# A coupled illness, practice and observation model: specification for review 17 September 2026. Codex research review. **Proposed mathematical architecture, not an empirically established universal theory, fitted model, clinical protocol or treatment-selection system.** No efficacy coefficients, simulated treatment benefits or patient recommendations are supplied. This specification develops the seven-source [interacting-processes comparison](/research/program-landscape/interacting-processes-evidence-2026-09-17.md) and the [Measurement Roadmap](/measurement-and-experimentation/), especially §§2, 6.3–6.7 and 7–10. It reuses that bounded primary-source review; no additional literature search was performed. The purpose is to make the proposed integration precise enough to disagree with, measure and simplify. ## 1. What the model is for The model asks how a person's ongoing bodily processes, learned responses, explanatory commitments, available practices and care choices produce trajectories of experience, function, disease measures and burden. It also asks how observing those trajectories changes subsequent action. A course is one source of instructions, expectations, skills and social context within that system. It does **not** assume that disease must have ended before psychological or behavioral help is possible, that a regulatory process is necessarily learned or erroneous, or that an available practice can change every nervous-system process. Conversely, it does not assume every distressing experience is an accurate measure of disease activity. These possibilities require separate variables and condition-specific tests. The proposed model is hybrid: continuous-valued person states evolve alongside discrete curriculum stages. It is stochastic and driven by external inputs. Nonlinearity is permitted but must improve prediction relative to simpler alternatives. Feedback, persistence and an attractive diagram do not establish an attractor, a reset or an optimal controller. ## 2. States, inputs and outcomes For person `i`, omit the person subscript below. At decision epoch `t`, define: | Symbol | Meaning and boundary | |---|---| | `P_t` | Vector of condition-specific pathological and tissue states: relevant inflammatory activity, tissue damage, metabolic constraints, or pathogen burden **where that variable is justified**. A replicating pathogen, if modeled, is an evolving state rather than automatically a one-time trigger. An unmeasured component is uncertain, not zero. | | `R_t` | Vector of specified regulatory states and configurations: selected autonomic, endocrine, immune-resolution or response-gain variables. Include fast and slow components when identifiable. This is not one scalar called safety or dysregulation. | | `L_t` | Learned response dispositions, operationalized by defined cue-response, avoidance, anticipation or generalization tasks. Distinct from immune memory and from consciously endorsed explanations. Do not prespecify that a disposition is maladaptive. | | `B_t` | Explicit explanatory beliefs, expected consequences, confidence, trust and perceived available choices. A measured proposition about illness differs from a computational prior, a physiological parameter and agreement with a teacher. | | `K_t` | Acquired skills or ability to perform a defined practice, distinguished from attendance, confidence and symptom improvement. | | `A_t` | Attention and current sampling of bodily/environmental information. Its meaning, duration and purpose matter; attention is not a single harmful or helpful dose. | | `Q_t` | Vector of accumulated burden/resources affected by participation: time, fatigue, financial cost, emotional load and displaced opportunities/care. Keep components visible rather than silently summing incompatible units. | | `s_t` | Discrete course stage or participation status. It changes the offered repertoire, material and support, not biology by definition. Pausing, leaving and continuing independently are valid states. | | `W_t`, `C_t` | External exposures and context, respectively: environmental inputs, sleep/social setting, relationships and access constraints. Avoid treating chosen activity or chosen care as exogenous. | | `u_t` | Actual enacted inputs, separately logged: practice ingredient/dose, activity/rest/exposure, medication/clinical care, device intervention and relational/community contact. Offered, intended and enacted doses are different records. | | `v_t` | Explanatory/instructional material encountered. Record its content, timing and source separately from doing a practice. | | `m_t` | Monitoring action: what is