EXP-2845 cluster: route hypothesis + unified triage + time-of-day refinement

Three woven deliverables closing the EXP-2843/2844 phenotype line.

EXP-2845 (route hypothesis):
- 17 significant patients, S0/S1 delivery decomposed by route
- Loop and Trio deliver IDENTICAL total compensation in S1
  (Δ=0.254 U/h both, p=0.82) but route oppositely:
  Loop +basal, Trio -basal +SMB
- H2 PASS (p=0.032): Loop basal uplift > Trio
- H3 PASS (p=0.82): total delivery similar
- H1 FAIL (p=0.41): SMB share similar (~33%) - SMB *delta* differs 6×
- Decisive: phenotype is controller-software route choice, not biology
- chart: exp-2845_route_by_controller.png

EXP-2845b (unified triage cross-tab):
- Outer-join EXP-2812 + 2831 + 2844 flags into single triage table
- Patient b (Loop, flat) = ONLY triple-flag in cohort
  (recovery=0.00, post_high=30.9%, wear delta=-31.5%)
- 2 Trio patients carry double-flags (wear + post-high)
- chart: triage_unified_table.png

EXP-2845c (time-of-day refinement, pairs with EXP-2780):
- Per-(patient, hour, state) basal aggregation, S1-S0 % delta
- down_shifters cut hardest in dawn (-15%) and midday (-13%)
- up_shifters raise hardest in midday (+11%) and evening (+11%)
- flat phenotype is also window-flat (controller not adapting)
- charts: tod_basal_heatmap_cohort.png + tod_basal_panel_by_phenotype.png

Combined: profile audition recommendations must be
(controller, phenotype, time-of-day) aware. Same envelope demand,
same total compensation, opposite routes, localized to specific hours.

All Stream B; no biology numbers; charter Appendix V compliant.

