Modeling Radioresistance and Radiotropic Fitness — Simulation Suite

Kinder & Faulkner · manuscript draft, not submitted · branch research/observable-split-and-pilot-hardening

What this is

A mechanistic modelling framework for a seven-species biofilm community under ionising radiation, plus the calibration and transport machinery that decides what the framework is allowed to claim.

Four things run. A 3D Cellular Potts model (biofilms_potts.jl) evolves biomass parcels on a cylindrical lattice under a four-term Metropolis acceptance functional — adhesion, volume, radiation and melanin — with a static radial radiation field and a genuinely two-way melanin field. A cylindrical radiodialysis transport model (biofilms_radiodialysis.R, and a second implementation inside the CPM) solves a three-equation contaminant/sorbate/membrane system in which membrane permeability is a dose-driven field rather than a fixed material constant — the most interesting mechanism in the repository. An OpenMC photon-transport path (coupling/) builds transport models from a lattice snapshot, tallies heating, converts it to absorbed dose and attributes that dose to lineage labels. A calibration layer (calibration/) gates all three, and currently refuses to emit a single physical unit.

julia --project=. biofilms_potts.jl    # coupled CPM + radiodialysis, N = 40, 100 MCS, seed 42

(That command has one open defect — see Dependencies.)

The framework is uncalibrated. Its parameters are literature priors and declared model inputs. No lattice pitch, no seconds-per-Monte-Carlo-step and no material density has been selected, and no output has been compared to a measurement of a target system. Every simulation number below is reported in the dimensionless model units it was computed in.

On the manuscript's status: two web searches on 2026-08-15 found no record of a bioRxiv posting, and no DOI appears anywhere in the repository. Its own final section is titled "Notation Fixes Required Before Submission". Treat it as unposted — that is a search result, not proof.

---

Which radiation phenotype this model represents

Five phenomena travel under one loose word and they are not interchangeable. A survival curve cannot calibrate a directional term. The full argument is in docs/calibration/integration_contract.md.

Phenotype	What it means	Represented here?
radiotropic	directional redistribution toward or away from a source	yes — this is response.hamiltonian
radiotrophic	radiation measurably enhances growth or metabolism	no — requires growth, and there is none
radioresistant	elevated survival after exposure	no — requires death, and there is none
melanized_radioprotective	melanin measurably alters damage, survival or shielding	no — requires damage
radiation_responsive	molecular, morphological or material change after exposure	partly — response.melanin

compute_delta_H adds β_ion[s]·I_γ when a parcel gains a site, so a negative β_ion lowers the acceptance functional in high-dose sites and the parcel drifts up the gradient. That is a tropism. It was previously labelled "metabolic energy gain", which conflates a Metropolis acceptance functional in arbitrary units with an energy budget; the model creates no biomass for such a gain to act on.

Radiotrophy and radioresistance are therefore structurally absent here, for the same reason response.growth_survival is unsupported_by_current_model. They are not weakly supported. A model with no birth and no death cannot express either, and no quantity of data changes that without a model revision.

This has a consequence worth stating plainly: the published D10 evidence — which is strong, and is what anchors the per-species β_ion magnitudes — is radioresistance evidence being spent on a tropism coefficient. That substitution is legitimate as a declared modelling choice, and it is now declared rather than left implicit in a sign convention. The constant is RADIOTROPIC in biofilms_potts.jl:49, renamed from RADIOTROPHIC; biofilms_3d.R's species flag likewise.

---

The finding: the tropism is melanin-mediated, not β_ion-mediated

This is the most interesting single result in the repository, and it is a self-audit finding rather than a simulation output. At the shipped constants I0 = 1.0 and T_cpm = 5.0:

Term	ΔH	Metropolis acceptance bias
ΔH_rad, radiotropic species (β = −5e−5)	−5.0e−5	1.000010
ΔH_rad, most radiosensitive species (β = 7.5e−2)	+7.5e−2	0.985
ΔH_mel at the reported M = 1.44	−0.720	1.155

β_ion — the one parameter Table 2 tabulates per species, and the one the entire sign convention is written around — biases acceptance by one part in 10⁵ for exactly the species whose radial stratification is the headline result. The melanin term biases it by 15.5%: four orders of magnitude more, through a coefficient of 0.5 hard-coded at its call site in compute_delta_H, appearing in no table and in no configuration file.

Radiation still reaches the dynamics. It reaches them indirectly, because melanin_drive is copied from the radiation field:

radiation → production (α_M, tabulated) → M → ΔH_mel (0.5, untabulated) → tropism

Two consequences. A sensitivity analysis over β_ion would report the radiation response as negligible — right about the coefficient, wrong about the mechanism. And only the product α_M · k reaches the dynamics, so cpm.melanin_coupling and response.melanin are jointly identifiable at best and never separately. The coefficient is now ledgered as cpm.melanin_coupling in data/parameter_provenance.csv with status=blocked and sensitivity_rank=high. It is one of five rows in that ledger not ranked unknown, and the only one outside the A0 transport sweep — the other four (transport.mesh.base_dimension, transport.mesh.coarsening_factor, transport.batches, transport.particles) were ranked by that sweep and carry sensitivity_domain = transport_numerical, which is a numerical-convergence rank and not a biological one. Every biological-response row in the ledger is still unknown, because no sensitivity analysis has been run over any of them; this one was measured by hand.

For scale: adhesion differences in the hand-specified J matrix are O(4–20) per neighbour pair, so ΔH_adh remains the largest term and still sets the sorting structure. The ordering is adhesion ≫ melanin ≫ radiation.

---

What this repository cannot currently claim

Each refusal below is enforced in code and pinned by a test. None of them is a caveat added in prose.

