teleo-codex/agents/vida/research-journal.md

Vida Research Journal

Session 2026-03-18 (Continuation) — Food-as-Medicine Intervention Taxonomy and Political Economy

Question: Does the intervention TYPE within food-as-medicine (produce prescription vs. food pharmacy vs. medically tailored meals) explain the divergent clinical outcomes — and what does the CMS VBID termination mean for the field's funding infrastructure?

Belief targeted: Belief 2 (non-clinical determinants are intervenable) — specifically testing whether "better" FIM intervention types rescue the food-as-medicine clinical outcomes thesis that Session 1 challenged.

Disconfirmation result: The intervention-type hypothesis FAILS. Medically tailored meals — the most intensive FIM intervention, with pre-prepared food delivered to patients' homes PLUS dietitian counseling — also show null HbA1c improvement in a controlled trial (Maryland pilot, JGIM 2024: -0.7% vs. -0.6%, not significant). The simulation-vs-RCT gap is not resolved by increasing intervention intensity. Two controlled trials, two intervention types, same null glycemic finding.

However: a new complicating factor emerged. The control group in the Maryland MTM pilot received MORE medication optimization than the treatment group — suggesting medical management may be more glycemically impactful than food delivery in the short term. The MTM may be producing real benefit but the comparison arm is also improving through a different pathway.

Key finding: The food-as-medicine field has a fundamental taxonomy problem. "Food is medicine" simultaneously means:

1. Diet quality is causally important for health outcomes (strong evidence)
2. Produce voucher programs improve clinical outcomes (weak-to-null RCT evidence)
3. Medically tailored meals reduce hospitalizations in complex patients (strong observational, weak RCT for glycemic outcomes)
4. Food-as-medicine programs advance health equity by reducing food insecurity (consistent evidence)

These four claims have DIFFERENT evidence standards and DIFFERENT target outcomes. The KB has been treating them as one claim. They need to be disaggregated.

Critical policy event: CMS VBID model terminated end of 2025. VBID was the primary payment vehicle for food benefits in Medicare Advantage for low-income enrollees. The SSBCI replacement pathway excludes socioeconomic eligibility criteria — effectively removing food-as-medicine access for the core target population. The Trump administration announced the most rhetorically food-forward dietary guidelines in history (January 2026) ONE WEEK after VBID ended. Peak rhetoric, contracting infrastructure.

Pattern update: FIVE sessions (including both March 18 sessions) now confirm the same meta-pattern: the gap between VBC/FIM/non-clinical intervention THEORY and PRACTICE. Session 1-3: VBC payment alignment doesn't automatically create prevention incentives. Session 4 (March 18 Session 1): identifying non-clinical determinants doesn't mean intervening on them improves outcomes. Session 5 (March 18 Session 2): even the most intensive food intervention type (MTM) fails to show glycemic improvement in controlled settings. The pattern is not convergence — it's accumulation of disconfirmatory evidence.

New pattern: Selection bias as the unifying explanation across FIM evidence. Programs showing dramatic results (Geisinger n=37, Recipe4Health) are self-selected populations. RCTs enroll everyone. The control groups also improve significantly. This suggests: food interventions may work for the motivated subset, but population-level impact is smaller than pilot programs suggest. This parallels the clinical AI story: adoption metrics (80% of physicians have access) vs. active daily use (much lower). Access ≠ engagement ≠ outcomes.

Confidence shift:

• Belief 2 (non-clinical determinants): FURTHER COMPLICATED — two controlled FIM trials (JAMA Doyle RCT + Maryland MTM pilot) both show null glycemic improvement. The 80-90% non-clinical determinant claim stands as a correlational diagnosis. The intervenability is weaker than assumed even for the most intensive single-factor intervention. The KB claim needs scope qualification distinguishing: (a) observational correlation between food insecurity and outcomes [strong], (b) clinical effect of resolving food insecurity on outcomes [weak in RCTs], (c) population-level health equity improvement from FIM [moderate, better evidence for diet quality than clinical outcomes].
• Belief 3 (structural misalignment): Extended — VBID termination is the clearest example yet of payment infrastructure contracting while rhetorical support peaks. The structural misalignment pattern applies not just to VBC/GLP-1s but to food-as-medicine funding. MAHA is using "food not drugs" rhetoric while the payment mechanism for food benefits disappears.

Sources archived: 7 (HHS FIM landscape summary, CMS VBID termination, Trump dietary guidelines reset, AHA FIM systematic review, Health Affairs MTM modeling pair, Maryland MTM pilot RCT, Diabetes Care produce prescription critique, APHA FIM equity report, NASHP CHW policy update)

Extraction candidates: 4 claims: (1) FIM intervention taxonomy with stratified evidence, (2) null MTM glycemic result pattern across two controlled trials, (3) VBID termination removes low-income MA food benefit access, (4) equity-vs-clinical outcome distinction for FIM policy justification

Session 2026-03-18 — Behavioral Health Infrastructure: What Actually Works at Scale?

