Skip to content

Endexis v2.0.0

Choose a tag to compare

@spencernorton spencernorton released this 07 Nov 02:05
14bc486

Summary

  • Precision Column & DA2 Field Disambiguation (strict_precision=True): Adds grain-pruned column and DA_ELEMENT selection across 6 temporal anchor surfaces (CLINICAL_EVENT.EVENT_END_DT_TM vs. VERIFIED_DT_TM vs. UPDT_DT_TM; MED_ADMIN_EVENT administration start vs. ORDERS.ORIG_ORDER_DT_TM; ENCOUNTER.INPATIENT_ADMIT_DT_TM vs. REG_DT_TM/ARRIVE_DT_TM/DISCH_DT_TM; SCH_APPT.BEG_DT_TM vs. ACTION_DT_TM; CHARGE.SERVICE_DT_TM vs. POSTED_DT_TM), identifier alias disambiguation (PERSON_ALIAS MRN Code Set 4 vs. ENCNTR_ALIAS FIN Code Set 319 vs. internal f8 surrogate keys), and multi-1:N child join-explosion guardrails.
  • Clinical & SQL Safety Qualification Filters: Enforces Order Status Not In (Canceled, Voided) (Code Set 6004: 2542, 2553) preserving Discontinued (2545) for historical MAR/BCMA audits, CE Result Status In (Auth (Verified), Modified, Altered) (Code Set 8: 25, 34, 35) with VALID_UNTIL_DT_TM >= 31-DEC-2100, LEFT OUTER JOIN V500.DIAGNOSIS for active census admitting diagnosis queries, and test-patient prefix exclusion (< 'ZZ').
  • Context-Conditioned Filter-Value & Code Set Resolver (resolve_da2_filter_value): Scores 6,458 DISPLAY <-> CODE_VALUE pairs across 90 DA2 filter fields conditioned on (query_intent, domain, subentity, da2_field_name, code_set).
  • Zero-Leakage V2 Adversarial Benchmark & SQLite Edge-Case Harness: Adds a 20-scenario V2 adversarial benchmark (data/eval_adversarial_v2.jsonl, evaluate_adversarial_v2_benchmark: 100.0% Pass@1, mean F1 0.995) and a 5-scenario SQLite V500 edge-case execution suite (run_adversarial_sql_edge_case_suite), plus .endexis.toml discovery (endexis init-discover) and LocalNeuralCompiler GGUF client support.
  • Expanded Catalogs & Retrained Stage-2 RPO-DPO Model: Enriches all 2,606 DA2 elements and 230 V500 tables, rebuilds 4,501 6-step <reasoning> SFT examples and 189 contrastive DPO pairs, and updates NorviTech/endexis-qwen3-0.6b-dpo-merged (v2.0.0) with BF16 .safetensors and Q8_0 / Q4_K_M GGUF checkpoints.

Verification

  • python3 scripts/release_pipeline.py audit (release_audit_passed: true, 74 files scanned, 0 violations)
  • pytest -q (36 passed)