loadpath 0.1.1__tar.gz

loadpath-0.1.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Connor White
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

loadpath-0.1.1/PKG-INFO

@@ -0,0 +1,319 @@
+Metadata-Version: 2.4
+Name: loadpath
+Version: 0.1.1
+Summary: Generative design for printed parts: topology optimization from keep-outs, preserved geometry and loads, verified in real units.
+License-Expression: MIT
+Project-URL: Homepage, https://github.com/connorwhite-online/loadpath
+Project-URL: Documentation, https://github.com/connorwhite-online/loadpath/tree/main/docs
+Project-URL: Repository, https://github.com/connorwhite-online/loadpath
+Project-URL: Issues, https://github.com/connorwhite-online/loadpath/issues
+Keywords: topology optimization,generative design,3d printing,fea,cad
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Intended Audience :: Manufacturing
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Physics
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.26
+Requires-Dist: scipy>=1.11
+Requires-Dist: pyamg>=5.0
+Requires-Dist: scikit-image>=0.22
+Requires-Dist: trimesh>=4.0
+Requires-Dist: manifold3d>=3.0
+Requires-Dist: rtree>=1.0
+Requires-Dist: numba>=0.59
+Provides-Extra: render
+Requires-Dist: matplotlib>=3.8; extra == "render"
+Requires-Dist: shapely>=2.0; extra == "render"
+Requires-Dist: fast_simplification>=0.1; extra == "render"
+Provides-Extra: mcp
+Requires-Dist: mcp>=2.0; extra == "mcp"
+Requires-Dist: matplotlib>=3.8; extra == "mcp"
+Requires-Dist: shapely>=2.0; extra == "mcp"
+Provides-Extra: gpu
+Requires-Dist: torch>=2.2; extra == "gpu"
+Provides-Extra: flow
+Requires-Dist: numba>=0.59; extra == "flow"
+Provides-Extra: crosscheck
+Requires-Dist: gmsh>=4.11; extra == "crosscheck"
+Requires-Dist: pytetwild>=0.4; extra == "crosscheck"
+Requires-Dist: pyvista>=0.44; extra == "crosscheck"
+Provides-Extra: fast
+Requires-Dist: pypardiso>=0.4; extra == "fast"
+Provides-Extra: critic
+Requires-Dist: anthropic; extra == "critic"
+Requires-Dist: matplotlib>=3.8; extra == "critic"
+Requires-Dist: shapely>=2.0; extra == "critic"
+Provides-Extra: step
+Requires-Dist: cadquery-ocp>=7.7; extra == "step"
+Requires-Dist: embreex>=4.3; extra == "step"
+Provides-Extra: 3mf
+Requires-Dist: lib3mf>=2.4; extra == "3mf"
+Requires-Dist: pillow>=9; extra == "3mf"
+Dynamic: license-file
+
+# loadpath
+
+Generative design for 3D printed parts (and cast or milled ones), built to be
+driven by coding agents. Tell your agent what a part has to carry and where it
+mounts; loadpath grows the part with topology optimization, meshes it as a
+watertight solid with the interfaces cut exactly, and checks that mesh with a
+fresh finite element analysis. You get an STL, 3MF or STEP and a report: mass,
+safety factor, deflection, thinnest member, printability, pass or fail.
+
+<table>
+<tr>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/printed-hook.png" width="380"></td>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/printed-hook.png" width="380"></td>
+</tr>
+<tr>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/bell-crank.png" width="380"></td>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/bell-crank.png" width="380"></td>
+</tr>
+<tr>
+<td align="center" colspan="2"><sub>Left, the part loadpath delivered (blue: load, red: kept bore). Right, its
+stress from a fresh analysis of that mesh, blue (none) to red (the part's 98th percentile), on the shape bent by
+the worst load case. Top: a 300 N wall hook printed upright without supports, 24.2 g, SF 2.73. Bottom: a 500 N
+aluminium bell crank (LPBF), 23.4 g, SF 3.22. More in <a href="#examples">Examples</a>.</sub></td>
+</tr>
+</table>
+
+## Point your agent at it
+
+```bash
+pip install "loadpath[mcp]"
+claude mcp add loadpath -- loadpath mcp   # any MCP client: {"command": "loadpath", "args": ["mcp"]}
+```
+
+Then ask in plain words:
+
+> Design a PETG bracket for a 10 kg shelf. It screws to the wall with two
+> M5 bolts 40 mm apart and the shelf sits 60 mm out. Safety factor 2,
+> printed on an FDM printer.
