WA Health Summary — Patient Chart Explorer (OpenClinAI Sandbox)

Decoupled Knowledge Lever (Applies to Any Selected Model):
• 🧠 + 4-Layer Clinical KB: Injects deterministic pre-condensation clinical audit findings (RxNorm brand/generic duplicates, KDIGO/ADA stage mismatches, note "hold/stop" vs. active medication orders, and immunization schedules) alongside the condensed chart (~+400–900 tok).
• 📄 Raw Chart Only: Passes only the condensed FHIR chart ledger with zero KB hints so you can evaluate the model's standalone clinical reasoning at any token budget.

Progressive Context Allocation (256k → 1k):
• 256k / 128k (Max-Window Stress Test & Decadal ICU/HIE History): Expands up to the full 128k–256k ceiling of Gemma 4 12B/26B and Qwen 3 (e.g., Maurice expands to 128.3k tok at 128k and 252.0k tok at 256k) so you can directly test huge-context attention dilution and latency vs. compact budgets.
• 64k / 32k (Full Longitudinal Ledger): Fits 100% of normalized clinical records for multi-year charts (including US Core & MIMIC-IV/Synthea cohorts) with zero or minimal clinical records trimmed.
• 16k / 8k (Recommended Limit Window): Retains 100% of active conditions, meds, allergies, implants, clinical notes, and multi-year lab trajectories inside the 16k–24k model Recommended Limit.
• 4k / 2k / 1k (Compact Visit / Emergency IPS Capsule): Prioritizes the mandatory IPS Safety Core first; when connected to Med Agent Hub, any trimmed items needed for your question are dynamically recovered from Tier 0 via conditional overflow lookup.

Two Provenance-Preserving Modes:
• Lossless (Verbatim): Renders each extracted clinical resource as its own timestamped [N] line with zero semantic folding—ideal when you have a 32k–64k window or want exact per-visit records.
• Grouped (Fold Renewals): Uses the 5-Tier Progressive Folding Ladder (Tier 0 Full Ledger → Tier 1 Renewal-Folded → Tier 2 24m Horizon → Tier 3 IPS + Lab Trajectories → Tier 4 IPS Micro-Capsule) to fold stopped medication renewals into single [first..last, N renewals] spans and compress multi-year lab panels into First → Peak/Trough → Latest trajectory lines, preserving all underlying memberUuids.

Stage 0 Structural Hygiene & Privacy Segmentation:
• Auto-Cleaned Wrapper Bloat: Strips FHIR JSON syntax overhead, administrative Claim/ExplanationOfBenefit billing records, and byte-identical SHA-256 Base64 note duplicates before token counting (~97% token reduction with zero clinical facts lost).
• Clinic Room Equipment (Device): Hides facility hardware (Hospital bed, Examination table, Dental x-ray) while always keeping patient implants (ICD, Mechanical valve, Coronary stent).
• Experimental Sandbox Sensitivity Filter (42 CFR Part 2 / WA MHMDA): An OpenClinAI sandbox experiment (separate from WAHS's native IPS + SMART Health Link sharing) that withholds sensitive substance-use, behavioral health, and domestic safety findings and logs the redaction count in [Privacy Manifest].

How Out-of-Budget Records Stay Available:
Even when a tight budget (1k–8k) trims older routine labs or visits from the prompt:
1. [2] Completeness Manifest: Line [2] of the ledger tells the LLM the exact count of omitted items per section (budget_overflow_omitted: lab_observation=N) so it never confuses a trimmed test with "never performed."
2. Pre-Condensation 4-Layer KB Audit: RxNorm duplicate checks, KDIGO/ADA stage conflicts, and note "hold/stop" vs. active medication checks run on 100% of the uncondensed chart before trimming.
3. On-Demand Overflow Lookup Tools: search_chart_entries, get_lab_history, and read_clinical_note query the full uncondensed Tier 0 chart in memory whenever a question asks about deep history.