diff --git a/projects/ExampleMonteCarlo/.gitignore b/projects/ExampleMonteCarlo/.gitignore
new file mode 100644
index 0000000..ba39cc5
--- /dev/null
+++ b/projects/ExampleMonteCarlo/.gitignore
@@ -0,0 +1 @@
+Manifest.toml
diff --git a/projects/ExampleMonteCarlo/Project.toml b/projects/ExampleMonteCarlo/Project.toml
new file mode 100644
index 0000000..d7630d0
--- /dev/null
+++ b/projects/ExampleMonteCarlo/Project.toml
@@ -0,0 +1,34 @@
+name = "ExampleMonteCarlo"
+uuid = "00000000-0000-4000-8000-000000000003"
+version = "0.1.0"
+authors = ["Your Name <you@example.com>"]
+
+# COMPUTE env. No plotting backend and no Pinax — those are in `report/`, which reads the same vault
+# off disk. That separation is the reason this project can exist at all: the physics packages below
+# were each carrying Plots as a hard dependency until it was moved behind an extension, and a
+# `load=` of the work module ships whatever it pulls to every worker.
+
+[deps]
+ClassicalMonteCarlo = "faaa8aa3-7fc9-41cf-84af-9d0c24f78989"
+DataVault = "23f5f8f6-b4da-40ee-8c72-c53b6c5de94f"
+Lattice2D = "e8031020-b7ff-4f1d-bee4-f79aea4cf140"
+ParamIO = "938a3ac2-d340-473c-bcf1-88af577e4ccf"
+Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+SweepRunner = "be946ad2-3cb3-4b6e-8f7e-4a5ecc3c255b"
+
+[sources]
+# None of these four is in a public registry yet; ParamIO and DataVault resolve from General.
+ClassicalMonteCarlo = {rev = "main", url = "https://github.com/sotashimozono/ClassicalMonteCarlo.jl"}
+Lattice2D = {rev = "main", url = "https://github.com/sotashimozono/Lattice2D.jl"}
+SweepRunner = {rev = "main", url = "https://github.com/QAtlasHub/SweepRunner.jl"}
+
+[compat]
+DataVault = "0.7"
+ParamIO = "0.4"
+julia = "1.10"
+
+[extras]
+Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+
+[targets]
+test = ["Test"]
diff --git a/projects/ExampleMonteCarlo/README.md b/projects/ExampleMonteCarlo/README.md
new file mode 100644
index 0000000..d8cec93
--- /dev/null
+++ b/projects/ExampleMonteCarlo/README.md
@@ -0,0 +1,35 @@
+# ExampleMonteCarlo
+
+Ising on a square lattice, swept over size and temperature.
+
+```
+configs/   the sweep at three sizes — smoke / debug / production
+scripts/   compute.jl (the run), collect.jl (read it back as a table)
+batch/     SLURM submission; the only place site-specific lines belong
+report/    a SEPARATE environment that draws from the vault
+out/       DataVault writes here; gitignored
+```
+
+```bash
+julia --project=. -e 'using Pkg; Pkg.instantiate()'
+julia --project=. scripts/compute.jl configs/smoke.toml   # two points, seconds
+julia --project=. scripts/collect.jl configs/smoke.toml
+sbatch batch/run.sh configs/production.toml               # the same command, at size
+```
+
+## What this project shows that the other does not
+
+**A real work package.** The physics is `ClassicalMonteCarlo` over a `Lattice2D` lattice, both
+reached through the `LatticeCore` interface. `run!(…; load=ExampleMonteCarlo)` ships this project's
+module to the workers, and everything it reaches travels with it.
+
+**Why the compute environment has no plotting backend.** It is not hygiene. Both physics packages
+carried `Plots` as a hard dependency until it was moved behind an extension; with `load=`, that
+would have put a plotting stack on every worker.
+
+**Why every call is qualified.** `ClassicalMonteCarlo` exports `run!` and so does `SweepRunner`, and
+they mean different things — one advances a Markov chain, the other dispatches the sweep. Every
+module here is imported as `using X: X`, so neither can be called unqualified by accident.