+
+The agent writes the study, checks it, runs it in the background and reads
+the report back. Its guide is [loadpath/SKILL.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/loadpath/SKILL.md) (the
+`loadpath_guide` tool returns it); the tools and the loop are in
+[docs/agents.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/agents.md). The server runs locally and nothing
+leaves the machine. Units are millimetres, newtons and megapascals. MIT
+licensed.
+
+## What it does
+
+- **Topology optimization** on a voxel grid (SIMP with a density filter,
+  Heaviside projection and a robust minimum-member formulation), searching
+  for the lightest part that meets a safety factor and a deflection limit
+  across several load cases. A stress-constrained mode (MMA) is available.
+- **Verification of the delivered mesh**, not the optimizer's density
+  field: the STL is analysed again, on a tetrahedral mesh when the
+  `crosscheck` extra is installed and on a fine voxel grid otherwise. Keep-out
+  intrusion and lost interface geometry are measured with exact booleans.
+- **Process rules** for FDM (overhangs, bridges, layer bond), SLS and MJF
+  (trapped powder, escape holes, warp-prone panels), resin (supports, peel,
+  suction cups) and powder-bed metal (internal ceilings, powder
+  clean-outs), plus milling and casting constraints.
+- **Field modelling**: a part as a graph of implicit fields, so skins,
+  lattices (TPMS and strut) and fillets can be graded by stress or by
+  distance ([docs/fields.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/fields.md)).
+- **Other analyses**: large deflection and buckling, compliant mechanisms,
+  thermal loads, fatigue and creep estimates, lattice crush and impact,
+  vibration, lattice-Boltzmann flow and conjugate heat for heat exchangers
+  and cold plates, and layer-by-layer warp and residual stress
+  ([docs/BUILDSIM.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/BUILDSIM.md)).
+- **CAD in and out**: a STEP or STL part becomes a draft study with its
+  bores and bolt patterns found ([docs/IMPORT.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/IMPORT.md)); results
+  export as STL, 3MF, and STEP with exact planes and cylinders at the
+  interfaces; a lattice part (strut cells, TPMS, graded, skinned or open)
+  goes to STEP lattice and all, as smooth patches on its true surface ([docs/fields.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/fields.md#the-lattice-itself-in-the-step-lattice-exact)).
+- **Mesh checks without a study** (`loadpath.checks`): wall thickness,
+  overhangs, trapped voids, fluid-network isolation, component fit.
+
+## How much to trust it
+
+Read [docs/VALIDATION.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/VALIDATION.md) before relying on a number.
+In short:
+
+- **The solvers match analytic answers** in the test suite: beam
+  deflection within 3-5%, stress at a hole within about 13%, Lamé pressure
+  vessels within 5-11%, Poiseuille flow exactly.
+- **The structural check is cross-checked**: CalculiX and loadpath's own
+  tetrahedral solver agree within 0.1% on the same mesh. The voxel check
+  reads 24-49% higher stress than the tet check, so it errs safe.
+- **The optimizer is benchmarked**: 1-5% less stiff than top88 and top3d at
+  equal material on three of their textbook problems, stiffer on the fourth.
+- **Printed lattices match published tests**: against 33 published
+  compression tests of printed lattices, the stiffness check has a median
+  error of 1.23x (held out) and the strength check stays under the measured
+  value on all 33.
+- **Warp prediction is fitted to measured data** (NIST AM-Bench 2018 for
+  LPBF IN625, a published ABS study for FDM) and misses other alloys and
+  materials by 30-60%.
+- **Everything else is simulation only**, including every part mass and
+  safety factor in the case studies. The analysis is linear elastic with
+  datasheet material values; printed parts vary with the printer.
+
+## Case studies
+
+Each states the problem, the setup, the numbers against a named baseline,
+and the limitations ([docs/case-studies/](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/docs/case-studies/)).