Not claimable	Why	What would clear it	Where enforced
Any length in µm or cm	No lattice pitch is selected. A pitch is not identifiable from a cell volume alone: V_physical = V_sites · a³ constrains only a product. select_pitch() refuses without declared tolerances, refuses when any tolerance is unset, and refuses when no evaluation is marked held-out. An undeclared tolerance is never read as infinitely permissive.	A calibrated 3-D stack of a target system with a genuine held-out acquisition, plus populated tolerances in config/cpm_spatial_acceptance_template.toml.	docs/calibration/cpm_spatial_calibration.md, calibration/biofilm_calibration/spatial/pitch_selection.py
Any duration in seconds, or any dose in Gy	No seconds/MCS conversion exists. CPMParams.seconds_per_mcs ships NaN and accrue_dose! raises rather than run without it. The repository has two independent clocks — MCS count and radiodialysis seconds — with no declared conversion.	The dynamic observable is already selected (biomass_autocorrelation_decay, status provisional). dt_MCS = a² · S_sim / S_exp still needs a pitch, so this is blocked on the other half.	docs/calibration/time_observable_contract.md, data/calibration/spatial/time_observable.csv, biofilms_potts.jl
Any mass, concentration, or g cm⁻³	No wet bulk density and no closed elemental composition. data/calibration/voxel_binding.csv carries both numeric fields blank; blank means NOT MEASURED and is never read as zero. wet_bulk_density() raises when a wet mass is divided by a cells_only volume.	Replicate wet and dry mass with matched hydrated and dry blanks, a hydrated volume on a whole_biofilm_envelope basis, ash, and CHNS + ICP closing to 1 on a wet-bulk basis — on one specimen under one condition.	docs/calibration/material_basis_contract.md, calibration/biofilm_calibration/materials/
Growth, division, survival, or extinction	response.growth_survival is unfittable in principle, not merely unmeasured: the CPM has no growth dynamics and divide_cell! has no trigger.	A model change. No quantity of data clears it.	docs/calibration/integration_contract.md, calibration/biofilm_calibration/schema.py
Anything per cell	One CPM cell ID is a computational biomass parcel, with literal_generation_claim_allowed = false for all seven species. Reading a parcel-volume distribution as a cell-size distribution, or comparing parcel counts to colony-forming units, is refused.	A model in which one ID is one organism — again a model change, not data.	docs/calibration/cpm_entity_semantics.md, data/calibration/spatial/entity_semantics.csv
A per-organism spatial prediction for B. subtilis, C. sphaerospermum or A. niger	Three of the seven species have morphologies the model cannot hold at any pitch — rods and filaments need surface-area, elongation and connectivity terms compute_delta_H does not have. Verdict: MODEL_REVISION_REQUIRED.	Those terms.	docs/calibration/cpm_spatial_calibration.md, calibration/biofilm_calibration/spatial/representability.py
A physical apparatus	Nothing here defines a physical system. The legacy R visualisation uses normalised R = 1, L = 2, so no value in it becomes centimetres by assigning a unit. The CPM additionally forces L = 2R — a topology fact, not a resolution one — so domain.semantics = "microvolume" is the only honest declaration available.	A declared apparatus with an independent length scale, which would also relax L = 2R.	docs/physical_reference_system.md
A coupled dose-driven remediation loop	STOP CONDITION RETAINED: no OpenMC dose enters the radiodialysis model until the constitutive choice is explicitly selected. In the current code m and P_eff are independent open-loop functions of time; neither reads the other.	An explicit constitutive selection for how damage maps to permeability, then import_dose_field! on a real path.	docs/online_coupling_design.md, docs/audit_biofilms_potts.md §6, §10

Two rules the whole repository runs on:

• Empty is not zero. A blank numeric cell means "not measured" and reads back as None. Nothing here substitutes 0.0 for a missing measurement — the single most dangerous default in a calibration ledger.
• blocked is not unsupported_by_current_model. A blocked quantity awaits a measurement; an unsupported one awaits a model change, and no quantity of data will clear it.

---

Selected structure

The repository root also carries unsorted media and data artifacts — PDFs, MP4s, Tableau workbooks, TSV exports, loose GIFs and PNGs — that are not part of the framework and are not listed here. One of them is no longer unsorted: geolocator/ and the power-plant dataset beside it are a maintained subsystem with its own provenance and tests, described under Primary Fuel Source Geolocator. It is still outside the framework, and it is documented so that sharing a repository is not mistaken for coupling. .gitmodules declares an unrelated submodule (Lumped-Uncertainty-SLS-MPC) that is not checked out and is referenced by nothing; git clone --recurse-submodules pulls an MPC repository you did not ask for.

Biofilms/
├── biofilms_potts.jl              # 3D Cellular Potts Model + radiodialysis coupling (authoritative)
├── biofilms_potts_jacc.jl         # JACC/AMDGPU port (8-colour checkerboard)
├── biofilms_radiodialysis.R       # Radiodialysis membrane PDE — Shiny app
├── biofilms_3d.R                  # Cylindrical bioreactor — Shiny app (legacy)
├── biofilms.R                     # Flat-domain particle Langevin SDE + k-means (legacy)
├── reactor_decision_tree.R        # k-NN / logistic sketch over 4 fuel candidates (illustrative)
├── validate_serial.jl             # Serial-vs-JACC statistical comparison
├── export_checkpoint.jl           # HDF5 checkpoint / transport-snapshot export
├── scoby_3d.jl                    # empty (0 bytes)
├── Project.toml, Manifest.toml, LocalPreferences.toml
│
├── calibration/                   # biofilm-calibration (Python, numpy-only)
│   ├── biofilm_calibration/
│   │   ├── spatial/               # entity semantics, scale candidates, biomass-field
│   │   │                          #   coarse-graining, structure, occupancy, pitch selection,
│   │   │                          #   morphology, representability, time observable,
│   │   │                          #   ingest/readers, dataset screening
│   │   ├── materials/             # basis-tagged Quantity arithmetic, mixture/mass closure,
│   │   │                          #   seeded Monte-Carlo uncertainty, export gate
│   │   ├── integration.py         # the voxel binding where the two branches join
│   │   ├── acquisition.py         # acquisition contract, biosafety gate
│   │   ├── precision.py           # derived replicate counts, detectability
│   │   └── schema.py              # shared status / evidence vocabulary
│   ├── scripts/                   # vcholerae_pilot (surrogate), detectability_pilot,
│   │                              #   emit_synthetic_reference_config, reference_d_status
│   └── tests/                     # 14 modules, 237 tests collected
│
├── contract/                      # biofilm-contract (Python; NO dependencies)
│   └── physical_contract/         # vocabulary, MaterialSpec, composition closure,
│                                  #   source placement, git provenance, feedback
│                                  #   gate verdicts. calibration/ may not import
│                                  #   coupling/, so what both need lives here.
│
├── coupling/                      # biofilm-openmc (Python; numpy, h5py)
│   ├── biofilm_openmc/            # config, model, materials, mesh, dose, lineage,
│   │                              #   snapshot, results, fingerprint, convergence, drivers,
│   │                              #   feedback_uq, feedback_gate
│   ├── scripts/                   # a0_sweep, synthetic_e2e, import_dose_field.jl
│   ├── requirements.txt
│   └── tests/                     # 14 unit modules + tests/integration/ (5, OpenMC-gated);
│                                  #   122 collected (6 skip without OpenMC)
│
├── config/
│   ├── coupling_template.toml                    # 21 REQUIRED-but-unset keys
│   ├── cpm_spatial_acceptance_template.toml      # every threshold unset
│   ├── biofilm_material_acceptance_template.toml # every threshold unset
│   ├── reference_a0_water_phantom.toml           # the first populated configuration
│   ├── reference_synthetic_biofilm_e2e.toml      # synthetic; calibrates nothing
│   ├── reference_d_spatial_acceptance.toml       # FROZEN declarations
│   ├── reference_d_material_acceptance.toml      # FROZEN declarations
│   └── feedback_effect_policy_template.toml      # threshold UNSET on purpose
│
├── data/
│   ├── parameter_provenance.csv      # 44 rows × 21 columns, per-parameter provenance ledger
│   ├── sources.csv                   # 12 sources pinned by version / accessed_date / sha256
│   ├── a0_transport_convergence.csv  # 100 data rows: the whole A0 sweep, 5×4×5
│   ├── calibration/
│   │   ├── spatial/                  # entity semantics, structure, morphology, time observable,
│   │   │                             #   dataset_candidates.csv (+ .sha256),
│   │   │                             #   pilot_vcholerae.csv / .metadata.json / .preflight.json
│   │   ├── materials/                # blanks, bulk, components, elemental, metal loading (header-only)
│   │   ├── voxel_binding.csv         # the join; both numeric fields blank
│   │   └── baseline_condition.csv    # ships UNSET
│   └── research/                     # radiotrophic compatibility audit ledgers
│       ├── radiation_response_candidates.csv        # 64 data rows (116 preamble lines)
│       ├── radiotrophic_dataset_candidates.csv      # 64 data rows
│       └── radiotrophic_materialization_candidates.csv  # 26 data rows
│
├── docs/
│   ├── audit_biofilms_potts.md    # the contract the coupling work was built against
│   ├── physical_reference_system.md, openmc_stack.md, exchange_schema.md
│   ├── online_coupling_design.md, jacc_coupling_port.md, adder_applicability.md
│   ├── a0_transport_resolution_decision.md, branch_report.md
│   ├── calibration/               # 8 contracts: entity semantics, spatial, time observable,
│   │                              #   material basis, material calibration, biomass field,
│   │                              #   integration, acquisition
│   └── research/                  # radiotrophic compatibility audit, calibration map,
│                                  #   lab gap, pilot recommendations
│
├── tests/                         # runtests.jl, contract_csv.jl, deterministic_radiation.jl,
│                                  #   genealogy_tests.jl, checkpoint_io_tests.jl, fixtures/
├── preprint/                      # .tex (source of record), .md (revised derivative),
│                                  #   figures/ (4 × PDF + PNG). The built .pdf was REMOVED:
│                                  #   it was a pre-revision build carrying withdrawn claims
├── assets/                        # preview_bioreactor_3d.png, preview_radiodialysis.png
├── kmeans_species_trajectory (1).gif, biofilm_dynamics_7_species.gif   # embedded below
├── Citations.md                   # predates data/sources.csv, unmaintained
└── .github/workflows/             # coupling-tests.yml (4 jobs)

