Spec-ID: x07.spec.internal.x07-core@0.1.0 Status: draft Applies-to: toolchain >= v0.0.95 Related schemas: []
X07 Core (Track B): build a compiler, not a VM
X07 is an experiment in discovering programming languages optimized for LLMs to write, not for humans to read.
The feedback loop is circular:
- A language is only “good” if an LLM can reliably write correct programs in it.
- You only learn what the LLM needs by observing successes/failures across many tasks.
This repo makes that loop explicit, deterministic, and measurable.
Language and programs (Track B)
The core artifact is the compiler + its embedded language definition (versioned by lang_id).
In Track B:
- The language is defined by core built-ins plus built-in stdlib modules (no macro/rewrite layer).
- The compiler (
x07c) turns the program into C source. - The runner compiles the C source to a native executable solver artifact.
- The runner executes the solver deterministically with strict fuel + memory limits.
Seed language references:
- Guide snapshot:
docs/spec/language-guide.md - Default benchmark bundle:
labs/benchmarks/bundles/regression.json - Optional suite:
labs/benchmarks/solve-pure/phase4-holdout.json
Determinism and safety
To keep fitness stable and evaluation safe:
- Evaluation uses deterministic fuel-based halting for solver runs.
- Capability access is deny-by-default and world-scoped (
solve-pure,solve-fs,solve-rr,solve-kv,solve-full). - Solver artifacts run with a fixed environment (no inherited args/env) and Unix resource limits as kill-switches.
Implementation details: docs/spec/internal/x07-c-backend.md.
Regression suites
Benchmark suites live under labs/benchmarks/ and are executed against committed reference solutions in labs/benchmarks/solutions/.
To run the default regression bundle:
python3 labs/scripts/bench/run_bench_suite.py --suite labs/benchmarks/bundles/regression.json