Skip to content

ROADMAP.md

Latest commit

 

History

History
77 lines (62 loc) · 3.35 KB

ROADMAP.md

File metadata and controls

77 lines (62 loc) · 3.35 KB
Raw

SOM code corpus for Python roadmap

Purpose

This roadmap orders the work that turns the Python corpus into the training material the four-layer generator in som-core needs. The order is data first: finish the oracle fixtures, then harden the near misses and cover the ISA's long tail so more of each gold is expressed as snippets. The subject stays the Python ecosystem written as code its community accepts. The current support state is in STATUS.md.

Near-term outcomes

Oracle fixtures complete

  • ID: oracle-fixtures-complete
  • Outcome: Every family under data/curated/families has a fixture, the harness reports no skipped family, it exits non-zero when a family has no fixture instead of skipping it, and no oracle or caught_by field is null.
  • Boundary: The three missing fixtures for 08-asyncio-concurrency-limiter, 09-pydantic-field-cross-validation, and 10-security-timing-constant-auth, plus the harness's handling of a missing fixture; not new families.
  • Completion evidence: The harness run over 08 09 10 shows each gold exiting 0 and each near miss failing on exactly its declared caught_by, the curation check reports no null field, and a run with one fixture renamed away exits non-zero.
  • Tracking: Not assigned.

Later outcomes

Non-minimal near misses

  • ID: non-minimal-near-misses
  • Outcome: Families gain near misses that are not a one-line edit of the gold candidate, refactored or carrying two defects, so preference data built from them cannot be separated from the gold by edit distance alone.
  • Boundary: New near-miss candidates and their family.json entries under data/curated/families, each still failing its fixture on an assertion with a measured caught_by; not new families and not the harness.
  • Completion evidence: The harness and the curation check exit 0 over the changed families, and diff between each new candidate and its gold shows more than one hunk.
  • Tracking: Not assigned.

Snippet long tail

  • ID: snippet-long-tail
  • Outcome: Route handlers with branching bodies, exception handlers, CLI commands, and async code become snippets where at least two families share the idiom, so the verifier's coverage rises without a template line that is only a placeholder.
  • Boundary: New snippet definitions under data/snippets/<library>/ and new families that give a one-family idiom its second use; not the template syntax and not the assembler.
  • Completion evidence: scripts/verify_snippets.py exits 0 with a higher coverage: line than STATUS snippet-isa records, and scripts/verify_roundtrip.py still rebuilds every family it did before.
  • Tracking: Not assigned.

Non-goals

Algorithms, data structures, and design patterns

  • ID: algorithms-data-structures-and-design-patterns
  • Reason: Families that differ in an algorithm, a data structure, or a design pattern as such are language-neutral and belong to the som-code corpus; this project teaches how the Python ecosystem is used and written, and a family here is judged by library behaviour and community style.

Platform pilot families

  • ID: platform-pilot-families
  • Reason: Docker, Terraform, and Git scenarios have no executable oracle in Python, so a family for them cannot meet the corpus contract; they belong to a corpus project for that platform, if one is ever opened.