---

Mathematical framework

The equations in this section are the paper's target formalism. They are stated because they define the model being aimed at; what each program actually integrates is stated in Simulations and differs.

The fitness of each species $s$ is written as a partial stochastic differential equation coupling diffusion, directed transport, radiation, melanin-mediated transduction and inter-species forces:

$$ \partial_t F_s = \nabla \cdot (D_s \nabla F_s) - \nabla \cdot \left( \mu_s \sum_j P_{sj}(t) F_j \right) + R_s + \sigma_s \xi(t,\mathbf{x}) - \beta_{s,ion} \dot{D}_\gamma F_s + \gamma_s \Delta_s - \alpha_{s,nir} N F_s + \theta_s H_s + C_s $$

No program in this repository integrates this equation. No file defines a field $F_s$. biofilms.R and biofilms_3d.R integrate particle positions, not a field; the CPM is a Metropolis Markov chain. Three of the nine terms are dimensionally checkable as written; the remainder contain a symbol with no declared unit ($\Delta_s$, $\theta_s$, $H_s$, $C_s$, $\alpha_{s,nir}$).

The multi-species Hamiltonian:

$$ H = \sum_i \left[ \frac{1}{2} \rho_i v_i^2 + U_i(\mathbf{x}_i) \right] + \sum_{i \neq j} V_{ij}(r_{ij}) + \sum_k W_k(t,\mathbf{x}) $$

$\rho_i$ is intended as the wet bulk mass density of a computational biomass parcel; it is density_g_cm3 in data/calibration/voxel_binding.csv, currently blank and blocked on the material gate. There is no velocity or momentum state anywhere in the repository, so the kinetic term has no counterpart in any implementation, and the equation closes dimensionally only if every term is re-declared as an energy density. This is an open defect, carried forward from the April draft's own notation list.

Mutualistic pairwise interaction potential:

$$ V_{ij}^{mutual} = -\gamma \exp\left( - \frac{r_{ij}^2}{\sigma^2} \right) $$

In the CPM this is a reporting statistic, not a force: total_pairwise_energy is called once per snapshot and γ_mutual / σ_mutual appear nowhere in compute_delta_H. The two R models each use a different ad-hoc attraction law.

Radiation field (Beer–Lambert, cylindrical source), as implemented in the CPM:

$$ I(r) = I_0 \exp(-\kappa r / N) $$

$\kappa$ is dimensionless per lattice span, $I_0 = 1$, and $r/N$ is a lattice fraction. The field is maximal on the axis and decays monotonically outward — pinned by tests/deterministic_radiation.jl. No attenuation length in cm exists. biofilms_3d.R uses $I_0 e^{-\kappa r}$ with $\kappa$ per normalised radius, a different quantity under the same name.

Melanin reaction–diffusion:

$$ \frac{\partial M}{\partial t} = D_M \nabla^2 M + \alpha_M , n_{RF}(\mathbf{x}) , \dot{D}_\gamma(t,\mathbf{x}) $$

Implemented in form. In the code $n_{RF} \in \lbrace 0,1 \rbrace$ is a per-site indicator over the three melanin producers, not a density; $\nabla^2$ is a 6-point unnormalised stencil with no $\Delta x$; $D_M$ and $\Delta t$ are lattice-and-step units; and the drive field is the dimensionless radiation field, not a dose rate. The resulting $M$ is a field value, not µg. The equation has no saturation, decay or substrate-limitation term.

Radiodialysis membrane transport

The most carefully derived component in the repository, implemented twice — biofilms_radiodialysis.R (method of lines + deSolve::lsoda) and biofilms_potts.jl (method of lines + forward Euler with an automatic stability substep) — with the two implementations in agreement. It postdates the April manuscript and appears in no version of it.

Mobile contaminant (cylindrical reaction–diffusion):

$$ \frac{\partial c}{\partial t} = \frac{1}{r} \frac{\partial}{\partial r} \left( r D_{eff} \frac{\partial c}{\partial r} \right) - (k_{ads} X + k_{red} X_{red}) c + k_{des} s $$

Immobile phase (biosorption + bioreduction):

$$ \frac{\partial s}{\partial t} = (k_{ads} X + k_{red} X_{red}) c - (k_{des} + k_{loss}) s $$

Membrane damage and dose-driven permeability:

$$ \frac{dm}{dt} = -k_{dam} \dot{D} m, \qquad P_{eff}(t) = P_0 \exp(\alpha_P D_{cum}(t)) $$

Robin boundary condition at the membrane wall $r = R$:

$$ -D_{eff} \left. \frac{\partial c}{\partial r} \right|_{r=R} = P_{eff}(t) \left( c(R,t) - c_{ext} \right) $$

Permeability as an evolving field driven by cumulative dose, rather than a fixed material property, is the interesting mechanism here and it survives intact. What does not survive is the units in the Julia coupling, where every physically-labelled slot receives a non-physical value: $R$ is params.N/2 — 20 lattice units — fed into a cm² s⁻¹ operator; $X_{total}$ and $X_{red}$ are site-occupancy fractions placed in g cm⁻³ slots and multiplied by rates in cm³ g⁻¹ s⁻¹; $X_{red}$ counts one species' sites over all interior sites, so it is neither a biomass fraction nor a reducer fraction; and the radial biomass profile is collapsed by an unweighted mean() before it reaches the solver, so the uptake sink is spatially uniform.