requested, sampled, attended to or reported; modality, timing, wording, duration and whether feedback is shown. | Write `X_t = (P_t,R_t,L_t,B_t,K_t,A_t,Q_t)`. Partitioning `P` from `R` is a measurement decision for the specific disease; biology does not supply a universal pathology/regulation dividing line. Experienced symptoms `E_t` are a vector, for example sensation intensity, unpleasantness, fatigue and distress. They are distinct from recorded symptom reports `Y^S_t`. This distinction is an observation-model requirement, not an assertion that reports are false or that genuine benefit requires a biomarker. Function `F_t` concerns capacity and participation in context; activity performed is also an input and is not synonymous with capacity. Disease outcomes `D_t` are condition-specific. Burden `Q_t` and harms are additional outcomes. Keep all four outcome families separate, including when a clinical composite deliberately combines them. ## 3. Time-ordered generative specification Choose timestamps and a resolution appropriate to the hypothesized mechanism. The following order prevents future outcomes from causing earlier decisions. It is a model family: before fitting, specify dimensions, functions, initial distributions, noise laws and which candidate links are included or fixed to zero. ### 3.1 Before measurement: experience and the decision to monitor ```text E^-_t = g_E(P_t, R_t, L_t, A_t, B_t, C_t; theta, eta^E_t) F_t = g_F(P_t, R_t, E^-_t, K_t, participation_and_opportunity_history; eta^F_t) m_t ~ pi_M(. | H^P_t, E^-_t, s_t, constraints_t, Z^M_t) ``` `H^P_t` is information already available to the participant, including previously returned clinical results. The researcher has a different history `H^R_t`; neither party observes the full latent state. A symptom-triggered diary entry can depend on unrecorded experience. `Z^M_t` denotes a scheduled or randomized measurement assignment when present, not an assumed instrument in ordinary observational data. Function here is an epoch-level outcome; if functional capacity has additional memory beyond these arguments, include that capacity in the augmented state. ### 3.2 Measurement is an action, then an observation ```text X^m_t = M(X_t, m_t, C_t, duration^M_t; eta^M_t) E^m_t = g_E(P^m_t, R^m_t, L^m_t, A^m_t, B^m_t, C_t; eta^{Em}_t) Y^S_t = h_S(E^m_t, B^m_t, K^m_t, wording_t, recall_window_t; epsilon^S_t) Y^phys_t = h_phys(P^m_t, R^m_t, sensor_conditions_t; epsilon^phys_t) Y^D_t = h_D(P^m_t, R^m_t, relevant_report_components_t; epsilon^D_t) Y^F_t = h_F(F_t, opportunity_t, measure_t; epsilon^F_t) Y^Q_t = h_Q(Q^m_t, measure_t; epsilon^Q_t) ``` `M` allows a question or body scan to change attention, experience or burden, and allows an invasive measurement to have specified physical effects. Identity transformation is a testable nonreactive special case. The direction of these effects is not assumed: monitoring can be useful, neutral or burdensome. Information produced here can improve subsequent choices even when measurement also has costs. Each requested observation has a completion/quality indicator `O_{jt}`. Its probability may depend on severity, fatigue, burden, access and the request itself. A missing or invalid observation is recorded as missing, **not zero**. Model or assess this process rather than assume random missingness. The equations show same-epoch readings for clarity. For delayed results, attach each value to its **collection** state at `t-d_j` and add it to the participant's information only at its **return** time. For recall questionnaires, `h_S` maps the stated interval of experience, not just the current instant. A passive sensor can enter research history without entering participant history if feedback is withheld. ### 3.3 Choice of practices, activity and care Let `H^{P,+}_t` contain the old history and only observations actually returned before the decision. Let `v_t` be material encountered before action selection. ```text intended_u_t ~ pi_U(. | H^{P,+}_t, E^m_t, B^m_t, K^m_t, s_t, v_t, C_t, available_choices_t, Z^U_t) u_t ~ delivery(. | intended_u_t, capacity_t, access_t, interruption_t, adherence_context_t) ``` `pi_U` is a descriptive choice policy, jointly influenced by symptoms, expectations, skill, advice, costs and opportunity. It is not assumed rational, optimal or accurately informed. Ordinary clinical care remains separately governed; it is not limited to the course repertoire. `Z^U_t` is a randomized assignment only when the actual study provides one. Intention-to-treat and actual-dose estimands then answer different questions. Discrete stage `s_t` determines which course materials and practices