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

Author: bewest

docs/60-research/exp-2845-cluster-route-triage-tod-2026-04-22.md

@@ -0,0 +1,172 @@
+# EXP-2845 Cluster: Route, Triage, and Time-of-Day Refinement
+
+**Date**: 2026-04-22
+**Stream**: B (operational)
+**Charter**: two-stream-methodology-charter-2026-04-22.md (V1–V8)
+**Predecessors**: EXP-2843 (envelope coupling), EXP-2844 (phenotype split)
+
+Three woven deliverables that close the EXP-2843/2844 line: route
+hypothesis test, unified triage cross-tab, and time-of-day refinement
+of the phenotype audit.
+
+---
+
+## EXP-2845: SMB-route hypothesis (Loop vs Trio in S1)
+
+**Question**: Does Loop's "never down-shift" (0/6) and Trio's "always
+down-shift" (5/6) reduce to *route* of insulin delivery (basal vs SMB)
+rather than total quantity?
+
+**Method**: For each of 17 significant patients, compute hourly delivery
+rate in S0 and S1 windows for: actual basal, bolus, SMB, total.
+Test H1 (Trio SMB share > Loop), H2 (Loop basal uplift > Trio),
+H3 (total delivery similar).
+
+**Results**:
+
+| Controller | n | Δ basal U/h | Δ SMB U/h | Δ total U/h | SMB share S1 |
+|------------|--:|------------:|----------:|------------:|-------------:|
+| Loop       | 6 | **+0.084** | +0.033 | **+0.254** | 0.34 |
+| Trio       | 6 | **−0.048** | +0.208 | **+0.254** | 0.36 |
+| OpenAPS    | 5 | +0.024 | 0.000 | −0.019 | 0.00 |
+
+| Hypothesis | Test | p | Result |
+|---|---|---|---|
+| H1 SMB share Trio > Loop | M-W (greater) | 0.41 | FAIL |
+| H2 Basal uplift Loop > Trio | M-W (greater) | **0.032** | **PASS** |
+| H3 Total delivery similar | M-W (two-sided) | **0.82** | **PASS** |
+
+**Interpretation (Stream B, decisive)**: Loop and Trio deliver
+**identical total compensation** in S1 (0.254 U/h both, p=0.82) but
+**route it oppositely**: Loop pushes via basal uplift; Trio cuts basal
+and pushes via SMB. SMB share at S1 is ~similar (~33%) between Loop and
+Trio, but the *delta* in SMB is 6× larger for Trio (+0.21 vs +0.03).
+
+This is the cleanest possible Stream B finding: **same envelope demand,
+same closed-loop net response, different controller-software route
+choice**. EXP-2844's phenotype split is now fully explained as a
+controller-route artifact, not a per-patient biological property.
+
+**Audition implication**: Profile basal recommendations must be
+controller-aware. A "raise scheduled basal" recommendation for a Loop
+up-shifter and a "lower scheduled basal" recommendation for a Trio
+down-shifter may both be the correct way to *retire* the same envelope
+demand the open-loop profile is missing.
+
+OpenAPS shows zero SMB delta in this cohort (5/5 patients) — likely
+configuration- or version-dependent; flagged for follow-up.
+
+**Source**: `tools/cgmencode/exp_smb_route_2845.py`,
+`docs/60-research/figures/exp-2845_route_by_controller.png`.
+
+---
+
+## EXP-2845b: Unified triage cross-tab (flat-phenotype investigation)
+
+**Question**: Does the flat phenotype (n=5, worst recovery 0.00) overlap
+with EXP-2812 (recovery flags) and EXP-2831 (wear flags)?
+
+**Method**: Outer-join EXP-2844 phenotype × EXP-2812 recovery flags ×
+EXP-2831 wear flags into a unified triage table; sort by total flag
+count.
+
+**Results** (top rows):
+
+| Patient | Controller | Phenotype | Δ basal % | Recovery | Post %high | Wear Δ% | Flags |
+|---------|------------|-----------|----------:|---------:|-----------:|--------:|:-----:|
+| **b**   | Loop  | **flat** | +0.3   | 0.00 | 30.9 | −31.5 | **3** |
+| ns-d444c120c23a | Trio  | down_shift | −24.8 | 0.25 | 34.7 | −14.2 | 2 |
+| ns-dde9e7c2e752 | Trio  | up_shift   | +23.6 | 0.25 | 30.9 | +16.0 | 2 |
+| ns-6bef17b4c1ec | Trio  | down_shift | −27.7 | — | — | −20.6 | 1 |
+| i               | Loop  | up_shift   | +15.8 | — | — | −44.7 | 1 |
+
+**Interpretation**:
+- Patient `b` is the only **triple-flag** case across the entire cohort.
+  It is also the only flat-phenotype patient with EXP-2812 transition
+  data — combined evidence is overwhelming for site rotation/profile
+  re-audit.
+- The other 4 flat patients lack S0→S1 transition data in EXP-2812
+  (insufficient transitions ≥2). Their flat status remains a phenotype
+  observation but cannot be cross-validated against recovery yet.
+- 2 Trio patients carry double-flags via wear + post-high; both have
+  significant S1 phenotype shifts.
+
+**Source**: `tools/cgmencode/viz_unified_triage.py`,
+`docs/60-research/figures/triage_unified_table.png`,
+`externals/experiments/exp-2845b_unified_triage.parquet`.
+
+---
+
+## EXP-2845c: Time-of-day refinement
+
+**Question**: Does the S1 basal shift concentrate in particular
+time-of-day windows? (V8: pairs viz with EXP-2780 circadian basal.)
+
+**Method**: For each (patient, hour, state), aggregate mean actual and
+scheduled basal. Compute hourly Δbasal% (S1 − S0 normalized to
+scheduled). Render cohort heatmap (24 hours × 17 patients) and
+phenotype-aggregated window panel.
+
+**Results** (median Δbasal % by phenotype × window):
+
+| Window | down_shift | flat | up_shift |
+|--------|----------:|-----:|---------:|
+| Overnight (0–3) | −9% | −2% | +6% |
+| **Dawn (4–9)** | **−15%** | −3% | +7% |
+| Midday (10–15) | −13% | −1% | **+11%** |
+| Evening (16–23) | −5% | +2% | **+11%** |
+
+**Interpretation (Stream B)**:
+- **down_shift** patients: controller cuts hardest in **dawn and
+  midday** (−15%, −13%). The open-loop profile is over-basaled in the
+  morning hours; controller corrects by cutting.
+- **up_shift** patients: controller raises hardest in **midday and
+  evening** (+11%, +11%). The open-loop profile is under-basaled in
+  active hours; controller corrects by adding.
+- **flat**: window-flat as well (≤±3% in every window) — these
+  controllers really are not adapting time-of-day demand at all.
+
+**Audition implications by (phenotype, window)**:
+
+| Phenotype | Dawn | Midday | Evening |
+|-----------|------|--------|---------|
+| down_shift | **lower scheduled basal here first** | also high-priority cut | mild cut |
+| up_shift | mild raise | **raise scheduled basal here** | **also raise** |
+| flat | (no signal) | (no signal) | (no signal) — investigate controller capacity |
+
+**Source**: `tools/cgmencode/viz_time_of_day_audit.py`,
+`docs/60-research/figures/tod_basal_heatmap_cohort.png`,
+`docs/60-research/figures/tod_basal_panel_by_phenotype.png`,
+`externals/experiments/exp-2845c_tod_summary.parquet`.
+
+---
+
+## Combined synthesis
+
+The EXP-2843/2844/2845 line now provides a complete operational story:
+
+1. **Envelope state matters** (EXP-2843): 17/22 patients show
+   significant S0/S1 basal differences.
+2. **Direction is a controller-route choice** (EXP-2844 + 2845):
+   Loop routes via basal-up, Trio routes via SMB+basal-down,
+   *same total compensation*.
+3. **Time-of-day localizes the recommendation** (EXP-2845c): the
+   actionable schedule edit lives in dawn (down-shifters) or
+   midday/evening (up-shifters).
+4. **Unified triage** (EXP-2845b) makes patient `b` the unambiguous
+   highest-confidence intervention candidate (3 flags).
+
+All findings are Stream B operational. No biology number is asserted
+anywhere in this cluster; profile-vs-actual gap is the audition unit
+throughout. Cohort overlays use percentile bands (V3) and "you are
+here" callouts (V4). Phenotype direction is foregrounded in every
+chart (V5). Each research question shipped with a paired chart (V8).
+
+## Open follow-ups
+
+- OpenAPS zero-SMB observation: configuration vs version vs cohort
+  artifact?
+- Flat-phenotype patients without EXP-2812 transition coverage:
+  back-fill with looser n_trans criterion?
+- Sensor-gap orthogonal Stream A test (still open from prior plan)
+- 24h state windows (EXP-2843b) to catch transitions earlier