The material gate records this as RADIODIALYSIS: BLOCKED — "blocked by a units error in the model, not by missing data." The gate blocks the biomass basis fed into the Julia coupling, not the PDE system or its discretisation, both of which are sound and cross-validated between the two implementations.

Also: $m$ decays while $P_{eff}$ grows, and neither reads the other. That is the retained STOP CONDITION.

---

Simulations

1 · Cellular Potts model — biofilms_potts.jl

Julia, stdlib only for the simulation core. Cylindrical lattice, Metropolis Monte Carlo, and a four-term effective energy in the acceptance path:

$$ \Delta H_{CPM} = \Delta H_{adhesion} + \Delta H_{volume} + \Delta H_{radiation} + \Delta H_{melanin} $$

The pairwise mutualistic energy is computed once per snapshot as a diagnostic and does not influence the dynamics. H_CPM is a Metropolis acceptance functional in arbitrary units, dissipated toward a minimum — not a conserved thermodynamic energy, and there is no symplectic integration anywhere in this repository.

Field coupling is partial: melanin is genuinely two-way (occupancy drives $M$; $M$ enters acceptance); the radiation field is static after init_radiation! and nothing feeds back into it; the nutrient field is written and never read by the dynamics.

julia --project=. biofilms_potts.jl                 # coupled CPM + radiodialysis
julia --project=. biofilms_potts.jl --no-radiolysis # CPM only

Run provenance, both branches: the coupled default is N = 40, 6 parcels per species, 100 MCS, seed 42. --no-radiolysis runs main(), which is N = 60, 8 parcels per species, 200 MCS, seed 42. All reported numbers below come from the coupled N = 40 run.

Two stale stdout banners are an open defect. biofilms_potts.jl:1088 and :1529 both print H = H_adh + H_vol + H_rad + H_pair + H_mel — five terms, where the acceptance path has four and H_pair is a diagnostic. :1541 prints Ḋ(R)=%.1f Gy/s for a quantity that is a dimensionless placeholder and for which seconds_per_mcs is NaN. Nothing enforces agreement between the banners and the code, which is why they drifted; a reader who trusts the terminal is told two things this README refuses.

Fig 1 — radial mean position by species over 100 MCS (N = 40, 6 parcels/species, seed 42). The plotted observable is the mean parcel distance from the cylinder axis, in lattice units normalised by R, sampled every 20 MCS. The radiation field is maximal on the axis, so the high-radiation region is the core, not the wall; β_ion is negative for the two radiotropic fungi and positive for B. subtilis.

Spatial sorting is dominated by the hand-specified adhesion matrix J. Within the radiation-derived pathway, melanin dominates β_ion by four orders of magnitude (see the melanin finding above). tests/deterministic_radiation.jl demonstrates the inertness directly: it must deliberately amplify $\beta$ to make multi-seed drift detectable, because the production values are not. The local ΔH sign per species is a contract; the collective drift is a result; the test keeps them separate.

Fig 2 — mean melanin field value over occupied sites, per producing species, to MCS 100. C. sphaerospermum reaches 1.44. The ordering among the three producers is set by the input α_M (0.14 versus 0.10 and 0.065), and the linear rise is structural — the field equation has no saturation, decay or substrate-limitation term. What is not set anywhere in the parameters is the spatial distribution and magnitude of $M$: those are outcomes of where parcels went under the acceptance test, multiplied by the field, integrated over time. That is a genuine emergent quantity, and it is the quantity that then feeds back into acceptance at 15.5% bias.

biofilms_potts_jacc.jl is a JACC/AMDGPU port using 8-colour checkerboard updates. Its contract is statistical, not pathwise, and docs/jacc_coupling_port.md records that it is never the first causal comparison vehicle. The literal comment # 13. Figure export in biofilms_potts.jl is a load-bearing split marker: tests/runtests.jl:14, validate_serial.jl:12 and biofilms_potts_jacc.jl:454 all load the serial monolith by splitting the source on that exact string and evaluating the first half in a sandbox module whose imports are hard-coded to LinearAlgebra, Statistics, Random, Printf. The two rules are: never edit that line, and add nothing above it that needs a non-stdlib package.

2 · Radiodialysis membrane transport — biofilms_radiodialysis.R

Shiny app. Method-of-lines finite-volume solver for the three-equation system, LSODA adaptive stiff integration (deSolve), Robin condition at the wall and L'Hôpital treatment at $r = 0$. Four tabs: $c(r,t)$ heatmap, $s(r,t)$ heatmap, membrane integrity / $P_{eff}$ time series, radial snapshots.

Rscript biofilms_radiodialysis.R    # serves on 127.0.0.1:${SHINY_PORT:-7799}, no browser launch

Rscript is not present on the machine this README was verified against, so both Rscript commands here are untested as written. The R app's units are declared but unvalidated. The Julia coupling's units are wrong by construction (above).

Fig 3 — membrane integrity $m(t)$ and permeability ratio $P_{eff}/P_0$ over 100 MCS. $m$ falls to 0.779 and $P_{eff}/P_0$ reaches $e \approx 2.72$. Neither is a physical dose response: $\dot{D} = 1.0$ is a hard-coded placeholder never derived from state.radiation, dt_rd = 0.5 carries the source comment "tune to match CPM time scale", and seconds_per_mcs is NaN, so no conversion to grays exists. The permeability ratio is exactly $e$ because $\alpha_P \dot{D} \Delta t_{rd} n_{MCS} = 1$ by construction from four hand-set constants.

Fig 4 — radial contaminant profile $c(r,t)$ and the sorbed phase $s$. This is the numerical solution of a correctly derived cylindrical PDE under a Robin wall condition, and as a transport result it stands. It is not a biological localisation result, because the uptake sink is spatially uniform: the radial biomass profile is averaged to a scalar before it reaches the solver. Three further readings the figure does not support on its own: wall concentration $c(R,t)$ is dominated by the imposed Robin condition, so it reads the boundary rather than transport; the interior mean is an unweighted mean over radial nodes, which under-weights the outer, contaminant-rich annulus of a cylinder; and $c(t=0) \equiv 0$ everywhere, so a low interior value measures non-penetration, not removal.

Stale figure and label notice. All four committed figures predate the 2026-08-14 correction that neutralised the reversed biological zone labels and removed the hard-coded dose annotation, so they differ from current export_figures output. In addition, the embedded Fig 4 prints (1.0 − c_mean) * 100 under the label "% depleted" — the figure says depleted where the model supports only non-penetrated. The preview image above assets/preview_bioreactor_3d.png likewise predates the 2026-08-14 radiation sign fix.

3 · Legacy R models

Both are declared legacy and non-authoritative. The Julia seven-species registry is authoritative.

biofilms_3d.R — cylindrical bioreactor, Shiny app. Overdamped dynamics in a cylinder of radius $R = 1$, length $L = 2$ in normalised units, with a radial Beer–Lambert field. Radiotropic species (C. neoformans, C. sphaerospermum) are attracted toward the high-radiation central axis; others drift outward. Five sliders: $I_0$, $\kappa$, nutrient $C_0$, thorium intensity, heat intensity.

