Correctness is established by proof rather than by measurement. The obligation is discharged once, and re-verified automatically on every build.
The Futamura projections establish that interpreters, compilers, and compiler generators are not different kinds of artifact. They are points on one continuum, connected by specialization, a result set out by Yoshihiko Futamura in 1971. LOGOS acts as a single executable specification at every level, so each lowering step is a transformation between representations of the same specification rather than a translation between unrelated languages.
Source
LOGOS, the executable specification
Proof obligation dischargedsemantics preserved
Interpreter
Reference semantics, software
Proof obligation dischargedsemantics preserved
Register VM
Sequential execution model
Proof obligation dischargedsemantics preserved
Dataflow tiles
Spatial placement, physical fabric
The tip lands on the chip.
ProvenZero counterexamples
Every lowering discharges its proof obligation in our proof kernel, and the 8x8 run of the int8 matmul primitive matches an independent reference oracle, checksum 510,720. The kernel behind that proof finishes families Z3 and Kissat do not.
Four vectors that fail independently
Correctness is established by four independent validation vectors, one at each layer of the lowering above, all served by one toolchain. The independence is the point: each can fail without the others noticing, so agreement across all four is evidence no single method provides.
LOGOS Supercompiler
one language, English. one tool, LOGOS.
Specify
One executable source
Compile
Placement onto the fabric
Prove
Obligations discharged
Check
Re-derived independently
Layer 01.
Source
LOGOS, the executable specification
checked by Formal SAT solving
Exhaustive across the primitive's input boundaries
from the toolchain: the specification parsed to logic, and the obligations the lowering opens
Layer 02.
Interpreter
Reference semantics, software
checked by Reference oracle
Continuous comparison, independent implementation
from the toolchain: the interpreter's run of the same source, held against a separately written implementation
Layer 03.
Register VM
Sequential execution model
checked by Soft core RTL
Simulated against the same reference
from the toolchain: the same program in its sequential form, and the reference to hold it to
Layer 04.
Dataflow tiles
Spatial placement, physical fabric
checked by Physical fabric
Executed on hardware, results read back over UART
from the toolchain: the placement onto the fabric, flashed over JTAG
one language, one toolThe specification is written in English. LOGOS carries it down through every layer, and the meaning does not change on the way. What every layer computes is the int8 quantized matmul primitive: the same program, in another form.
agreementFour layers, four independent readings of one primitive. The methods share no failure mode, so agreement across them is evidence that no single method establishes on its own.
For the int8 quantized matmul primitive, lowering preserves the semantics of the reference, verified across four independent vectors. Manufacturing signoff comes back with zero DRC, LVS, antenna and routing violations under Sky130 signoff parameters.
Certificates in milliseconds, not Z3
An obligation is only worth stating if it can be discharged inside a build. LOGOS carries its own prover, written in Rust with no Z3 dependency, and it runs the same in a browser as it does in CI. Every verdict it returns is a certificate that is re-checked, not a result taken on trust.
On structured problems it's not competing on tuning. Resolution based solvers inherit a lower bound proved by Haken in 1985: certain families require proofs of exponential size, so no amount of engineering finishes them. The prover works in proof systems where those families are small, which is a difference in kind rather than in speed.
Measured on an Intel Core i9-14900K against byte identical input, median of five runs, with each competing solver writing its proof to disk. Kissat 4.0.4, CaDiCaL 3.0.0, CryptoMiniSat 5.11.21.
The same benchmark set includes a control: random 3-SAT, where there is no structure to exploit.
Source: LOGOS prover benchmarks, full families and methodology, logicaffeine.com(opens in a new tab)
| Solver | Pigeonhole, 16 holes |
|---|---|
| LOGOS | 13.7 ms, with a 25 KB certificate |
| SaDiCaL | 56.5 ms, with a 242 KB proof |
| Z3 | did not finish within 10 s |
| Kissat, CaDiCaL, CryptoMiniSat | did not finish within 15 s |
Formally verified compilation is an established qualification credential. CompCert, a formally verified C compiler commercialized by AbsInt, is used to earn certification credit under DO-178C in avionics. The proposition that a compiler can carry a proof is settled. Applying it to spatial dataflow architectures is the contribution here.
Source: CompCert, AbsInt Angewandte Informatik(opens in a new tab)
Does this apply to the silicon you are building?