Question: How did Medicare Advantage become the dominant US healthcare payment structure, what are its actual economics (efficiency vs. gaming), and how does the US senior care system compare to international alternatives?

Key finding: MA's $84B/year overpayment is dual-mechanism (coding intensity $40B + favorable selection $44B) and self-reinforcing through competitive dynamics — plans that upcode more offer better benefits and grow faster, creating a race to the bottom in coding integrity. But beneficiary savings of 18-24% OOP ($140/month) create political lock-in that makes reform nearly impossible despite overwhelming fiscal evidence. The $1.2T overpayment projection (2025-2034) combined with Medicare trust fund exhaustion moving to 2040 creates a fiscal collision course that will force structural reform within the 2030s.

Confidence shift:

• Belief 2 (non-clinical determinants): strengthened — Commonwealth Fund Mirror Mirror 2024 shows US ranked 2nd in care process but LAST in outcomes, the strongest international validation that clinical quality ≠ population health
• Belief 3 (structural misalignment): strengthened and deepened — MA is value-based in form but misaligned in practice through coding gaming, favorable selection, and vertical integration self-dealing (UHC-Optum 17-61% premium)
• Belief 4 (atoms-to-bits): complicated — PACE's 50-year failure to scale (90K out of 67M eligible) despite being the most integrated model suggests structural barriers beyond technology

Sources archived: 18 across three tracks (8 Track 1, 5 Track 2, 5 Track 3) Extraction candidates: 15-20 claims across MA economics, senior care infrastructure, and international benchmarks

Session 2026-03-12 — GLP-1 Agonists and Value-Based Care Economics

Question: How are GLP-1 agonists interacting with value-based care economics — do cardiovascular and organ-protective benefits create net savings under capitation, or is the chronic use model inflationary even when plans bear full risk?

Key finding: GLP-1 economics are payment-model-dependent in a way the existing KB claim doesn't capture. System-level: inflationary (CBO: $35B additional spending). Risk-bearing payer level: potentially cost-saving (ASPE/Value in Health: $715M net savings over 10 years for Medicare). The temporal cost curve is the key insight — Aon data shows costs up 23% in year 1, then grow only 2% vs. 6% for non-users after 12 months. Short-term payers see costs; long-term risk-bearers capture savings. But MA plans are RESTRICTING access (near-universal PA), not embracing prevention — challenging the simple attractor state thesis that capitation → prevention.

Pattern update: This session deepens the March 10 pattern: MA is value-based in form but short-term-cost-managed in practice. The GLP-1 case is the strongest evidence yet — MA plans have theoretical incentive to cover GLP-1s (downstream savings) but restrict access (short-term cost avoidance). The attractor state thesis needs refinement: payment alignment is NECESSARY but NOT SUFFICIENT. You also need adherence solutions, long-term risk pools, and policy infrastructure (like the BALANCE model).

Cross-session pattern emerging: Two sessions now converge on the same observation — the gap between VBC theory (aligned incentives → better outcomes) and VBC practice (short-term cost management, coding arbitrage, access restriction). The attractor state is real but the transition path is harder than I'd assumed. The existing claim "value-based care transitions stall at the payment boundary" is confirmed but the stall is deeper than payment — it's also behavioral (adherence), institutional (MA business models), and methodological (CBO scoring bias against prevention).

Confidence shift:

• Belief 3 (structural misalignment): further complicated — misalignment persists even under capitation because of short-term budget pressure, adherence uncertainty, and member turnover. Capitation is necessary but not sufficient for prevention alignment.
• Belief 4 (atoms-to-bits): reinforced — continuous monitoring (CGMs, wearables) could solve the GLP-1 adherence problem by identifying right patients and tracking response, turning population-level prescribing into targeted monitored intervention.
• Existing GLP-1 claim: needs scope qualification — "inflationary through 2035" is correct at system level but incomplete. Should distinguish system-level from payer-level economics. Price trajectory (declining toward $50-100/month internationally) may move inflection point earlier.

Sources archived: 12 across five tracks (multi-organ protection, adherence, MA behavior, policy, counter-evidence) Extraction candidates: 8-10 claims including scope qualification of existing GLP-1 claim, VBC adherence paradox, MA prevention resistance, BALANCE model design, multi-organ protection thesis

Session 2026-03-16 — GLP-1 Adherence Interventions and AI-Healthcare Adoption

Question: Can GLP-1 adherence interventions (digital behavioral support, lifestyle integration) close the adherence gap that makes capitated economics work — or does the math require price compression? Secondary: does Epic AI Charting's entry change the ambient scribe "beachhead" thesis?

Key finding: Two findings from this session are the most significant in three sessions of GLP-1 research: (1) GLP-1 + digital behavioral support achieves equivalent weight loss at HALF the drug dose (Danish study) — changing the economics under capitation without waiting for generics; (2) GLP-1 alone is NO BETTER than placebo for preventing weight regain — only the medication + exercise combination produces durable change. These together reframe GLP-1s as behavioral catalysts, not standalone treatments. On the AI scribe side: Epic AI Charting (February 2026 launch) is the innovator's dilemma in reverse — the incumbent commoditizing the beachhead before standalone AI companies convert trust into higher-value revenue.

