Code harvest & verification

Two related tracks: pointers (repos / harvested samples attached to methods) and artifacts (verified reference implementations with a verification ladder).

Pointer / harvest lane (continuous)

Typical chain:

1. discover_code_surfacing_candidates.py / discover_paper_level_candidates.py
2. classify_code_surfacing_batch.py (GitHub)
3. stream_harvest_follower.py → CodeSample nodes, state in stream_harvest_state.json
4. synthesize_probe_inputs.py / synthesize_fixtures.py
5. probe_harvested_contracts.py
6. assign_roles_harvested.py → attach_harvested_to_methods.py
7. Loop wrappers: run_contract_pipeline_loop.sh, launchd KeepAlive

Edges: HAS_HARVESTED_IMPL, IMPLEMENTS (with roles), plus repo-level IMPLEMENTED_BY / PapersWithCode loaders.

Artifact / verification ladder

groundwork/13_impl_generation_verification/ — batch cycles, V0–V4 ladder, verification_ladder.py. groundwork/14_code_ontology_surface/ — load_*.py, graph_write_guard.py (refuses edge regression / role downgrade). 21_reconcile_served_levels.py writes down verification already held.

Served via MCP/REST: get_reference_implementation, list_reference_implementations (min verification floor; refuse rather than under-serve). Human page: /code.

Risks

• Continuous writers need writer credentials and ledger discipline.
• State JSON files must not be hand-edited.
• Unbacked verification_level is invalid (R2); readers treat as 0.