Patient C therapy-discussion vignette + 6 visualizations

Sample clinician/patient discussion report demonstrating the full
audition matrix applied to one well-characterized case. Pulls every
production audition signal + raw 180-day grid stats for patient c
and emits a Stream B operational triage report.

Patient C profile: GMI 7.2%, TIR 51%, CV 43%, clean carb log
(95%% real meals). Four audition signals fire at P=1.00:
  - basal_mismatch (mult=0.0; controller suspends ~100% in fasting)
  - isf_over_correction
  - low_recovery
  - post_high_envelope

Joint reading: coherent over-aggressive-basal cascade. Report frames
4 conversation starters for clinic review (basal trial, ISF softening,
post-high investigation, site rotation) with explicit EXP-2738 safety
caveat - the gap IS the EGP margin; recommended trial reduction
(5-10%) is intentionally much smaller than observed multiplier.

Figures: AGP, TIR bar, scheduled-vs-actual basal, carb log distribution,
post-meal excursion samples, audition signal panel.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: bewest

docs/60-research/figures/patient-c-vignette/01_agp.png

@@ -0,0 +1,164 @@
+# Patient C — Therapy-Discussion Vignette (2026-04-22)
+
+> **Operational triage report (Stream B).** Every recommendation is a
+> *conversation starter* for clinician/patient review — not an autonomous
+> setting change. Per the safety memory (EXP-2738): **the gap between
+> scheduled and delivered insulin often IS the controller's safety
+> margin.** Lowering settings to "match" the gap can increase
+> hypoglycemia.
+
+## Summary
+
+| Metric | Value |
+|--------|------:|
+| Observation window | 180 days |
+| Mean glucose | 162 mg/dL |
+| GMI (eA1c estimate) | **7.2 %** |
+| Coefficient of variation | 43% (target <36%) |
+| Time in range (70–180) | **51%** (target ≥70%) |
+| Time below range (<70) | 3.9% (target <4%) |
+| Time very low (<54) | 1.29% (target <1%) |
+| Time above range (>180) | 28% (target <25%) |
+| Time very high (>250) | 10% (target <5%) |
+| Carb-log quality | 2.2 events/day · 95% ≥10 g (**clean log**) |
+| Median scheduled basal | 1.40 U/h |
+| Median delivered basal | 0.00 U/h |
+
+## Glycemic profile
+
+![AGP](figures/patient-c-vignette/01_agp.png)
+
+![TIR](figures/patient-c-vignette/02_tir.png)
+
+**Read:** TIR is just above 50% — clinically meaningful improvement
+target. CV 43% (>36%) suggests glycemic instability is the
+dominant driver, not just elevated mean.
+
+## Bootstrap-confident audition signals
+
+![Audition panel](figures/patient-c-vignette/06_audition_panel.png)
+
+| Signal | P | Tier | Source |
+|--------|--:|------|--------|
+| **flat_low_recovery** | — | high | matrix |
+| **isf_over_correction** | — | medium | matrix |
+| **site_degradation** | — | low | matrix |
+| **post_high_envelope** | — | medium | matrix |
+| **basal_mismatch** | — | high | matrix |
+
+**Raw bootstrap probabilities:**
+
+| Signal | P | Source |
+|---|---:|---|
+| Basal mismatch | 1.00 | EXP-2869 🔴 high |
+| ISF over-correction | 1.00 | EXP-2861 🔴 high |
+| Low recovery | 1.00 | EXP-2862 🔴 high |
+| Post-high envelope | 1.00 | EXP-2864 🔴 high |
+| Site degradation | 0.39 | EXP-2863 🟠 boundary |
+| ISF under-correction | 0.00 | EXP-2861 ⬜ clean |
+| Simpson paradox | 0.00 | EXP-2859 ⬜ clean |
+
+
+## The story (4 high-confidence signals point to one mechanism)
+
+Patient C shows a **coherent over-aggressive-basal cascade**:
+
+1. **Basal mismatch (P=1.00, mult=0.00):**
+   in fasting equilibrium the controller delivers ~0% of the scheduled
+   basal — it is suspending almost continuously to defend against the
+   schedule.
+
+2. **ISF over-correction (P=1.00):** when
+   the controller does correct, BG drops more than the ISF predicts
