rtSD Explorer

An interactive system-dynamics explorer for resistance-training adaptation — implementing the GET-PAID framework, with a 7-test physiological-plausibility suite that gates the model.

RM1 (dose-response): doubling weekly training volume produces a larger steady-state Fitness gain. One of seven physiological reference modes the simulator must satisfy.

Try it live (Shiny): get-paid.shinyapps.io/rtsd · Run locally · What's inside · Theory · Citation

---

What is this?

rtSD Explorer is the public interface to rtSD — a system-dynamics model of resistance-training adaptation built on the MEDv4 fitness-fatigue-signal ODE. The three-tier picture:

• GET-PAID is the theoretical framework: Generating Expected Timelines of Predicted Adaptations to Imposed Demands. It claims that imposed training demands produce predictable adaptation trajectories.
• rtSD is the specific system-dynamics implementation of GET-PAID for resistance training — a three-stock model (Fitness / Fatigue / Signal) originally built in Vensim and ported to R + Python here.
• rtSD Explorer is this tool — the Shiny + marimo interface that lets anyone interact with the model.

Drag the sliders, change the training program, watch:

• Fitness (long-run adaptation) rise as you train,
• Fatigue (short-timescale recovery debt) accumulate and decay,
• Signal (the secondary mediator that gates adaptation) ride between them,
• Performance = Fitness − α · Fatigue trace out the predicted progression.

It ships with a reference-mode validator — seven forward-simulation unit tests anchored to canonical RT literature (dose-response, detraining, ADL floor, saturation, …). The validator is the structural gate before anything more ambitious gets built on top of the model.

Why this exists

There's a long lineage of fitness-fatigue models in exercise science — Banister 1975, Calvert 1976, Busso 1991/2003, Fitz-Clarke, Morton, the PerPot family — but most live in figures, not in code you can actually run.

This repo packages one variant of that lineage so anyone can:

1. Click through the dynamics without compiling C++ or installing Stan.
2. Confirm the model behaves the way the science says it should — by re-running the seven-test validator.
3. Adapt the simulator to their own programming (custom periodization, RP-style mesocycles, return-from-break) using a simple event-list API.

It's an open-science explainer, packaged as a tool.

What's inside

Path	What it is
inst/shiny/rtsd/	The Shiny app. Self-contained R; deployable on shinyapps.io. Live at the URL above.
python/rtsd/rtsd.py	Marimo notebook port. Same model + UI; exports to WASM so it runs in a browser with no server.
python/rtsd/sd_diagram.py	A Vensim-style stock-and-flow diagrammer (matplotlib).
R/med_ode.R	The MEDv4 ODE right-hand side. Two variants: original (quadratic) + linear.
R/simulate.R	High-level simulator wrapper around deSolve::ode.
R/scenarios.R	Composable scenario builders: weekly training, detraining, return-from-break, custom event lists.
R/training_schedule.R	Forcing-function constructors — both Vensim's PULSE TRAIN parameterization and arbitrary event lists.
R/reference_mode_tests.R	The seven-test validator with a self-contained Euler integrator (no external solver).
tests/validate_reference_modes.R	Driver: runs the validator and writes results.csv + figures.
inst/vensim/MEDv4_secondary_signal.mdl	Read-only copy of the source Vensim model (provenance).
docs/theory.md	Plain-English explainer of the MEDv4 equations and what the seven reference modes test.
docs/figures/	RM1, RM4, RM5 validator output PNGs.
Dockerfile + docker-compose.yml	Reproducible runtime — R 4.5.2 + Python + all deps, one image, three entrypoints (Shiny / marimo / validator).

Run locally

Docker (recommended — one image, three runtimes)

Everything in one container; no host R or Python install needed.

# Build once:
docker compose build

# Pick what to run:
docker compose up shiny           # Shiny at http://localhost:3838
docker compose up marimo          # marimo editor at http://localhost:2718
docker compose run --rm validate  # Reference-mode validator (7-test gate)

The image is rocker/r-ver:4.5.2 + slim apt deps + Posit Public Package Manager binaries (so R packages install in seconds, not minutes) + marimo + numpy + matplotlib via pip. Source is bind-mounted so edits hot-reload without rebuilding.