Rscript biofilms_3d.R    # auto-prints the shinyApp object, which starts the server (untested here)

This file carries Pseudoalteromonas, which appears nowhere else in the repository, and omits O. intermedium. Its species ordering differs from both other registries, so any positional cross-file mapping is silently mismatched. Two open defects: heat and nutrient contributions are scalars added identically to the $x$, $y$ and $z$ force components, producing a spurious drift along $(1,1,1)$ — an axial force from a radial quantity; and the noise increment scales with $\Delta t$ rather than $\sqrt{\Delta t}$.

biofilms.R — flat-domain particle Langevin SDE. Seven entities (one an unnamed placeholder; O. intermedium absent) on a dimensionless $[0,1]^2$ domain, with species-specific mobility, an ad-hoc quadratic attraction, a radiation term, additive Gaussian noise and k-means clustering ($k = 4$, 7 points, re-fit every 10 steps, unseeded) over 1000 steps. Two defects: the same $\Delta t$-versus-$\sqrt{\Delta t}$ noise mis-scaling, so this is not Euler–Maruyama and the realised diffusivity is $D \Delta t$, step-size dependent; and there is no radiation field in this file — gamma_intensity is a scalar constant added identically to both velocity components, producing a uniform drift along the diagonal. Results are qualitative only and support no gradient or niche-partitioning claim.

 

---

Calibration and coupling

Neither subsystem appeared in the April documents. Together they are now the largest part of the repository and they govern what the rest may claim.

calibration/ — biofilm-calibration

An installable Python package, deliberately numpy-only, which must not import the coupling package. Two branches meeting at one contract:

• spatial — entity semantics, scale-candidate enumeration, biomass-field coarse-graining, structural observables, occupancy mapping, pitch selection, morphology and representability diagnostics, synthetic validation fields, the dynamic-observable contract, format-agnostic ingest with optional microscopy readers, and a public-dataset screening ledger.
• materials — basis-tagged Quantity arithmetic, mixture and mass-closure checks, seeded Monte-Carlo uncertainty, and an export gate.
• integration — the voxel binding that joins them, and the only place where a whole class of error becomes visible. docs/calibration/integration_contract.md names six rejected combinations whose individual clauses each pass their own branch's gate — the canonical one being a literal cell fragment, whose density comes from bulk wet biofilm, whose lineage is claimed as a biological cell lineage. Neither branch can see the contradiction alone. Only the join does.
• campaign — the acquisition contract, derived replicate counts and detectability.

It selects no pitch, exports no material and emits no config, because no measured dataset of a target system exists.

The gates refuse, and the refusal is the tested behaviour. emit_transport_config() prints REFUSING to emit a biofilm transport config with the full reason list. apply_occupancy() refuses while the mapping is unset, and mass_preserving refuses without a declared seed — an unseeded lattice cannot be reproduced or reviewed. precision.detectability() refuses and names the right remedy: more biomass per coupon, or a substrate with a lighter, more reproducible blank — not more replicates. The material export gate blocks proxy, derived_proxy, synthetic, declared and unresolved evidence, because shipping it would launder it into a measurement.

The dataset ledger records rejections as first-class results. data/calibration/spatial/dataset_candidates.csv holds 12 screened candidates: 1 use_first, 3 inspect_first, 2 diagnostic_only, 6 rejected. The informative half is what was turned down — a search that only records what it accepted is not a provenance trail. Two none_found rows are recorded as findings: no volumetric dataset covers the seven-species community, and no public record pairs a calibrated 3-D volume with wet mass, dry mass and matched blanks for one system. That is a search result, not proof of non-existence — but it is enough to conclude that the target campaign should be designed rather than waited for. The most instructive rejection is an excellent exact-species S. oneidensis single-cell tracking dataset, rejected on representational grounds because it answers the wrong question — independent confirmation that single_cell_msd was correctly rejected as a seconds/MCS observable.

Biosafety follows strains, not species. Several components are commonly BSL-1 while C. neoformans strains are commonly BSL-2, and collection material under one species name can carry different assigned levels, so a species list cannot determine a facility class. The enforced chain is exact strain IDs → institutional review → containment decision → approved procedures, and the gate blocks on a missing institutional_approval_id. Nothing in this repository describes a field-deployable system.

The V. cholerae pilot — it ran, and it is a loss curve

calibration/scripts/vcholerae_pilot.py ran against Dryad doi:10.5061/dryad.zcrjdfnph (CC0), with reference_system_id = public_vcholerae_surrogate and target_calibration = false written into its own output so the distinction survives being copied out of context. V. cholerae is not one of the seven modelled species. A successful run licenses exactly one claim — that the spatial calibration machinery was exercised against a named public 3-D biofilm reference system — and never that a biological pitch was calibrated.

Preflight found no file in the Dryad set that is both 3-D and time-resolved. The spatial pass therefore ran on Dataset_3_Fig3c_VPS-Bap1.nd2, a static Z = 94 × 3-channel × 1024² stack with voxel size (0.065, 0.065, 0.13) µm read from the file rather than typed, has_time_axis: false, time_metadata_source: "absent". The temporal observable stayed blocked — not deferred, blocked, because the file has no time axis to observe.

Two rows in data/calibration/spatial/pilot_vcholerae.csv:

coarsening	pitch µm	passed	biovolume frac	porosity	interface area	corr. length	thickness	comp. size q50
1	0.065	true	0.0	0.0	0.0	0.0	0.0	0.0
2	0.130	false	0.0	0.0	0.463	0.260	0.235	15.0

Row 1 passes by construction — it is the field compared against itself, so every error is identically zero and it is not evidence of anything. Row 2 is the result: one factor-of-two coarsening destroys every structural observable, worst of all the median connected-component volume at a relative error of 15.0, while the two intensive volume-fraction observables — biovolume fraction 0.047702 and porosity 0.952298 — hold exactly. Volume fraction is preserved by block averaging; geometry is not. That is the loss curve, and it is steep.

Both rows carry segmentation_basis = unresolved (the channels are fluorescent reporters, and a microscope records what a channel detected, not what it is intended to represent) and holdout_kind = none, since no frame of one file is an independent held-out stack — so select_pitch refuses these rows regardless of the numbers.

The dataset download was declined rather than defeated on the first attempt: the host's proof-of-work is a deliberate control against exactly the kind of client this is. Four SHA-256 digests are recorded so a manual copy verifies byte-for-byte, and the input digest is recorded as sha256_verified: true in the run metadata.

A defect was found and fixed by this pilot. nd2.ND2File.voxel_size() returns (1.0, 1.0, 1.0) for a file that carries no calibration, and the reader's old v > 0 guard passed that straight through as a calibrated 1 µm voxel. The guard now consults axesCalibrated (calibration/biofilm_calibration/spatial/readers.py). This is the pilot's most durable output.

coupling/ — biofilm-openmc

An installable Python package (deps exactly numpy, h5py) that builds OpenMC photon-transport models from a CPM lattice snapshot, tallies heating, converts it to absorbed dose, attributes dose to lineage labels, and evaluates a two-way-feedback gate offline. One console script, biofilm-dose-scan. The online (two-way) driver is NOT IMPLEMENTED; its design and its gates are recorded instead.