+
+| Case study | Result | Baseline |
+| --- | --- | --- |
+| [Eval suite](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/evals.md) | 15 of 15 parts meet their requirements on a fresh analysis of the delivered mesh | Each part's own requirements |
+| [GRIP robot arm](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/grip.md) | A load sweep over the arm's URDF found 3 of 5 links below SF 1; the regrown set meets SF 2 and is 18% smaller | GRIP's committed parts, made by an earlier loadpath with hand-written loads |
+| [Arm link cores](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/arm-link-cores.md) | 42.2 g box section for FDM, 36.8 g hollow link for SLS, both at 0.5 mm and SF 3 | 54.9 g solid web; 54.2 g free-form topology optimization |
+| [Liquid cold plate](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/cold-plate.md) | A die-graded gyroid ties streamline channels; channels filled with a graded gyroid run 7% cooler | Streamline channels and pin fins, equal metal and pressure drop |
+| [TPU finger](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/tpu-finger.md) | up to 8 mm of tip curl for a 5 mm pull; a tendon finger grips 8x harder | A linear-model design (2.5 mm, then the solve fails); a plain bar of equal volume |
+| [Benchmarks](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/benchmarks.md) | 1-5% less stiff than top88/top3d at equal material, in 57-60 iterations instead of 130-410 | Ports of top88 and top3d |
+
+## Command line and Python
+
+```bash
+loadpath examples shelf-bracket > bracket.json
+loadpath check bracket.json            # every problem at once, with suggestions
+loadpath run bracket.json --out out/   # optimize, mesh, verify, render
+```
+
+`out/` gets an STL and a render per outcome and `report.json`. Every command
+takes `--quick` for a first look in a minute or two.
+
+Optional extras: `crosscheck` (tetrahedral verification; recommended),
+`render`, `step` (STEP in and out, OpenCascade), `flow` (lattice-Boltzmann,
+numba), `gpu` (PyTorch for large grids on CUDA, Apple Silicon or CPU
+threads).
+
+Studies are validated against a published JSON Schema
+([`schema/`](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/schema/)), and `loadpath check` returns every problem at once
+with its JSON path, so an agent can fix a study in one round trip. Long
+solves run as background jobs the agent polls. Each finished outcome has an
+`attention` list of numbers to look at before trusting it, and renders a
+vision model can inspect.
+
+From Python:
+
+```python
+from loadpath.spec import study_from_dict
+from loadpath.study import run_study
+
+result = run_study(study_from_dict(study))
+best = result.outcomes[result.best]
+best.mesh.export("part.stl")
+print(best.report["mass"], best.report["safetyFactor"])
+```
+
+## Documentation
+
+- [Quickstart](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/quickstart.md): install to a checked part.
+- [Tutorials](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/docs/tutorials/): eleven end-to-end workflows, each run by CI
+  (an FDM bracket to 3MF, an SLS lattice with drain holes, a metal heat
+  exchanger, an impact bumper, a STEP part round trip, LPBF distortion
+  compensation, and more).
+- [Reference](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/loadpath/README.md): the method, every study field, the
+  report, the limits.
+- [Validation](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/VALIDATION.md): what each model has been checked
+  against.
+- [Case studies](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/docs/case-studies/) and [benchmarks](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/BENCHMARKS.md).
+- Topics: [fields](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/fields.md), [materials](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/materials.md),
+  [build simulation](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/BUILDSIM.md), [calibration](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/CALIBRATION.md),
+  [CAD import](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/IMPORT.md), [slicer](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/slicer.md),
+  [multi-material](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/MULTIMATERIAL.md), [impact](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/IMPACT.md),
+  [mechanisms](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/MECHANISMS.md).
+
+<!-- gallery:start -->
+
+## Examples
+
+The eval suite (15/15 passing): robot-arm links, brackets, a bell crank and a hook, each optimized and then checked by a fresh analysis of the delivered mesh. "Under envelope" compares with the design space filled solid, not with a conventionally designed part. These are simulated results; see [docs/VALIDATION.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/VALIDATION.md). Blue arrows are loads, red is keep-out. Every view, stress map and number is in [docs/GALLERY.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/GALLERY.md).