are offered through a versioned repertoire `R_course(s_t)`. It does not guarantee their use or erase skills from previous stages. In a Primal Trust instantiation, Level 1 tools remain available during deeper work and after membership ends; the observed invitation to advance is not a verified physiological readiness classifier. ### 3.4 Evolution during the following interval Let `D_t^lag` be a causal history/queue of prior states, experiences, actions and exposures sufficient for the prespecified delays. A symbolic transition family is: ```text P_{t+1} = F_P(P^m_t, R^m_t, u_t, W_t, C_t, D_t^lag, Delta_t; theta, eta^P_t) R_{t+1} = F_R(R^m_t, P^m_t, L^m_t, B^m_t, A^m_t, u_t, W_t, C_t, D_t^lag, Delta_t; theta, eta^R_t) L_{t+1} = F_L(L^m_t, E^m_t, prior_cues_and_consequences_t, u_t, D_t^lag, Delta_t; theta, eta^L_t) B_{t+1} = F_B(B^m_t, v_t, H^{P,+}_t, E^m_t, C_t, Delta_t; theta, eta^B_t) K_{t+1} = F_K(K^m_t, u_t, v_t, C_t, Delta_t; theta, eta^K_t) A_{t+1} = F_A(A^m_t, E^m_t, B^m_t, u_t, C_t, Delta_t; theta, eta^A_t) Q_{t+1} = F_Q(Q^m_t, u_t, C_t, D_t^lag, Delta_t; theta, eta^Q_t) D_{t+1}^lag = shift_and_append(D_t^lag, X^m_t, E^m_t, u_t, W_t, v_t, m_t) ``` All right-hand sides use information available at `t`. Their innovations may share measured or explicitly latent causes; do not assume independence merely to obtain causal arrows. `F_L` can be a specified learning rule but is not automatically reinforcement learning or active inference. `F_B` may represent greater confidence, appropriate revision, uncertainty tolerance or increased commitment despite contrary results. It is not defined as improving accuracy. The next experience follows from the new state at the next epoch. Thus an action's outcome informs subsequent learning and choice when it becomes available, not retrospectively. The next-epoch equations below define the values reused in §3.1, not additional draws of the same experience and function. If an explanation, practice and feedback response occur within one session, represent them as successive epochs; do not collapse their order into one daily observation and infer within-session causality. ```text E^-_{t+1} = g_E(P_{t+1}, R_{t+1}, L_{t+1}, A_{t+1}, B_{t+1}, C_{t+1}; eta) F_{t+1} = g_F(P_{t+1}, R_{t+1}, E^-_{t+1}, K_{t+1}, participation_and_opportunity_history; eta^F) D_{t+1} = g_D(P_{t+1}, relevant_R_{t+1}; eta^D) s_{t+1} ~ pi_stage(. | s_t, H^{P,end}_t, completed_material_t, B_{t+1}, K_{t+1}, invitation_t, resources_t) ``` `H^{P,end}_t` contains only information received by the progression decision. The new stage affects the next offered repertoire. Progression, pausing or leaving can therefore select people on response, resources and belief. None is a recovery outcome by definition. A concurrent stage change and health change do not establish that advancement caused improvement. ### 3.5 Well-posedness and scale An implementation must give an initial joint distribution for `(X_0,s_0,D_0^lag)`, admissible state domains, proper transition/observation kernels, noise distributions, policy support and the timestamp/availability rules above. The specification can condition on recorded external inputs/context; a fully generative simulation would additionally need their transition laws and a content-delivery law for `v_t` conditional on stage, selected material and interpersonal contact. Variables affected by past actions must be modeled as such, not designated exogenous for convenience. Positive biological quantities require domain-preserving transitions. These definitions produce a forward generative model without algebraic loops. They do not establish parameter identifiability or an existing fitted solution. Fast recordings may require seconds/minutes; immune responses and recovery may require hours/days; learning, skills and treatment course may require weeks/months. These are study-design scales, not universal biological constants. Use a fine integration grid with coarser/asynchronous observations, or validated transition operators for unequal `Delta_t`. Do not mistake weekly module release, monthly worksheets or six-month commitment advice for biological time constants. For delayed effects one permissible form is a distributed input `sum_{ell=0}^L K_ell u_{t-ell}`. Its kernels, sign and horizon must be estimated or justified; no values are supplied here. Longer memory can instead use explicit compartments or an extended state. Prespecified delayed follow-up and recovery windows must be long enough for the condition under study. A short immediate observation cannot certify absence of delayed harm. ## 4. Where the seven papers constrain this model Evidence