Real transport has run against a pinned stack: conda-forge openmc 0.15.3, Python 3.13, HDF5 1.14.6, official ENDF/B-VIII.0 photoatomic data with recorded SHA-256.

The OpenMC result is a feasibility benchmark, not a physical calibration. config/reference_a0_water_phantom.toml says it of itself: "This is a numerical-validation geometry, NOT an apparatus. Every value below is either evaluated nuclear data or a DECLARED modeling choice — none of it is measured, and none of it describes a real reactor." All dose values are voxel-averaged absorbed dose under OpenMC's charged-particle local-deposition approximation, not single-cell microdosimetry.

What it demonstrated:

• Energy balance closes to a relative error of 5.2 × 10⁻⁹ (1 000 000.0 eV deposited per 1 MeV source photon in a reflective water box), and uncollided transmission through 1, 2 and 3 cm of water is exponential within 15%.
• The feasibility run reported its own bad news: 75% of occupied voxels scored no heating at the 10 µm demo tally pitch, median relative error ≈ 1. Recorded, not hidden. The 10 µm feasibility mesh is a tally pitch and never a biological pitch.
• The A0 sweep (100 runs + 1 reference: 5 coarsening factors × 4 history counts × 5 seeds, plus a 48³ / 6.4 × 10⁷-history reference at an unswept seed) sets the A0 mesh and history count and establishes a hard ceiling of coarsening_factor ≤ 4. Sparsity is a scale problem, not a method problem, and the A0 geometry does not have it. Two metrics are reported because one of them cannot see discretization: field_diff_vs_reference has zero discretization error by construction on nested meshes, resolution_loss does not — conflating them would have made every mesh look convergent.

Every physical value ships REQUIRED-but-unset in config/coupling_template.toml (14 keys, staged transport → dosimetry → cpm_feedback → membrane_feedback), and the loaders enumerate every missing key in one error and refuse. biofilm-dose-scan exits 2 with the complete missing-key list rather than evaluating.

Two ideas worth stealing from docs/

Block sum versus block mean, in two packages, pinned by tests in both. Deposited energy is extensive, so coupling/.../mesh.py:coarsen_field block-sums; a biomass volume fraction is intensive, so calibration/.../spatial/field.py:coarse_grain block-averages. Same concept, opposite operation, and using the wrong one is off by factor³. The two packages deliberately do not import each other, so the distinction is restated where it can be seen. (The V. cholerae pilot is the same fact from the other side: block averaging preserved volume fraction exactly and destroyed every geometric observable.)

The array boundary is not an interface, and an edge-truncated chord is not a chord. The first version of specific_interface_area did count the boundary, and a test caught it: the metric changed when the same slab was cropped.

---

Radiotrophic compatibility audit

docs/research/radiotrophic_compatibility_audit.md (2026-08-15) is an evidence audit across eleven data repositories and the sequence archives, run as seven parallel retrieval tracks each re-verified by a second pass, with the ledgers in data/research/. It populates no parameter. Nine access controls were encountered and none was circumvented.

Question	Verdict
Exact radiotrophic data for the modelled species	TARGETED_LAB_EXPERIMENT_REQUIRED
CPM calibration from public data	PUBLIC_DATA_CAN_VALIDATE_PIPELINE_ONLY
Materialization (density, composition)	TARGET_WET_BULK_MEASUREMENT_REQUIRED
Integrated (one system, all components)	PUBLIC_COMPONENTS_FOUND_BUT_NOT_PAIRED

The headline: radiotrophy is not established for any of the seven species. The claim rests on one 2007 paper whose exposure was characterised by analogy rather than by a stated dosimeter, whose growth effects are single-digit percentages, and whose integrated cell dose is not stated. Against it stand a same-species independent report of no substantial melanin protection under gamma, a same-species negative melanin-pathway result at 1 and 3 kGy, a same-genus null under a TLD-anchored ¹³⁷Cs field, an isogenic null in A. niger across three ionizing modalities, and a formal energy-budget critique. Peer review struck the word from the flagship ISS title. No author contact resolves this, because the discriminating experiment has not been run by anyone.

Radioresistance, by contrast, is well evidenced — and it is the evidence that anchors the per-species β_ion magnitudes, which is exactly why the phenotype table above matters.

Radiotropism has its own candidate and it is blocked twice. Zhdanova 2004 reports directed hyphal growth toward a source; whether its dose gradient was measured or calculated is unverified, and directed hyphal elongation has no CPM representation, because a parcel is isotropic and there is no elongation or connectivity term. One blocker awaits a measurement, the other awaits a model change.

Public exact-species volumetric imaging exists for exactly one of the seven (B. subtilis), and for none of the four with radiation data. Two of the seven have no ionizing-radiation record in the audit at all. O. intermedium AM7 has a total public footprint of 1086 base pairs and one paywalled abstract, and has never been irradiated in any public experiment.

---

Key results

From the coupled CPM + radiodialysis run in biofilms_potts.jl: N = 40, 6 parcels per species, 100 MCS, seed 42, single realisation, no replicates. Not from the preprint — none of these numbers appears in it. Every value below is dimensionless; nothing in this table is in grays, seconds, centimetres or grams, because no conversion to any of those exists.

Result	Value	Status
C. sphaerospermum melanin, MCS 100	1.44	Dimensionless field units. Mean of a dimensionless field over occupied sites. The ordering among the three producers is set by the input α_M; the magnitude and spatial distribution are emergent.
Melanin acceptance bias at that value	1.155 (15.5%)	Dimensionless. Computed from the shipped 0.5 coefficient and T_cpm = 5.0; four orders of magnitude above the β_ion bias of 1.000010. The dominant radiation-derived term, and undeclared until 2026-08-15.
Membrane integrity after 100 MCS	m = 0.779	Dimensionless. No physical dose exists: Ḋ = 1.0 is a placeholder, dt_rd = 0.5 is an uncalibrated MCS→second knob, seconds_per_mcs is NaN, and accrue_dose! raises rather than run without it. There is no value of Gy this corresponds to.
Permeability ratio	P_eff / P₀ = e ≈ 2.72	Dimensionless, and exact by construction. A closed-form function of four hand-set constants ($\alpha_P \dot{D} \Delta t_{rd} n_{MCS} = 1$); the simulation computes no permeability in cm s⁻¹. The former "0.010 → 0.027 cm s⁻¹" was the Nafion-117 literature prior P₀ rescaled by that factor, and a literature prior is not a calibration.
Interior contaminant, unweighted node mean	c / c_ext = 0.024	Dimensionless ratio. Measures non-penetration, not depletion: c(t=0) ≡ 0 everywhere, so the interior never held contaminant to remove. The mean is unweighted over radial nodes, not volume-weighted. The embedded figure labels this "% depleted"; the figure is wrong and the model supports only non-penetration.
Radial position of parcels	C. neoformans mean r/R = 0.65; B. subtilis mean r/R = 0.50	Dimensionless, and not yet evidence. Both terms are lattice units, so the ratio is legitimate. But the figures come from main_coupled() at N = 40, six parcels, 100 MCS, so R = 20 and parcel centres are rejection-sampled uniformly over radial_dist < R − 4 = 16 = 0.8R, giving a centres-only null of E[r/R] = ⅔(0.8) ≈ 0.53 — the B. subtilis value sits 0.43 SE from it, indistinguishable from the initial condition; n = 6 parcels on one seed gives SE ≈ 0.077 against a 0.15 gap, ≈ 1.4 SE of the difference; and the direction is opposite to the implemented physics, since a negative β_ion biases C. neoformans toward the high-radiation axis. Note that 0.58 is itself only an approximation: parcels occupy volume and grow toward V_target, so the mean over occupied sites is not the mean over seed centres, and the true null has to be simulated rather than derived. No multi-seed run against a uniform-seeding control exists in the repository.
Pairwise community energy	−34.4 → −41.5	Model units, diagnostic only. total_pairwise_energy never enters Metropolis acceptance, so nothing in the simulation minimises it. It records mutualistic-pair proximity under adhesion, not a selected-for outcome.
Parcel persistence	42 / 42 seeded biomass parcels, where 42 = 7 species × 6 parcels is the ceiling as well as the floor	A property of the model, not a survival result. There is no division (divide_cell! has no trigger) and the only death path is volume erosion under a volume constraint that opposes it. Parcels are not cells. response.growth_survival is unsupported_by_current_model.