+
+The grids crowd `kbT ≈ 2.269`, the exact 2D Ising transition, because a sweep earns its cost when
+the interesting thing happens between two of its points.
diff --git a/projects/ExampleMonteCarlo/batch/run.sh b/projects/ExampleMonteCarlo/batch/run.sh
new file mode 100755
index 0000000..b74f544
--- /dev/null
+++ b/projects/ExampleMonteCarlo/batch/run.sh
@@ -0,0 +1,14 @@
+#!/bin/bash
+#SBATCH --job-name=ising-sweep
+#SBATCH --nodes=1
+#SBATCH --ntasks-per-node=16
+#SBATCH --time=02:00:00
+#SBATCH --output=out/log/%x-%j.out
+#
+# Site-specific lines (partition, account, module load) belong here and nowhere else.
+#     sbatch batch/run.sh configs/production.toml
+set -euo pipefail
+CONFIG="${1:-configs/smoke.toml}"
+mkdir -p out/log
+julia --project=. -e 'using Pkg; Pkg.instantiate()'
+julia --project=. scripts/compute.jl "$CONFIG"
diff --git a/projects/ExampleMonteCarlo/configs/debug.toml b/projects/ExampleMonteCarlo/configs/debug.toml
new file mode 100644
index 0000000..0eadeae
--- /dev/null
+++ b/projects/ExampleMonteCarlo/configs/debug.toml
@@ -0,0 +1,24 @@
+# Ising on a square lattice, swept over size and temperature. The exact transition of the 2D Ising
+# model is at kbT/J = 2/ln(1+√2) ≈ 2.269, which is why the grids below crowd there: a sweep is worth
+# running when the interesting thing happens between two of its points.
+#
+# A LIST value is a swept axis; a SCALAR is fixed. In DataKey.params the entries appear under their
+# DOTTED names, e.g. "system.kbT".
+
+[study]
+project_name  = "ising"
+total_samples = 1
+outdir        = "out"
+
+[datavault]
+path_keys = ["system.L", "system.kbT"]
+
+[[paramsets]]
+
+[paramsets.system]
+L   = [8, 12]
+kbT = [2.0, 2.27, 2.5]
+
+[paramsets.numerics]
+nsteps = 2000
+burn   = 1000
diff --git a/projects/ExampleMonteCarlo/configs/production.toml b/projects/ExampleMonteCarlo/configs/production.toml
new file mode 100644
index 0000000..656c769
--- /dev/null
+++ b/projects/ExampleMonteCarlo/configs/production.toml
@@ -0,0 +1,24 @@
+# Ising on a square lattice, swept over size and temperature. The exact transition of the 2D Ising
+# model is at kbT/J = 2/ln(1+√2) ≈ 2.269, which is why the grids below crowd there: a sweep is worth
+# running when the interesting thing happens between two of its points.
+#
+# A LIST value is a swept axis; a SCALAR is fixed. In DataKey.params the entries appear under their
+# DOTTED names, e.g. "system.kbT".
+
+[study]
+project_name  = "ising"
+total_samples = 4
+outdir        = "out"
+
+[datavault]
+path_keys = ["system.L", "system.kbT"]
+
+[[paramsets]]
+
+[paramsets.system]
+L   = [8, 16, 32]
+kbT = [1.6, 1.8, 2.0, 2.15, 2.27, 2.4, 2.6, 2.8, 3.2]
+
+[paramsets.numerics]
+nsteps = 20000
+burn   = 5000
diff --git a/projects/ExampleMonteCarlo/configs/smoke.toml b/projects/ExampleMonteCarlo/configs/smoke.toml
new file mode 100644
index 0000000..9a418b4
--- /dev/null
+++ b/projects/ExampleMonteCarlo/configs/smoke.toml
@@ -0,0 +1,24 @@
+# Ising on a square lattice, swept over size and temperature. The exact transition of the 2D Ising
+# model is at kbT/J = 2/ln(1+√2) ≈ 2.269, which is why the grids below crowd there: a sweep is worth
+# running when the interesting thing happens between two of its points.
+#
+# A LIST value is a swept axis; a SCALAR is fixed. In DataKey.params the entries appear under their
+# DOTTED names, e.g. "system.kbT".
