The current M2 flow library has 357 approved blocks at
data/blocks/current/blocks.json. The earlier CRUD and query library has 16 blocks indata/blocks/v1/blocks.json. The original 14-block file is preserved indata/blocks/crud-v1/blocks.jsonfor historical model checks. Its six date-filter blocks are stored indata/blocks/v4/blocks.json. The older corpus status below describes research tooling. The production CLI and HTTP service are in som-rs.
This document describes the support contract in the current repository source. It does not claim that every listed gate ran in this working session.
The rows are the corpus surfaces: the oracle families and their harness, the curated rationale and its check, the code-quality gate, the snippet ISA and its verifier, the project's own test suite, and its packaging. Use the README for the workflow and capabilities. Use the roadmap for future outcomes and explicit non-goals.
| State | Meaning |
|---|---|
| Supported | The current source has a public contract, an implementation, and a named executable gate for the stated scope. |
| Limited | The current source supports the stated scope, but the Limits cell names a material boundary. |
| Not supported | The behavior is not part of the current product contract. The Evidence cell points to a future outcome or a non-goal. |
| Surface | ID | State | Supported scope | Limits | Evidence |
|---|---|---|---|---|---|
| Executable oracle corpus | oracle-corpus |
Limited | The harness discovers every family under data/curated/families, runs each candidate against its fixture, and exits 0 only when every gold exits 0, every near miss fails on exactly the fixture tests its caught_by declares, and every declared oracle is the fixture the harness ran; the current run evaluates all 101 families with every gold and every near miss passing that contract. |
Families 08-asyncio-concurrency-limiter, 09-pydantic-field-cross-validation, and 10-security-timing-constant-auth have no fixture under data/curated/fixtures; the harness prints that no test file matched, skips them, and still exits 0, so their candidates are unproven. See Oracle fixtures complete. |
uv run --project som-code-python python som-code-python/scripts/verify_harness.py |
| Curated rationale | curated-rationale |
Limited | Every family.json carries rationale.teaches, rationale.why, and oracle, every near miss carries why_wrong and caught_by, and every family carries a caption of 80 to 400 words and a decompiled block measured from the gold AST; the check refuses a caption that omits a raised exception, declared status code, or imported library, a decompiled block that drifted from the source, a missing or trivial field, a why_wrong that repeats its failure mode, a caught_by naming a test the fixture does not define, a label word in candidate source, and a comment or docstring that differs across the six candidates; the current run reports 0 defects over 101 families. |
Families 08-asyncio-concurrency-limiter, 09-pydantic-field-cross-validation, and 10-security-timing-constant-auth have no fixture, so their oracle and every caught_by under them is null and their why_wrong sentences are unmeasured prose. See Oracle fixtures complete. |
uv run --project som-code-python python som-code-python/scripts/verify_curation.py |
| Code quality | code-quality |
Supported | Every gold candidate passes ruff check with E,W,F,I,UP,B,SIM,N,PL,RUF, ruff format --check, mypy --strict, and pylint with zero messages, and every near miss passes ruff format --check and ruff check --select E,W,I,UP, all with the tool tables in pyproject.toml; the current run reports 0 defects over 101 gold candidates and 505 near misses. |
The configuration ignores PLR2004, whose literals are the contract the fixture asserts, allows FastAPI and Typer parameter calls in defaults, accepts the X feature-matrix name, disables pylint's too-few-public-methods, duplicate-code, and not-callable, and lets mypy skip scikit-learn, which ships no types; each exception carries its reason in pyproject.toml. |
uv run --project som-code-python python som-code-python/scripts/verify_quality.py |
| Decompiled layer records | decompiled-layer-records |
Limited | Every family.json carries an authored plan and a decompiled block with topology, ops, and coverage measured from the gold; som assemble rebuilds each gold from its ops to the same AST and the family's fixture passes on the rebuilt file, and the multi-file TODO fixture under tests/fixtures/multi_file_todo/ round-trips across its files; the curation check refuses drift and a plan that omits a raised exception or status code; 100 of 101 rebuilt golds and all 5 multi-file sources are byte-identical. |
259 of 400 blocks are still INSERT_BLOCK literals; the snippet share is recorded under snippet-isa. 06-sqlalchemy-atomic-order-rollback has 21 blocks, over the 20-block limit, so som assemble refuses it as SCOPE_TOO_LARGE and it is not round-tripped. 08, 09, and 10 have no fixture and are checked on AST alone. The scope limit is reported by the round trip, not enforced by curation. See Snippet long tail. |
uv run --project som-code-python python som-code-python/scripts/verify_roundtrip.py |
| Snippet ISA | snippet-isa |
Limited | 17 snippets across fastapi, pydantic, sqlalchemy, and typer, each cut from gold text, express 141 of the corpus's 400 blocks as INSERT_SNIPPET in 20 of 101 families; model classes, SQLAlchemy tables, and a create route's keyword list are one snippet each through a list section, and every match renders back to its gold block byte for byte. The verifier refuses a snippet used by fewer than two families, a template line that is only a placeholder, a malformed or nested section, an operation naming an id the ISA lacks, and a first use that does not render to Python. |
259 blocks remain literals: route and function bodies with branching logic, exception handlers, CLI commands, test instrumentation such as query counters, and the code of libraries only one family uses; an idiom seen in one family is kept literal by the two-family rule. See Snippet long tail. | uv run --project som-code-python python som-code-python/scripts/verify_snippets.py |
| Project test suite | python-test-suite |
Supported | tests/ holds the snippet verifier's adversarial cases, one per rule it refuses plus a clean ISA and the committed one, which run under the project virtualenv and pass: 14 passed. |
None. | uv run --project som-code-python python -m pytest som-code-python/tests -q |
| uv-runnable packaging | uv-runnable-packaging |
Supported | uv sync --project som-code-python resolves from uv.lock and installs every third-party library the candidates and fixtures import, plus ruff, mypy, pylint, and their stub packages, into the project .venv; every gate here runs through uv run --project som-code-python. |
The project is not a package ([tool.uv] package = false) and declares no console script; the engine's som command comes from som-core. |
uv sync --project som-code-python |
- A Supported or Limited row names the command that measures it, run from the
repository root through
uv run --project som-code-python. - A Limited row's Limits cell states the boundary as an observation and names every tolerated failure verbatim, never as a count, and links the ROADMAP outcome that removes it.
- A Not supported row links exactly one ROADMAP outcome or non-goal and claims no gate.
- The harness exiting 0 is not proof for a family it skipped; the Limits cell of the oracle row is where a skipped family is recorded.