On the "self-regulating remediation loop." The intended design is: radiation damages the membrane → P_eff rises → more contaminant enters → metal-reducing S. oneidensis, co-located by CPM dynamics, immobilises it. Every arrow in that chain is currently open. m and P_eff are independent laws and neither reads the other; P_eff is a function of $t$ alone, so it is open-loop; the radial biomass profile is averaged to a scalar before reaching the transport model, so species location cannot influence uptake; and no feedback runs from biology to radiation or to the membrane anywhere in the repository. import_dose_field! exists to make a real link explicit, and no run currently invokes it. Wiring the loop is gated on the retained m-versus-P_eff stop condition. It is a design hypothesis; nothing in this repository tests it.

---

Gate verdicts

The vocabulary is the code's.

spatial:    READY_FOR_TIME_CALIBRATION | PROVISIONAL | MODEL_REVISION_REQUIRED | NOT_EVALUATED
material:   READY_FOR_OPENMC_BIOFILM_TRANSPORT | READY_FOR_RADIODIALYSIS_MAPPING
            | PROVISIONAL | BLOCKED | NOT_EVALUATED
feedback:   PASSED | FAILED | NOT EVALUATED
row status: ready | provisional | blocked | unresolved | not_applicable
            | unsupported_by_current_model

Gate	Verdict
CPM spatial	PROVISIONAL — entity semantics are declared for all seven species; everything else is missing
Material, OpenMC path	OPENMC: PROVISIONAL
Material, radiodialysis path	RADIODIALYSIS: BLOCKED — blocked by a units error in the model, not by missing data
Voxel binding	coherent, incomplete · Emission: REFUSED — a coherent sentence with blanks in it is still a sentence with blanks
Offline feedback gate	NOT EVALUATED — by design
Online (two-way) coupling	NOT IMPLEMENTED
JACC coupling port	design only — no JACC coupling code exists on this branch, deliberately
Time observable	SELECTED: biomass_autocorrelation_decay, row status = provisional (runner-up interface_rearrangement_rate); seconds/MCS still blocked on the pitch half
Baseline condition	Ships UNSET — a protocol contract, not a claim about a specimen that exists
Reference D declarations	FROZEN — all twelve modelling decisions made, reference_d_acceptance_v1; three declared criteria carry enforcement = not_implemented and may not close a gate
Reference D campaign	CAMPAIGN_READY: no — blocked by exactly one thing, D-APPROVAL, which only a biosafety committee can issue
Offline feedback gate	NOT_EVALUATED — no effect threshold declared, and none may be defaulted
Online feedback	DISABLED, fail-closed — every prerequisite defaults to its refusing value

Provenance ledger distribution (data/parameter_provenance.csv, 49 rows × 21 columns): status — 22 ready, 10 not_applicable, 9 blocked, 7 provisional, 1 unsupported_by_current_model. model_compatibility — 34 direct, 10 requires_transform, 5 unsupported. A row whose status is ready must carry a source_id; unresolved and blocked rows must not carry a value.

---

Tests

julia --project=. tests/runtests.jl              # 135 passed, 0 failed (re-run 2026-08-15 at HEAD)
julia --project=. biofilms_potts_jacc.jl --selftest

pip install -e "coupling[dev]"    && (cd coupling    && pytest -rs tests)   # 86 collected
pip install -e "calibration[dev]" && (cd calibration && pytest -rs tests)   # 173 collected

The Julia figure is 2 + 34 + 57 + 42 across the four suites, re-run on this tree rather than quoted from docs/branch_report.md, which records the older 2026-08-13 branch-end figures and is stale on the Python counts.

Run together in the coupling venv, the two Python suites give 259 passed, 4 skipped (re-run 2026-08-15 at HEAD). All four skips are coupling modules that skip on import openmc in a bare venv — the three in coupling/tests/integration/ plus coupling/tests/test_model_build.py. No calibration test skips: the pilot ND2 file is present on this machine, so test_pilot.py runs. The OpenMC integration tier is manual opt-in: activate the openmc-biofilms environment, set OPENMC_CROSS_SECTIONS, then run the coupling suite. CI is .github/workflows/coupling-tests.yml — julia-tests, python-unit, calibration-unit, and openmc-integration behind a workflow_dispatch input.

What the tests pin:

• tests/contract_csv.jl — byte-identical CSV output against a golden fixture, pinning the serial RNG stream. The fixture is Julia-1.12-dependent.
• tests/deterministic_radiation.jl — field shape (maximum on the axis, monotone radial decay), the exact local ΔH sign per species, and multi-seed drift direction with a deliberately amplified β, because the production β is too small to be detectable. The local sign is a contract; the collective drift is a result.
• tests/genealogy_tests.jl — the lifecycle/dose contract, and that the windowed API is identical to the legacy loop.
• tests/checkpoint_io_tests.jl — snapshot orientation probes and bit-exact restart resume.
• coupling/tests/test_mesh.py — the nesting invariant and the coarsening trap: taking the dose denominator from voxel_pitch_cm after coarsening understates every dose by exactly factor³.
• coupling/tests/test_fingerprint.py — that transport_state_hash excludes source.photons_per_second while dose_state_hash includes it.
• calibration/tests/test_spatial_schema.py — test_gate_is_provisional_not_ready, test_gate_returns_model_revision_required_for_a_literal_cell_claim, test_gate_is_not_evaluated_without_semantics.
• calibration/tests/test_material_basis.py — that site_occupancy is refused everywhere a concentration is required.
• calibration/tests/test_integration_binding.py — the six incoherent bindings.
• calibration/tests/test_precision.py — that extra imaging fields buy nothing when between-batch variance dominates.

