Carcara is a proof checker and elaborator for SMT proofs in the Alethe format.
To build Carcara, you will need Rust and Cargo 1.67 or newer. Build the project with cargo build
.
When running on large proofs, we recommend compiling with optimizations enabled: cargo build --release
.
To build and install Carcara, run cargo install --profile release-lto --path cli
. This will build
the project with all optimizations enabled, and install the CLI binary in $HOME/.cargo/bin
.
To check a proof file, use the check
command, passing both the proof file and the original
SMT-LIB problem file.
carcara check example.smt2.proof example.smt2
If the problem file name is exactly the proof file name minus .proof
, you can omit it:
carcara check example.smt2.proof
By default, Carcara will return a checking error when encountering a rule it does not recognize. If
instead you want to ignore such rules, pass the --skip-unknown-rules
flag.
The --strict
flag will enable a "strict checking" mode. See the strict
checking section for more details.
See carcara help check
for more options.
You can elaborate a proof file using the elaborate
command.
carcara elaborate example.smt2.proof example.smt2
This command will check the given proof while elaborating it, and print the elaborated proof to
standard output. The --print-with-sharing
flag controls whether the elaborated proof will be
printed using term sharing.
By default, elaboration of lia_generic
steps using cvc5 is disabled. To enable it, pass the
--lia-via-cvc5
flag. You will need to have a working binary of cvc5 in your PATH.
Many of the same flags used in the check
command also apply to the elaborate
command. See
carcara help elaborate
for more details.
The bench
command is used to run benchmarks. For example, the following command will run a
benchmark on three proof files.
carcara bench a.smt2.proof b.smt2.proof c.smt2.proof
The command takes as arguments any number of proof files or directories. If a directory is passed,
the benchmark will be run on all .proof
files in that directory. This command assumes that the
problem file associated with each proof is in the same directory as the proof, and that they follow
the pattern <filename>.smt2
/<filename>.smt2.proof
.
The benchmark will parse and check each file, and record performance data. If you pass the
--elaborate
flag, the proofs will also be elaborated (though the resulting elaborated proof is
discarded).
The benchmark results are simply printed to the screen by default. Instead, if you pass the
--dump-to-csv
flag, they will be recorded in two csv files, runs.csv
and by-rule.csv
.
By default, Carcara will check/elaborate each file only once. You can increase the number of runs
using the -n
/--num-runs
option. By default, all benchmarks are run on a single thread. You can
enable multiple threads using the -j
/--num-threads
option.
See carcara help bench
for more options.
Strict checking mode can be enabled by using the --strict
flag when checking. Currently, this only
affects a few rules.
For the assume
and refl
rules, if strict checking is enabled, the implicit reordering of
equalities is not allowed in those steps.
For the resolution
and th_resolution
rules, if strict checking is enabled, the steps must
provide the resolution pivots as arguments. The expected format is that, for each binary resolution
step, two arguments must be provided: the pivot used, and a boolean argument indicating whether the
pivot is in the left-hand (true
) or right-hand (false
) clause. For example:
(step t1 (cl p q) :rule hole)
(step t2 (cl (not q) (not r)) :rule hole)
(step t3 (cl (not s) r) :rule hole)
(step t4 (cl (not (not s))) :rule hole)
(step t5 (cl p) :rule resolution :premises (t1 t2 t3 t4)
:args (q true r false (not s) true))
The intended invariant of strict checking is that any proof that has been elaborated by Carcara can be checked strictly. Strict checking may also improve perfomance.