tags should attach to the **relationship and its population, intervention and timescale**, not to an entire node or the whole diagram. | Primary source | Specific model relationship it informs | What remains untested here | |---|---|---| | [Bartsch et al., 2015](https://doi.org/10.1371/journal.pone.0142143): 36 healthy adults during overnight sleep | Joint temporal structure among `Y^phys` signals conditional on sleep context. Time-delay stability can constrain observable coordination and motivate candidate `R` dynamics. | It does not identify directed `R_j → R_k` causation, learned illness maintenance, course effects or a unique nonlinear law. Sleep stage is context here, not curriculum stage. | | [Dobreva et al., 2021](https://doi.org/10.1113/JP280883): acute human endotoxin data | Explicit candidate `W_LPS → inflammatory states → thermal/pain-related/cardiovascular outputs`, with fitted coupling and an independent acute comparison. | Sustained pathological input and combined treatment scenarios are model predictions. No learned-response state or mind-body input was tested. Detailed equations from the author preprint have not been verified against final full text in this review. | | [Relouw et al., 2024](https://doi.org/10.1038/s41540-024-00473-y): bolus and prolonged LPS in healthy men | Nonlinear inflammatory/mRNA transition structure and links to heart rate/temperature, tested across doses and a three-hour infusion. Useful precedent for `F_P/F_R`, lag structure and held-out input predictions. | Infection input profiles were constructed for simulations, not independently observed infections. No cognition, treatment effects or `B/L → inflammation` test. Biochemical persistence does not instantiate `B` or learned `L`. | | [Queirazza et al., 2024](https://doi.org/10.1016/j.bbi.2024.03.044): 19 healthy men, vaccine/placebo crossover | Assigned inflammatory challenge changed selected model-associated neural decision/feedback signatures, suggesting a candidate physiological-input-to-learning-system relationship to measure. | Task performance and fitted behavioral parameters did not significantly differ. This does not establish durable `P/R → L` change, reverse `L/B → P` effects, or a bidirectional clinical loop. Add a neural/task observation channel if replicating this test rather than relabel its signal as a belief. | | [Appelman et al., 2024](https://doi.org/10.1038/s41467-023-44432-3): long COVID with PEM, exercise and delayed follow-up | Condition-associated tissue/metabolic findings, activity input, before/after tissue measures and symptoms constrain candidate `u_activity → P/R → Y^phys/Y^S` trajectories and delayed windows. | The study did not identify a complete causal chain, nonlinear law or longer-term feedback loop. It neither establishes active replicating infection nor tests mind-body treatment; no individual safe-activity threshold follows. | | [Pradhan et al., 2007](https://doi.org/10.1002/art.23010): 63 adults with RA, MBSR plus reinforcement versus waitlist/usual care | Randomized package assignment to reported distress/well-being over six months; no disease-activity benefit detected. Directly motivates separation of experiential and disease endpoints while illness remains active. | It does not identify whether `A`, `B`, `K`, `L`, relationship, reporting or another component mediated benefit. A small null disease result does not fix every behavioral-to-disease coefficient to zero. | | [Tesser et al., online 2025/issue 2026](https://doi.org/10.1038/s41591-025-04114-7): active RA, implanted stimulation versus sham | Randomized `u_device → ACR20` difference at three months during continuing RA care. This tests a specific neural intervention and a composite clinical endpoint. | It does not identify all mediating `R → P` links, a learning process or a breathing/meditation dose. ACR20 includes clinical and reported components, so it is not a pure disease-state assay. Longer open-label follow-up has additional treatment/comparator limitations. | None of these seven directly estimates the full `course narrative → belief → chosen practice → learning/regulation → disease → reported outcome → progression` loop. The monitor-reactivity, belief-revision, skill-acquisition and progression functions are explicit research proposals here. The Primal Trust audit contributes the **content and offered sequence**, not evidence that these biological pathways occur. ## 5. Observability and causal identification Some measurements are direct for a limited quantity: a waveform, an analyte concentration, a joint examination or an enacted practice timestamp. They remain imperfect