---

Preprint

preprint/modeling_radioresistance_and_radiotropic_fitness.tex is the source of record: Modeling Radioresistance and Radiotropic Fitness in Multispecies Biofilms — A Computational Framework, Implementation Audit, and One-Way Radiation-Transport Foundation.

It reports executed computational work and an implementation audit. It infers no target physical parameter, and its transport section reports that the one-way CPM–OpenMC path executed rather than reporting results from it — no gate verdict, no lever magnitude, no dose in physical units. The Ethics statement records that the work involved no culture, no recombinant or synthetic nucleic-acid experiment and no irradiation, so no institutional approval was required for it, and that any future campaign waits on prospective strain-, protocol-, facility- and source-specific approvals.

It cites the parent of the commit that adds it. A document cannot cite the commit that contains it, which is the same structural limit the run artifacts record for their own provenance — git_commit there names the parent and carries a -dirty marker for the same reason.

The repository has moved past that revision, and some of what it moved past was our own published numbers. A false-positive control that could not fail, a squared difference reported as a variance, an unprovenanced sensitivity table, and one synthetic-gate verdict withdrawn as not resolution-converged. None of those numbers appears in the manuscript. The full record is data/claims_ledger.csv and docs/calibration/.

The superseded sources are in git history, not in the tree. modeling_radiotrophic_fitness.{md,tex} and the pre-revision PDF carried a fabricated Computational Methods section, a Th-232 comparison against data that exists in no data/ file, organisms named in no file here, and a symplectic integrator the Project.toml has never contained. Those claims are catalogued as 115 PP-* rows in the claims ledger, 34 of them delete.

And the ledger is now enforced. calibration/tests/test_claims_ledger.py asserts that no claim marked delete reappears in the manuscript. Until it existed, nothing in the repository checked the ledger at all.

Its real coverage is 20 of the 34, and the other fourteen are named in the test output as still needing a human. An earlier version of this note said "30 of 34", which counted rows that yield a searchable phrase rather than rows the guard can actually find — many ledger entries are reconstructions across the two superseded sources rather than verbatim quotes from either. The guard is calibrated against the pre-revision manuscript recovered from git history: it must detect at least 18 of the deleted claims there, because a guard that cannot find known-bad input is not a guard, and the first version of it detected 12.

DOI archiving is an author action: preprint-v1 is the tag to archive, and enabling Zenodo requires the account owner.

Figures are generated by biofilms_potts.jl below its split marker. The .tex includes them without an extension, so a pdflatex build resolves preprint/figures/*.pdf.

---

Primary Fuel Source Geolocator — a separate subsystem

This supports no claim in this repository. The geolocator is a visualization and infrastructure-data subsystem that shares this repository. It is not coupled to the Cellular Potts model, the radiodialysis transport stack, the calibration harness, or the claims gates. No gate reads it, no figure depends on it, and nothing it renders is evidence for anything the framework asserts.

It is documented for the opposite reason to an unexplained artifact. A stray file invites a reader to infer a relationship that does not exist; a named, maintained subtree invites the same inference unless the boundary is stated. This is the statement.

geolocator/ serves a globe of the world's power plants over the WRI Global Power Plant Database (v1.3.0, CC BY 4.0), vendored here as an immutable snapshot rather than fetched at run time, so a given commit always means a given dataset. power_plant_database_global.provenance.json records the publisher, version, licence, canonical source and the SHA-256 of the committed bytes. A test checks that hash against the file, because a provenance declaration that can drift from what it describes turns unknown into confidently wrong. The dataset is pinned -text, since under line-ending normalisation a content hash is not reproducible across platforms and the declaration would be unverifiable by construction.

The part worth borrowing is the freshness model, which refuses to let one signal stand for another. Reload detection and freshness authority are separate: a stat fingerprint decides when to reread a file, while a timestamp inside the data decides whether it is current. Touching a file rereads it and changes nothing else, because touching a file does not make the world newer. Sources declare what kind of thing they are — a reference dataset has a declared vintage and is never stale by clock, since a versioned snapshot does not drift toward wrong; a live source is measured against a clock; an authored source carries no semantic timestamp at all and reports that absence rather than letting a file modification time impersonate one. Absence of an authority is a state to publish, not a gap to paper over.

The stack is FastAPI and three.js with no build step; run it with python3 -m uvicorn geolocator.api:app. Its own tests live in geolocator/tests/. An optional local market adapter is not included in this repository, and its absence is reported explicitly rather than rendering as an empty result — an uninstalled producer must not read as a producer that found nothing.

The Tableau workbook and screen recording at the repository root are this subsystem's ancestor, retained for provenance. They plot the same dataset coloured by primary fuel. They are not the implementation, and neither they nor it bear on the biofilm work.

Dependencies

Julia 1.12 — the tested version; the CSV golden fixture pins the RNG stream to it.

Project.toml [deps] is exactly AMDGPU, HDF5, JACC. HDF5 is required by export_checkpoint.jl; JACC (with AMDGPU) by biofilms_potts_jacc.jl.

Open defect: CairoMakie is undeclared. biofilms_potts.jl:1874 does using CairoMakie at top level, below the # 13. Figure export split marker, so julia --project=. biofilms_potts.jl fails at that line unless CairoMakie happens to be present in the default environment. That is a missing dependency declaration, not a design choice: the split-marker rule constrains only what appears above the marker, and the sandbox module the splitters build has hard-coded imports that never consult Project.toml. Until it is declared, run the simulation from an environment that also has CairoMakie.

R — deSolve, shiny, plotly, ggplot2, dplyr, MASS, class (the last two for reactor_decision_tree.R).

install.packages(c("deSolve", "shiny", "plotly", "ggplot2", "dplyr", "MASS", "class"))

No R version floor is enforced or evidenced in the code, and no R command in this README has been executed on the machine it was verified against.

Python ≥ 3.11 (both pyproject.toml files) — two independent packages, neither importing the other. The pinned OpenMC stack is Python 3.13.

pip install -e "calibration[dev]"   # numpy only
pip install -e "coupling[dev]"      # numpy, h5py

OpenMC is optional and out-of-band: a pinned conda-forge stack (openmc 0.15.3, HDF5 1.14.6, ENDF/B-VIII.0 photoatomic data with recorded SHA-256), documented in docs/openmc_stack.md. Until it is installed, every OpenMC-dependent result is reported as blocked or skipped, never as passed.

---

Citation

The manuscript is unpublished and has no DOI. Cite the repository at a commit.

Kinder, H., Faulkner, B. (2026). Modeling Radioresistance and Radiotropic Fitness:
Spatiotemporal Dynamics of Microbial Communities Under Ionizing Radiation Stress.
Unpublished manuscript.
https://github.com/aurascoper/Biofilms, commit 5d1b777.

data/sources.csv is the working source registry, pinning each cited document by document_version, accessed_date and sha256, and it records honest failures alongside successes — two rows read NOT LOCATED and NOT VERIFIED against the primary report. Citations.md predates that registry and is not maintained.

License

See LICENSE.txt.