Derivative Forge v0.2.0
Derivative Forge v0.2.0
Release date: 2026-08-22
Summary
This release turns Forge from an initial typed pipeline into an evidence-grounded, fail-closed software synthesis system for greenfield Python artifacts.
Forge compiles natural-language requirements into traceable contracts, plans through the Derivative substrate, generates deterministic or model-backed candidates, validates them in an isolated environment, performs bounded evidence-targeted repair, and packages only fully validated artifacts.
Highlights
- Atomic requirement preservation from source text through assertions and validation evidence.
- Typed acceptance, obligation, public import, implementation blueprint, and quality contracts.
- Deterministic CLI, REST service, data pipeline, and library adapters.
- Composable capability renderers with provenance and manifest enforcement.
- OpenAI-backed complete-candidate compilation for uncovered plans in
hybridandremote-onlymodes. - Three independent validation layers: runtime, contracts, and adversarial checks.
- Requirement-level semantic assertion evidence and stronger anti-stub detection.
- Bounded repair driven by observed failure signatures and artifact deltas.
- Docker execution policy with no network, read-only root, bounded resources, and secret isolation.
- Fail-closed packaging with explicit
verified,validation_failed, andinfeasible_proventerminal states. - Oracle contract preflight that rejects incoherent external harnesses before Forge or model execution.
- Capability-specific dependency profiles and lazy scientific/model runtime loading.
- A reorganized README, technical reference, and benchmark evidence ledger.
Evidence Checkpoint
The release repository passes 470 tests with 2 skips locally and is gated by minimal-runtime and full Docker-backed GitHub Actions jobs.
Blind V7 remains an honest limitation, not a marketing score:
- frozen baseline external Verified@1: 0/6 (
0.000); - frozen baseline legacy false-verified field:
0.000with denominator 0; schema-v3 adjudicated valuenull; - targeted post-fix replay:
0/5externally accepted; - one targeted oracle was independently invalid;
- four cases remained genuine fail-closed candidate-generation failures.
Known V5, V6, and V7 requirements are regression corpora after first execution. Generality must be measured on a newly frozen schema-v3 blind bundle.
See Benchmark Evidence for immutable receipts, hashes, denominators, and replay labels.
Scope
This release supports greenfield Python CLI, service, pipeline, and library synthesis.
It does not claim:
- arbitrary existing-repository modification;
- arbitrary language or frontend generation;
- formal verification of general software;
- proof of universal properties beyond implemented contracts;
- wheel/container distribution, SBOM, or supply-chain attestation.
Upgrade Notes
The minimal deterministic host is installed with:
python -m pip install -r requirements/forge.txtModel-backed compilation and repair require:
python -m pip install -r requirements/model.txtProduction verification requires the sandbox image:
docker build --file Dockerfile.forge-sandbox --tag derivative-forge-sandbox:py311 .requirements.txt remains a compatibility alias for the complete requirements/all.txt environment.