delete the old assume_specification in the same change;
  • ensure no duplicate specification remains;
  • strengthen helper contracts only as much as the body proof requires.

An external helper may remain trusted when Verus cannot translate its syntax, but its contract must expose all facts needed by verified callers.

4. Keep Proofs Separate From Execution

Prefer this shape:

pub fn operation(input: i32) -> Result<Number> {
    proof! {
        // Input-only lemmas, cast equalities, and arithmetic facts.
    }

    // Original executable body.
}

Use local proof blocks later only when facts genuinely depend on an executable value produced at that point.

Do not introduce executable temporaries solely to help a proof. If a temporary is ghost-only, keep it inside proof!.

5. Handle Casts and Boundaries Explicitly

Verus often needs explicit facts connecting machine integers and mathematical integers/naturals:

assert((e as u32) as nat == e as nat);

For negative signed values, widen before negating:

let magnitude = (-(e as i64)) as u32;

Then prove:

  • the magnitude is positive;
  • its cast equals the intended mathematical magnitude;
  • required power/division lemmas apply;
  • remainder is nonzero when the runtime should choose floating division.

Check memory implications separately. A mathematically valid BigInt may be very large. Prove extreme paths, but do not execute resource-heavy regression tests unless the cost is acceptable and intentional. Test the conversion and a smaller representative behavior instead.

For signed division and remainder, model Rust semantics with rust_div and rust_rem; mathematical / and % do not capture truncation toward zero for all negative inputs. Bridge primitive operator specs such as RemSpec to those models with focused lemmas. Handle MIN / -1 before either / or %, because both machine operations overflow, and prove the exact quotient fits before connecting a mathematical result to checked_div or a narrowing cast.

6. Use Verification Attributes Deliberately

  • #[verus_verify] on an impl applies to all methods in that impl.
  • Do not split adjacent inherent impls merely to change verification scope when one impl-level annotation plus narrow method overrides is clearer.
  • Use #[verus_verify(external)] only when an item must remain entirely outside verification and has a separate specification.
  • Use #[verus_verify(external_body)] when Verus should trust a stated contract but cannot verify the implementation body.
  • Method-level attributes can override the impl-wide default.

Before diagnosing missing internal markers or macro bugs, inspect braces and attributes. Confirm the method is actually inside the annotated impl.

7. Preserve Production Macros

Do not replace bail!, anyhow!, or formatting in the executable body merely to make translation easier. Inspect the macro expansion and specify the smallest unsupported pieces. For anyhow!, this may mean narrow specifications for Arguments::from_str, format_err, and must_use; if verified callers only rely on taking the error branch, those assumptions need not promise anything about the error value.

After a Verus upgrade, retry the original macro and previously externalized bodies. Translation support changes, so stale shims and external_body annotations should not become permanent trusted surface by inertia.

Diagnosing Verus Failures

Trigger Failure

Before repairing or replacing a rejected trigger, check whether the quantifier is semantically necessary. If its bound variables merely name fields of fixed arguments through equalities such as lhs == NumberView::Integer(integer_lhs), match on those arguments and state the branch-specific condition directly. This preserves the contract while removing the quantifier, its trigger, and needless solver instantiation. Use a natural or artificial trigger only when the contract genuinely ranges over multiple values that are not determined by fixed inputs.

Translation or Compiler Failure

  1. Reduce to the exact operator, type, attribute, and impl context.
  2. Test a one-file reproducer with the same relevant structure.
  3. Do not introduce macros, missing impl annotations, or different ownership patterns unless they exist in the failing code.
  4. If a small candidate passes, it is not a reproducer. Keep reducing the real context or state that the failure was caused by local annotation structure.
  5. Inspect ~/verus only after the local code path is understood.

A valid verifier bug report must:

  • fail on the stated Verus version;
  • contain no unrelated repository dependencies when avoidable;
  • reproduce the same failure mechanism;
  • document any workaround retained in Regorus.

Proof Failure

Treat the first focused failure as evidence:

  • failed arithmetic safety means the implementation has an unhandled machine boundary or needs a justified precondition;
  • failed postcondition may indicate a missing helper fact, a representation mismatch, or an incorrect contract;
  • unsupported library internals should be isolated in the narrowest helper, not used to externalize the verified caller.

Do not respond to a failed proof by immediately weakening the postcondition. First trace a concrete input through the runtime behavior.

Validation

After the first substantive edit, immediately run the narrowest check:

cargo verus verify \
  --fwd-verus-args-to roots -- --verify-module number

Use a fresh target directory when checking for stale macro or compiler behavior.

After the focused proof passes:

cargo test focused_test_name
cargo fmt --all -- --check
git diff --check

For broader or final validation, use repository commands as appropriate:

cargo xtask fmt
cargo xtask clippy
cargo xtask ci-debug

Report verification counts accurately. Distinguish:

  • body-verified functions;
  • external-body contracts;
  • trusted assumptions;
  • runtime tests actually executed;
  • extreme tests skipped due to resource cost.

Completion Checklist

  • Contract matches exact executable semantics.
  • Integer/float and Undefined distinctions remain intact where relevant.
  • Minimum/maximum signed values and casts were considered.
  • Original executable body is preserved unless a runtime bug was fixed.
  • Proof-only code is inside proof! and placed early when possible.
  • No redundant uninterpreted helper or trusted assumption remains.
  • Every external_body has the narrowest useful exact contract and a reason.
  • Impl-level verification annotations cover the intended methods without unnecessary splits.
  • Any claimed verifier reproducer is representative and independently fails.
  • Focused Verus verification passes.
  • Relevant runtime tests, formatting, and diff checks pass.