+
+[study]
+project_name  = "ising"
+total_samples = 1
+outdir        = "out"
+
+[datavault]
+path_keys = ["system.L", "system.kbT"]
+
+[[paramsets]]
+
+[paramsets.system]
+L   = [8]
+kbT = [2.0, 2.5]
+
+[paramsets.numerics]
+nsteps = 200
+burn   = 100
diff --git a/projects/ExampleMonteCarlo/note/README.md b/projects/ExampleMonteCarlo/note/README.md
new file mode 100644
index 0000000..5c61426
--- /dev/null
+++ b/projects/ExampleMonteCarlo/note/README.md
@@ -0,0 +1,4 @@
+# note/
+
+Working notes for this project — what was tried, what the numbers looked like,
+why a parameter range was chosen. Not documentation for others; a record for you.
diff --git a/projects/ExampleMonteCarlo/out/.gitignore b/projects/ExampleMonteCarlo/out/.gitignore
new file mode 100644
index 0000000..62ec09e
--- /dev/null
+++ b/projects/ExampleMonteCarlo/out/.gitignore
@@ -0,0 +1,6 @@
+# DataVault writes here. Results are not source; keep the directory, drop the data.
+*
+!.gitignore
+!log/
+log/*
+!log/.gitkeep
diff --git a/projects/ExampleMonteCarlo/out/log/.gitkeep b/projects/ExampleMonteCarlo/out/log/.gitkeep
new file mode 100644
index 0000000..e69de29
diff --git a/projects/ExampleMonteCarlo/report/Project.toml b/projects/ExampleMonteCarlo/report/Project.toml
new file mode 100644
index 0000000..31e8167
--- /dev/null
+++ b/projects/ExampleMonteCarlo/report/Project.toml
@@ -0,0 +1,20 @@
+# report/ — the render environment. SEPARATE from the compute env one level up: that one runs on a
+# cluster and carries no plotting backend, while this reads the same on-disk vault and draws it.
+# The reduction is not duplicated here — it is `../src/ExampleMonteCarlo.jl`, and both sides call it.
+#
+# No `name`/`uuid`: an environment, not a package.
+
+[deps]
+DataVault = "23f5f8f6-b4da-40ee-8c72-c53b6c5de94f"
+ExampleMonteCarlo = "00000000-0000-4000-8000-000000000003"
+ParamIO = "938a3ac2-d340-473c-bcf1-88af577e4ccf"
+Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
+
+[sources]
+ExampleMonteCarlo = {path = ".."}
+
+[compat]
+DataVault = "0.7"
+ParamIO = "0.4"
+Plots = "1"
+julia = "1.10"
diff --git a/projects/ExampleMonteCarlo/report/report.jl b/projects/ExampleMonteCarlo/report/report.jl
new file mode 100644
index 0000000..6288305
--- /dev/null
+++ b/projects/ExampleMonteCarlo/report/report.jl
@@ -0,0 +1,17 @@
+# Draw the sweep. Run with `--project=report`, never the compute env.
+#
+#     julia --project=report report/report.jl configs/production.toml
+
+using DataVault: DataVault
+using ExampleMonteCarlo: ExampleMonteCarlo
+using ParamIO: ParamIO
+
+const CONFIG = get(ARGS, 1, joinpath(@__DIR__, "..", "configs", "smoke.toml"))
+const OUTDIR = get(ENV, "DATAVAULT_OUTDIR", joinpath(@__DIR__, "..", "out"))
+
+vault = DataVault.Vault(CONFIG; run="phase1", outdir=OUTDIR)
+rows = ExampleMonteCarlo.summarise(vault)      # the same reduction scripts/collect.jl uses
+
+@info "reporting over" n = length(rows)
+# The Binder cumulant's crossing in L is where the transition shows up — draw it here, with Plots,
+# which is a dependency of THIS environment and of nothing that runs on a compute node.
diff --git a/projects/ExampleMonteCarlo/scripts/collect.jl b/projects/ExampleMonteCarlo/scripts/collect.jl
new file mode 100644
index 0000000..328dc59
--- /dev/null
+++ b/projects/ExampleMonteCarlo/scripts/collect.jl
@@ -0,0 +1,28 @@
+# The reader side. No plotting dependency, so it runs where the compute did.