indicators of the broader latent states. HRV is not a direct reading of safety; a cytokine is not whole disease activity; agreement with an explanation is not neural learning; a symptom score is not a direct measure of tissue damage. Clinical composites can mix layers, so do not count their components as independent confirmation of one another. From symptom reports alone, experience-generation `g_E` and reporting `h_S` are generally inseparable: several combinations can produce the same record. Better wording, interviews, anchoring and multiple outcomes can constrain interpretation; they do not provide an objective readout of all subjective experience. Non-identifiability should narrow the mechanism claim, not erase participant-reported benefit or harm. Other problems are equally consequential: - People practice or seek care because they worsen; a negative practice–outcome association can reflect that indication. Participants who improve may advance, and those who deteriorate may leave or stop reporting. Conditioning on completion can select on consequences of treatment and illness. - Natural fluctuation, regression to the mean, concurrent care, changing exposure, expectation, reporting and persistent pathological input can mimic slow retraining. A fitted state called learning cannot resolve that ambiguity by its name. - Correlated innovations, unknown input doses, unmeasured prior load, sparse biomarkers and flexible observation models can leave many parameterizations observationally equivalent. More sampling of the same ambiguous variable does not necessarily identify the missing process. - An intervention RCT estimates a scoped assignment effect under its design. Mediation, actual-dose effects and transport to another disease or branded package need further assumptions and measurements. Predicting outcomes accurately also does not by itself identify the cause. Begin with the smallest condition-specific subset; calibrate instruments and repeatability; compare baseline/autoregressive and input-response models before adding hidden states. Record actual inputs and co-interventions. Use randomization or bounded perturbations only when authorized and appropriate, and account for time-varying confounding when analyzing endogenous choices. Preregister candidate lags and mediators, examine parameter recovery/identifiability, and test on new episodes/participants. Do not interpret a forecast as a validated treatment-selection rule. ## 6. Four testable restrictions of the model family These are separable contrasts, not four mutually exclusive worldviews. Each requires a named condition, outcome, intervention and follow-up horizon. Failure to detect an effect is informative only relative to precision, adherence, measurement validity and the size the study could detect. | Restriction / rival | Discriminating prediction and possible test | |---|---| | **Adjunctive benefit without an additional disease-modifying pathway:** for the specified behavioral ingredient, fix its paths into `P` to zero except through explicitly modeled changes in activity/exposure and ordinary care. Permit effects on experience, choices and selected regulation. | Symptoms, distress or participation may improve without disease change; disease could also improve through the allowed care/load routes. An adequately identified effect beyond those routes challenges this restriction. A randomized total disease effect alone cannot identify that path. These are structural restrictions, not instructions to adjust naively for post-treatment mediators. The RA MBSR result motivates the scoped contrast; it does not prove a universal zero. | | **Acute regulation without durable learned change:** set practice-driven persistent changes in `L` and its target response parameters to zero; retain temporary action effects and ordinary natural history. | Effects track sessions or current input. The competing durable-learning model predicts prespecified changes on later, unpracticed probes or generalized responses after acute effects dissipate. Repeated safe probes, reliable response measures, comparators and follow-up must distinguish lasting learning from exposure/care changes; reported improvement alone does not. | | **Nonreactive/invariant observation:** restrict `M` to no relevant state change and keep `h_S` invariant across the compared monitoring conditions, apart from modeled random error and requested recall interval. | Randomized low-burden cadence/wording/feedback conditions should not systematically change the targeted trajectory or report mapping. A difference rejects some part of this restriction. Experience