+   (over-correction). This is consistent with the basal already being
+   "too much" — additional bolus drops too far.
+
+3. **Low recovery (P=1.00):** after a low,
+   recovery to 100 mg/dL within 60 minutes is rare — also consistent
+   with continuous basal pressure.
+
+4. **Post-high envelope (P=1.00):** after
+   a high, BG sustains above target without a quick recovery — the
+   wide IQR (110–203) confirms oscillation rather than tight control.
+
+## Conversation starters for clinic review
+
+> ⚠️ **Do not change settings autonomously.** These are hypotheses for
+> the clinician to discuss with the patient.
+
+### 1. Review basal schedule
+The 4 signals above are individually noisy but jointly point to
+**scheduled basal being too high**. Recommended discussion:
+- Audit overnight TBR pattern (figure 1, hours 0–6).
+- Consider: was the basal schedule last reviewed? Has weight,
+  activity, or insulin sensitivity changed?
+- A **conservative trial** of a 5–10% basal reduction (clinician-
+  supervised) tests the hypothesis safely. The controller's 0%
+  delivery suggests there is room.
+
+### 2. Review ISF (correction factor)
+Over-correction (P=1.00) on top of an already-suppressed basal
+suggests ISF may be too aggressive. Discuss whether to **soften ISF
+by 10–15%** as a paired adjustment with #1.
+
+### 3. Investigate post-high recovery
+P=1.00 for sustained post-high envelope. Discussion points:
+- Is bolus timing pre-meal (not at meal start)?
+- Are corrections being delivered at high BG, or is the patient
+  waiting for the controller alone?
+
+### 4. Site rotation (low-confidence)
+Site-degradation P=0.39 (boundary, not
+confirmed). Worth asking about rotation cadence but not a primary
+finding.
+
+## Carb-log quality (no concern)
+
+![Carb log](figures/patient-c-vignette/04_carb_log.png)
+
+Patient C's log is clean: 95% of events ≥10 g, only 3% are <5 g
+(treat-of-low / noise). The audition signals are **not** confounded
+by data-quality issues for this patient.
+
+## Basal pattern detail
+
+![Basal](figures/patient-c-vignette/03_basal_schedule.png)
+
+The black line is the patient's scheduled basal across the day; the
+red band is what the controller actually delivered (P25–P75). The
+controller is **continuously suppressing** delivery across all hours
+— most pronounced during the afternoon/evening peak hours of the
+schedule.
+
+## Post-meal excursion examples
+
+![Excursions](figures/patient-c-vignette/05_meal_excursions.png)
+
+Six largest real meals in the dataset, plotted ±1 hr / +4 hr from the
+meal time. These illustrate the high variance in meal response that
+drives the wide IQR in the AGP.
+
+## Methodology
+
+* All audition signals computed from production loaders
+  (`tools/cgmencode/production/*_facts_loader.py`).
+* Bootstrap confidence (P) computed per the EXP-2859/2861/2862/2863/
+  2864/2869 protocols.
+* Carb-event quality assessed against the EXP-2866 conventions
+  (`meal_filter.py`).
+* Source data: `externals/ns-parquet/training/grid.parquet`,
+  patient_id = `c`, 51,841 rows, 180 days.
+
+## Caveats
+
+* **Not medical advice.** Audition signals are statistical patterns;
+  every recommendation must be filtered through the clinician's
+  judgment and the patient's full history.
+* **Stream B operational, not Stream A causal.** These signals describe
+  *what the controller is doing*, not *what biological ISF/basal
+  needs are*. The closed-loop system response is observed; the
+  underlying biology is inferred only loosely.
+* **Per EXP-2738**: the basal-mismatch gap IS the EGP safety margin
+  the controller needs. The recommended trial reduction (5–10%) is
+  much smaller than the observed multiplier (~0%) because the gap is
+  protective.