Pattern update: Three sessions now converge on the same observation about the gap between VBC theory and practice. But this session adds a partial resolution: the CMS BALANCE model's dual payment mechanism (capitation adjustment + reinsurance) directly addresses the structural barriers identified in March 12. The attractor state may be closer to deliberate policy design than the organic market alignment I'd assumed. The policy architecture is being built explicitly. The question is no longer "will payment alignment create prevention incentives?" but "will BALANCE model implementation be substantive enough?"

On clinical AI: a two-track story is emerging. Documentation AI (Abridge territory) is being commoditized by Epic's platform entry. Clinical reasoning AI (OpenEvidence) is scaling unimpeded to 20M monthly consultations. These are different competitive dynamics in the same clinical AI category.

Confidence shift:

• Belief 3 (structural misalignment): partially resolved — the BALANCE model's payment mechanism is explicitly designed to address the misalignment. Still needs implementation validation.
• Belief 4 (atoms-to-bits): reinforced for physical data, complicated for software — digital behavioral support is the "bits" making GLP-1 "atoms" work (supports thesis). But Epic entry shows pure-software documentation AI is NOT defensible against platform incumbents (complicates thesis).
• Existing GLP-1 claim: needs further scope qualification — the half-dose finding changes the economics under capitation if behavioral combination becomes implementation standard, independent of price compression.

Sources archived: 9 across four tracks (GLP-1 digital adherence, BALANCE design, Epic AI Charting disruption, Abridge/OpenEvidence growth) Extraction candidates: 5-6 claims: GLP-1 as behavioral catalyst (not standalone), BALANCE dual-payment mechanism, Epic platform commoditization of documentation AI, Abridge platform pivot under pressure, OpenEvidence scale without outcomes data, ambient AI burnout mechanism (cognitive load, not just time)

Session 2026-03-18 — Behavioral Health Infrastructure: What Actually Works at Scale?

Question: What community-based and behavioral health interventions have the strongest evidence for scalable, cost-effective impact on non-clinical health determinants — and what implementation mechanisms distinguish programs that scale from those that stall?

Key finding: Non-clinical health interventions are NOT a homogeneous category. They fail for three distinct reasons: (1) CHW programs have strong RCT evidence (39 US trials, $2.47 Medicaid ROI) but can't scale because only 20 states have reimbursement infrastructure; (2) UK social prescribing scaled to 1.3M referrals/year but has weak evidence (15/17 studies uncontrolled, financial ROI only 0.11-0.43 per £1); (3) food-as-medicine has massive simulation projections ($111B savings) but the JAMA Internal Medicine RCT showed NO significant glycemic improvement vs. control. The exception: EHR default effects (CHIBE) produce large effects (71%→92% statin compliance), reduce disparities, and scale at near-zero marginal cost by modifying the SYSTEM rather than the PATIENT.

Pattern update: Four sessions now reveal a consistent meta-pattern: the gap between what SHOULD work in theory and what DOES work in practice. Sessions 1-3 showed this for VBC (payment alignment doesn't automatically create prevention incentives). Session 4 shows the same gap for SDOH interventions (identifying non-clinical determinants doesn't automatically mean fixing them improves outcomes). The food-as-medicine RCT null result is particularly important: observational association (food insecurity → disease) ≠ causal mechanism (providing food → health improvement). The confounding factor may be poverty itself, not any single determinant.

Cross-session pattern deepening: The interventions that WORK (CHW programs, EHR defaults) modify the system or provide human connection. The interventions that DON'T reliably work in RCTs (food provision, social activities) provide resources without addressing underlying mechanisms. This suggests that the 80-90% non-clinical determinant claim is about the DIAGNOSIS (what predicts poor health) not the PRESCRIPTION (what fixes it). The prescription may require fundamentally different approaches — system architecture changes (defaults, workflow integration) and human relational models (CHWs, care coordination) — rather than resource provision (food, social activities).

Confidence shift:

• Belief 2 (non-clinical determinants): COMPLICATED — the 80-90% figure stands as diagnosis but the intervenability of those determinants is much weaker than assumed. Food-as-medicine RCTs show null clinical results. The "challenges considered" section needs updating.
• Existing SDOH claim: needs scope qualification — "strong ROI" applies to CHW programs but NOT to food-as-medicine or social prescribing (financial ROI). Should distinguish intervention types.

Sources archived: 6 across four tracks (CHW RCT review, NASHP state policy, Lancet social prescribing, Tufts/JAMA food-as-medicine, CHIBE behavioral economics, Frontiers social prescribing economics) Extraction candidates: 6-8 claims: CHW programs as most RCT-validated non-clinical intervention, CHW reimbursement boundary parallels VBC payment stall, social prescribing scale-without-evidence paradox, food-as-medicine simulation-vs-RCT causal inference gap, EHR defaults as highest-leverage behavioral intervention, non-clinical interventions taxonomy (system modification vs. resource provision)