Shiny (host R)

# Install the runtime deps:
install.packages(c("shiny", "bslib", "ggplot2", "patchwork",
                   "dplyr", "tidyr", "deSolve", "DiagrammeR"))

# Run from this repo's root:
shiny::runApp("inst/shiny/rtsd")

The Shiny app is self-contained — inst/shiny/rtsd/R/ carries its own copy of the simulator so the app can be deployed standalone.

Marimo (Python, also exports to in-browser WASM)

pip install marimo numpy matplotlib

# Interactive edit mode:
python -m marimo edit python/rtsd/rtsd.py

# Read-only run mode:
python -m marimo run python/rtsd/rtsd.py

# Export a static WASM build (Pyodide; runs in any browser):
python -m marimo export html-wasm python/rtsd/rtsd.py -o build/index.html --mode run

Reference-mode validator

Rscript tests/validate_reference_modes.R

Seven forward-simulation tests, each anchored to a citation from the RT literature. Writes data/validation/reference_modes/results.csv and renders illustrative PNGs. Runs in seconds; no external solver, no MCMC.

Expected output at v2 defaults: 7/7 PASS.

The seven reference modes (what the validator checks)

#	Reference mode	What it tests	Citation anchor
RM1	Dose-response	Doubling weekly training volume produces a larger steady-state Fitness gain (ratio > 1.3)	Schoenfeld 2017; Krieger 2010
RM2	Diminishing returns	Low-baseline subjects gain more relatively than high-baseline (relative-gain ordering)	Ahtiainen 2003; Suchomel 2016
RM3	Ceiling	Asymptotic Fitness under constant load (final-50wk CV < 0.05)	Bickel & Bamman 2011
RM4	Detraining decay	Fitness loss after training stops sits in 0.5–3 %/wk	Mujika & Padilla 2000; Bickel & Bamman 2011
RM5	ADL floor	With ADL, untrained Fitness floors at ≥ 85 % of baseline (vs collapse without)	Thom 2005; LeBlanc 2000
RM10	Plateau under constant load	After plateau, additional Fitness gain ≤ 10 % at t=200	DeLorme 1945; Kraemer & Ratamess 2004 ACSM
RM11	Adaptive heterogeneity	Cross-subject 5–95 % relative-gain range ≥ 0.3 (hyper-responders vs non-responders)	Hubal 2005 MSSE

Failing any of these at sensible parameter values means the structural model is wrong, not just the calibration. The validator runs before any downstream work.

Roadmap

This repo is the deterministic explorer + structural-plausibility gate. Companion work in progress:

• Hierarchical Bayesian calibration of the MEDv4 family against open RT cohort data — separate project; a peer-reviewed publication is in preparation.
• sdviz — a planned standalone Python library extracting the Vensim-style diagrammer + reference-mode dashboard plotting into a reusable visualization package for system-dynamics models.

When those land, this repo will gain a posterior/ layer that lets rtSD Explorer ride along a calibrated parameter cloud instead of just point estimates.

How to cite

@software{bowie_rtsd_2026,
  author = {Bowie, Jacob},
  title  = {rtSD Explorer: an interactive system-dynamics explorer for resistance-training adaptation},
  year   = {2026},
  url    = {https://github.com/JacobBowie/rtsd},
  version = {0.1.0}
}

See CITATION.cff for the structured form.

License

MIT.

Acknowledgements

• MEDv4 source model: in-lab Vensim system-dynamics build (provenance at inst/vensim/).
• Reference-mode validation idiom: Forrester 1961, Sterman 2000.
• FFM lineage: Calvert 1976, Banister 1975/1991, Busso 1991/2003.
• Built with: Shiny, marimo, deSolve, matplotlib.

Provenance note: rename from synthesim

This project was originally released as synthesim (v0.1.0, 2026-05-12). It was renamed to rtSD Explorer on 2026-05-26 after discovering that "SyntheSim" is a trademarked feature mark of Ventana Systems' Vensim DSS. The model and code are unchanged; only the name moved.