+#
+#     julia --project=. scripts/collect.jl configs/smoke.toml
+
+using DataVault: DataVault
+using ExampleMonteCarlo: ExampleMonteCarlo
+using ParamIO: ParamIO
+using Printf
+
+const CONFIG = get(ARGS, 1, joinpath(@__DIR__, "..", "configs", "smoke.toml"))
+const OUTDIR = get(ENV, "DATAVAULT_OUTDIR", joinpath(@__DIR__, "..", "out"))
+
+vault = DataVault.Vault(CONFIG; run="phase1", outdir=OUTDIR)
+rows = ExampleMonteCarlo.summarise(vault)
+
+@printf("\n  %-5s %-7s %-11s %-11s %s\n", "L", "kbT", "E/N", "|M|", "Binder")
+println("  ────────────────────────────────────────────────────")
+for r in rows
+    @printf(
+        "  %-5d %-7.3f %-11.5f %-11.5f %.4f\n",
+        r.L,
+        r.kbT,
+        r.energy,
+        r.magnetization,
+        r.binder
+    )
+end
+@printf("\n  %d points\n\n", length(rows))
diff --git a/projects/ExampleMonteCarlo/scripts/compute.jl b/projects/ExampleMonteCarlo/scripts/compute.jl
new file mode 100644
index 0000000..9a9bce1
--- /dev/null
+++ b/projects/ExampleMonteCarlo/scripts/compute.jl
@@ -0,0 +1,36 @@
+#==============================================================================
+ compute.jl — the three-layer driver, over an Ising sweep.
+
+   ParamIO    : config TOML   -> Vector{DataKey}   (what to compute)
+   DataVault  : (study, run)  -> file storage      (where it goes)
+   SweepRunner: run!(work_fn) -> parallel runtime  (do it, lock-safe, resumable)
+
+     julia --project=. scripts/compute.jl configs/smoke.toml
+
+ Every module is imported as `using X: X` and every call is qualified. Here that is load-bearing
+ rather than tidy: `ClassicalMonteCarlo` exports `run!` and so does `SweepRunner`, and the two mean
+ different things — one advances a Markov chain, the other dispatches the sweep.
+==============================================================================#
+
+using DataVault: DataVault
+using ExampleMonteCarlo: ExampleMonteCarlo
+using ParamIO: ParamIO
+using SweepRunner: SweepRunner
+
+const CONFIG = get(ARGS, 1, joinpath(@__DIR__, "..", "configs", "smoke.toml"))
+const OUTDIR = get(ENV, "DATAVAULT_OUTDIR", joinpath(@__DIR__, "..", "out"))
+
+spec = ParamIO.load(CONFIG)
+keys = ParamIO.expand(spec)
+vault = DataVault.Vault(CONFIG; run="phase1", outdir=OUTDIR)
+
+SweepRunner.init_workers!(; mode=:auto)
+opts = SweepRunner.RunOpts(; stop_flag=get(ENV, "PM_STOP_FLAG", nothing))
+
+result = SweepRunner.run!(
+    ExampleMonteCarlo.work_fn, vault, keys; opts=opts, load=ExampleMonteCarlo
+)
+@info "phase1 complete" result
+
+ledger = DataVault.build_ledger(vault)
+@info "ledger written" ledger
diff --git a/projects/ExampleMonteCarlo/src/ExampleMonteCarlo.jl b/projects/ExampleMonteCarlo/src/ExampleMonteCarlo.jl
new file mode 100644
index 0000000..5c778bc
--- /dev/null
+++ b/projects/ExampleMonteCarlo/src/ExampleMonteCarlo.jl
@@ -0,0 +1,106 @@
+module ExampleMonteCarlo
+
+# This project's own code: one equilibrated chain, and the reduction over finished points.
+#
+# `work_fn` lives here rather than in `scripts/` because `run!(…; load=…)` ships THIS module to the
+# workers. Everything it reaches travels with it, which is why the compute environment is kept free
+# of plotting.
+#
+# Note the qualified calls. `ClassicalMonteCarlo` exports `run!` and so does `SweepRunner`; the sweep
+# calls one and a chain calls the other, so neither is imported unqualified anywhere in this project.
+
+using ClassicalMonteCarlo: ClassicalMonteCarlo
+using DataVault: DataVault
+using Lattice2D: Lattice2D
+using Random: MersenneTwister
+
+export work_fn, summarise
+
+"""
+    work_fn(key) -> Dict{String,Any}
+
+One `(lattice, L, kbT)` point of the Ising sweep: burn in with no observers, then accumulate a
+`ThermodynamicObserver` over `nsteps` sweeps and return the thermodynamics.
+
+Returns the payload; it never saves. The runtime writes what comes back, plus the `.done` marker.