and reporting effects may still be inseparable; separately record burden, function and any suitable physiology rather than calling all differences reinforcement. | | **Additive dynamics without the proposed state-dependent interaction:** retain input, lag and main effects but set a preregistered cross-term, such as `practice dose × baseline inflammatory state`, to zero. | The richer model must predict a reproducible interaction in held-out data, with adequate overlap or randomization to separate state dependence from who chooses a practice. A nonlinear curve, threshold or hysteresis claim needs its own defined prediction and comparison against simpler lag/clearance models. Statistical interaction alone does not establish the claimed physiological coupling. | ## 7. Two deliberately different instantiations ### Active inflammatory rheumatoid arthritis Define `P` using the specific inflammatory/joint disease processes and damage relevant to the cohort; choose a small measured `R` subset rather than assume a general neural cause. Observe symptoms and distress, participation, clinical disease components, care doses and burden separately. Distinguish beliefs about illness from measurable practice skill. Treat ongoing rheumatological care as an actual input, not a nuisance to delete. An MBSR package can enter as instruction, attention/practice dose and community context. The Pradhan trial supplies a bounded package-to-outcome precedent, with eight weeks of teaching and later reinforcement; it does not validate Primal Trust or populate its individual pathways. The implanted device in Tesser is a different input with its own dose, mechanism questions and adverse-event profile. It cannot calibrate a vagal-toning exercise parameter. A useful model comparison would ask whether a scoped intervention changes distress/function, disease measures, or both, and whether adding measured regulatory/learning states improves held-out predictions beyond care, disease state and natural fluctuation. An observed clinical benefit would still require mediation work before assigning it to the advertised mechanism. This is a study question, not advice to modify RA treatment. ### Long COVID with post-exertional malaise Keep tissue/metabolic constraints and prior exertional load available in `P` and the delay queue; do not set them to zero because the intervention discusses learned threat. Do not instantiate infection burden as known active infection from Appelman: that study does not establish it. Conversely, the absence of a confirmed measure is not proof that every ongoing biological input has ended. Record actual ordinary activity, symptoms over the relevant delayed window, participation, recovery and care. A coping or attention practice might change distress, reporting, activity decisions or a physiological response; these remain distinct hypotheses. Increased immediate activity with later functional loss cannot be labeled improvement by using only the immediate endpoint. Appelman supplies a reason to retain tissue processes and delayed measurements, not a fitted activity-threshold law, a mind-body mechanism or a safe exertional progression algorithm. Start from naturalistic activity and any separately justified clinically acceptable procedures; this model does not authorize replicating a maximal-exercise challenge. The informative comparison is whether a practice changes experience alone, changes exposure/load, or changes a reproducible response at comparable inputs without offsetting delayed harm. ## 8. Review conclusion and next concrete deliverable This specification makes the integration coherent as a **research bookkeeping and model-comparison framework**. It does not yet make it estimable as one large model. The next useful artifact is one condition-specific preregisterable specification: variable dictionary and units, minimal state dimension, actual ingredient and measurement schedule, candidate link masks, lag windows, primary outcomes, missingness assumptions, competing models and held-out evaluation plan. A prototype may visualize these relationships without manufacturing coefficients or efficacy curves. For the site, the attributed course map integrates illness, practice, observation and progression in one system. A companion research map uses the same kind of integrated representation to express competing possibilities beyond any one course. Label the connecting edges as instruction, choice, observation, hypothesized biological change or measured outcome. Confidence belongs to particular relationships and scoped sources; it should not color the entire combined model as established.