Skip to content

Endexis v2.3.0 — Neuro-Symbolic Concept Isolation, Multi-Turn Refinements & 14x SFT+RPO-DPO Sync

Latest

Choose a tag to compare

@spencernorton spencernorton released this 12 Nov 13:38
f991c4d

Endexis v2.3.0 — Neuro-Symbolic Compiler Isolation, Multi-Turn Refinements & 14x Corpus SFT+RPO-DPO Overhaul

Highlights

  • 14x Corpus-Synchronized Neural Weights (5,277 SFT Examples + 648 Contrastive RPO-DPO Pairs):
    • Retrained endexis-qwen3-0.6b-dpo-merged (BF16, Q8_0, and Q4_K_M GGUF) from stock Qwen/Qwen3-0.6B across all 36 clinical, nursing/ADL, medication, scheduling, and financial revenue-cycle templates plus 768 compact NeuralCompilationIR examples and 648 preference pairs.
    • Stage-2 NLL-anchored RPO-DPO (beta = 0.1, lambda_nll = 0.15) achieved 100.0% preference accuracy (+3.4014 mean log-prob margin) with flat full-corpus SFT validation loss (0.1137 -> 0.1137) and bounded per-fold OOF SFT loss deltas (+0.0024 to +0.0058 vs. v2.0.0 +0.0158).
  • Clinical-vs-Billing Concept Isolation & Multi-Turn Refinements:
    • Added lexical_guards.py word-boundary disambiguation preventing clinical physician/provider requests from pulling CWF GS Adjustments (Adj Post Prsnl, Prsnl Position, Adj Trans *), preventing "discharged" from matching CHARGE columns, and preserving affirmative clauses trailing negated phrases ("without billing columns and include LOS").
    • Added structured and natural-language multi-turn negative/override refinements (exclude_columns, exclude_filters, prior_filters, override_filters across the 95-field da2_filter_value_catalog) across IR, <cclquery> XML, .daquery, EJS V500 SQL, CCL .prg, and Eclipse BIRT 4.4.2 .rptdesign, plus conversational-filler stripping in _derive_report_name().
  • Sealed N = 640 Post-Freeze Qualification & Sub-260ms GGUF Inference:
    • 5-Fold GroupKFold OOF (N = 132): Unverified Track-A Strict Pass@1 jumped from 0.00% (v2.0.0 322-row weights) to 50.76% (+50.76pp), with 99.24% CUDA Q8_0 <-> BF16 structural parity and 257.44 ms single-pass Q4_K_M CUDA p50 (vs. 2,352.61 ms legacy PyTorch HTTP mean).
    • Sealed Post-Freeze Holdout (data/eval_postfreeze_v3.jsonl, N = 640, 0 / 0 / 0 Exact / Token-Bag / Normalized 5-Gram Skeleton Overlap; 19 Generic Boilerplate Raw 5-Grams): Strict zero-rescue Track-A Q4_K_M / Q8_0 reached 95.00% / 97.19% SQLite V500 SQL + <cclquery> XML pass rates and 75.47% / 78.28% 4-surface (XML + SQLite + CCL + BIRT) compile rates (191.81 ms / 209.24 ms p50; 214.22 ms / 234.08 ms mean) vs. 0.00% v2.0.0; Track-B neuro-symbolic reached 100.00% (640 / 640) SQLite V500 + XML + contamination-free compliance and 99.69% (638 / 640) 4-surface verified Pass@1 (75.47% raw unedited neural IR + 19.38% symbolically repaired neural IR = 94.84% neural-led acceptance, 5.16% full symbolic fallback) at 210.12 ms full wall-clock p50 (232.82 ms mean).
    • Zero-Leakage & Privacy Audits: 100.00% domain-conditioned Acc@1 (89.58% unconditioned Acc@1) and 100.00% CODE_VALUE exact match across 144 holdout filter values, 100.00% across 20 V2 adversarial traps + 14 frozen + 4 holdout queries (53/53 unit tests), and 0 violations across repository tree (including decoded Base64 literal checks), tenant-agnostic invariants, binary .safetensors/.gguf scans, and 17 behavioral canary prompts.