+"""
+function work_fn(key)
+    L = Int(key.params["system.L"])
+    kbT = Float64(key.params["system.kbT"])
+    nsteps = Int(key.params["numerics.nsteps"])
+    burn = Int(key.params["numerics.burn"])
+
+    # `build_lattice` takes the topology TYPE, not an instance:
+    # `build_lattice(::Type{<:AbstractTopology{2}}, Int, Int)`.
+    lat = Lattice2D.build_lattice(Lattice2D.Square, L, L)
+    model = ClassicalMonteCarlo.IsingModel(; J=1.0, h=0.0)
+    alg = ClassicalMonteCarlo.LocalUpdate(;
+        rule=ClassicalMonteCarlo.Metropolis(),
+        selection=ClassicalMonteCarlo.RandomSiteSelection(),
+    )
+
+    N = Lattice2D.num_sites(lat)
+    rng = MersenneTwister(hash((L, kbT, key.sample)))
+    grids = rand(rng, [-1, 1], N)
+
+    # Burn-in carries no observer: measuring the approach to equilibrium would bias the average.
+    ClassicalMonteCarlo.run!(
+        rng,
+        grids,
+        lat,
+        model,
+        alg,
+        ClassicalMonteCarlo.AbstractObserver[];
+        kbT=kbT,
+        nsteps=burn,
+    )
+    obs = ClassicalMonteCarlo.ThermodynamicObserver(; interval=10)
+    ClassicalMonteCarlo.run!(
+        rng,
+        grids,
+        lat,
+        model,
+        alg,
+        ClassicalMonteCarlo.AbstractObserver[obs];
+        kbT=kbT,
+        nsteps=nsteps,
+    )
+    res = ClassicalMonteCarlo.get_thermodynamics(obs, kbT, N, model)
+
+    return Dict{String,Any}(
+        "L" => L,
+        "kbT" => kbT,
+        "N" => N,
+        "energy" => res["Energy"],
+        "magnetization" => res["Magnetization"],
+        "specific_heat" => res["SpecificHeat"],
+        "susceptibility" => res["Susceptibility"],
+        "binder" => res["BinderParam"],
+    )
+end
+
+"""
+    summarise(vault) -> Vector{NamedTuple}
+
+Every finished point, sorted. `scripts/collect.jl` and `report/report.jl` both call this, so the
+table printed on the cluster and the figure drawn afterwards cannot disagree.
+"""
+function summarise(vault)
+    rows = NamedTuple[]
+    for key in DataVault.keys(vault; status=:done)
+        d = DataVault.load(vault, key)
+        push!(
+            rows,
+            (;
+                L=d["L"],
+                kbT=d["kbT"],
+                energy=d["energy"],
+                magnetization=d["magnetization"],
+                binder=d["binder"],
+            ),
+        )
+    end
+    return sort(rows; by=r -> (r.L, r.kbT))
+end
+
+end # module ExampleMonteCarlo
diff --git a/projects/ExampleMonteCarlo/test/runtests.jl b/projects/ExampleMonteCarlo/test/runtests.jl
new file mode 100644
index 0000000..818bf15
--- /dev/null
+++ b/projects/ExampleMonteCarlo/test/runtests.jl
@@ -0,0 +1,27 @@
+using ExampleMonteCarlo
+using Test
+
+@testset "ExampleMonteCarlo" begin
+    @test isfile(joinpath(@__DIR__, "..", "configs", "smoke.toml"))
+
+    # One point, small enough to run in a test. The assertion is physics, not a threshold: deep in
+    # the ordered phase the Ising magnetisation is near saturation, and well above the transition it
+    # is near zero. A chain that silently stopped equilibrating fails both ways.
+    mk(L, kbT) = (;
+        params=Dict(
+            "system.L" => L,
+            "system.kbT" => kbT,
+            "numerics.nsteps" => 400,
+            "numerics.burn" => 200,
+        ),
+        sample=1,
+    )
+
+    cold = ExampleMonteCarlo.work_fn(mk(8, 1.2))
+    hot = ExampleMonteCarlo.work_fn(mk(8, 4.0))
+
+    @test cold isa Dict{String,Any}
+    @test cold["magnetization"] > 0.8          # ordered
+    @test hot["magnetization"] < 0.3           # disordered
+    @test cold["energy"] < hot["energy"]       # and the ordered phase is the lower-energy one
+end