Repository navigation
# [SCHEMA] Expand SQLite and Output Recorders for Class-Level Risk Vectors #110
Description
🎯 Objective
Ensure the newly calculated Class-level risk vectors are successfully persisted in the data translation layer.
🛠️ Implementation
record_keeper.py: Update theclass_datatable schema to include the 14 relevant metrics fromRISK_SCHEMA(currently it only holds basic physics likemethod_count,state_entanglement, andlcom_score).gpu_recorder.py&audit_recorder.py: Ensure theclassesarray inside the file payload retains its calculated risk vectors so the WebGL UI and forensic JSON can render them accurately.
Activity
squid-protocol commentedon Jul 25, 2026
Verified the "14 relevant metrics" against `08-01-methodology.md`'s hierarchy table and `RISK_SCHEMA` (`gitgalaxy/standards/analysis_lens.py:1047`). The number's exactly right, and here's the concrete list so schema work doesn't have to re-derive it:
Needs Class-level rollup (14): `cognitive_load`, `safety_score`, `tech_debt`, `verification`, `api_exposure`, `concurrency`, `state_flux`, `spec_match`, `documentation`, `algorithmic_dos`, `obscured_payload`, `logic_bomb`, `injection_surface`, `memory_corruption`.
Stays File-level only, N/A at Class (the other 5): `dead_code` (Graveyard), `stability` (File Stability/Heat), `churn` (Deep Churn), `tabs_vs_spaces` (Civil War), `secrets_risk` (Hardcoded Secrets). These are explicitly marked N/A at Class/Function in the doc table -- not an oversight, don't add columns for these.
Also confirmed the current `class_data` table description is accurate: `detector.py`'s class extractor (~line 466-522) only writes `method_count`, `state_entanglement`, and `lcom_score` today.
One thing worth deciding before the schema is finalized (see discussion on #109): per the methodology doc, Class-level values are "count based, 0-∞", not the 0-100 sigmoid-normalized scale File-level uses. Worth making sure the 14 new columns are typed/labeled to make that distinction obvious (e.g. not reusing the same column names/scale as the file-level risk vector), so downstream consumers don't accidentally treat a class's raw count as a percentage.