+
+<table>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/arm-link.png" width="250"><br><b>arm-link</b><br><sub>55.3 g · 56% under envelope · SF 8.92 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/arm-link-cast.png" width="250"><br><b>arm-link-cast</b><br><sub>67.8 g · 46% under envelope · SF 20.11 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/arm-link-driven.png" width="250"><br><b>arm-link-driven</b><br><sub>48.4 g · 61% under envelope · SF 11.17 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/bell-crank.png" width="250"><br><b>bell-crank</b><br><sub>23.4 g · 84% under envelope · SF 3.22 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-turret.png" width="250"><br><b>grip-turret</b><br><sub>140.4 g · 91% under envelope · SF 2.85 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-upper-arm.png" width="250"><br><b>grip-upper-arm</b><br><sub>207.7 g · 88% under envelope · SF 7.33 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-wrist-link-mjf.png" width="250"><br><b>grip-wrist-link-mjf</b><br><sub>40.2 g · 95% under envelope · SF 5.83 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-wrist-yoke.png" width="250"><br><b>grip-wrist-yoke</b><br><sub>17.3 g · 93% under envelope · SF 3.88 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/mbb-beam.png" width="250"><br><b>mbb-beam</b><br><sub>13.8 g · 79% under envelope · SF 1.51 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/motor-mount.png" width="250"><br><b>motor-mount</b><br><sub>27.6 g · 69% under envelope · SF 3.65 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/printed-hook.png" width="250"><br><b>printed-hook</b><br><sub>24.2 g · 84% under envelope · SF 2.73 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/wall-bracket.png" width="250"><br><b>wall-bracket</b><br><sub>26.9 g · 94% under envelope · SF 2.85 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/wall-bracket-fasteners.png" width="250"><br><b>wall-bracket-fasteners</b><br><sub>26.5 g · 94% under envelope · SF 2.57 · pass</sub></td>
+</tr>
+</table>
+
+Where the load goes: von Mises stress, blue (none) to red (the part's 98th percentile), on the part bent by its worst load case:
+
+<table>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/bell-crank.png" width="250"><br><sub>bell-crank</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/grip-upper-arm.png" width="250"><br><sub>grip-upper-arm</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/printed-hook.png" width="250"><br><sub>printed-hook</sub></td>
+</tr>
+</table>
+
+Two-fluid heat exchangers: two isolated channel networks through one sheet gyroid, with headers that print without support.
+
+![box-crossflow](https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/heatx/box-crossflow.png)
+
+Printed in metal, nothing inside can be supported: the channels end in peaked roofs instead of flat ceilings (red).
+
+![metal ceilings](https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/heatx/lpbf-ceilings.png)
+
+Made another way: the same arm link printed, cast (no undercuts along z), 3-axis milled from both sides with a 6 mm cutter, and with members capped at 6 mm. Each meets SF 3 and 0.5 mm.
+
+![manufacturing constraints](https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/manufacturing/arm-link.png)
+
+<sub>Rendered 2026-11-03 from `6c97b57` by `python docs/gallery.py`.</sub>
+
+<!-- gallery:end -->
+## Limits
+
+- Linear elastic, small deflection in the main check; large deflection and
+  buckling are separate analyses (`nonlinear.py`), with contact (rigid
+  obstacles and self contact, optionally with Coulomb friction) and
+  tendons there only.
+- Printed layers are modelled as transversely isotropic stiffness plus a
+  tension-across-layers strength check. Interlayer shear is not checked
+  separately.
+- Material cards are datasheet values until you calibrate your printer
+  (`loadpath calibrate`).
+- Results depend on the grid: mass at a target safety factor moves about
+  5% with small changes of grid size or starting point.
+- Pure Python with NumPy and SciPy. A 40,000-cell study takes a few minutes
+  on one laptop core; a safety-factor search runs several of those, and the
+  larger eval parts take 15-60 minutes on four cores.
+
+## Tests
+
+```bash
+python tests/run_tests.py && python tests/test_loadpath.py        # core
+python tests/test_tutorials.py                                    # every tutorial, quick mode
+python -m loadpath.evals.run --tier coarse --corpus 6 \
+    --baseline loadpath/evals/baselines/coarse.json --gate regressions   # under 30 min
+python -m loadpath.evals.run --out evals-out/                     # full eval suite, hours
+python -m loadpath.benchmarks.run --out bench-out/                # top88 / top3d benchmarks
+python docs/gallery.py --evals evals-out/                         # refresh the Examples renders
+```
+
+`.github/workflows/tests.yml` lists the full set CI runs.
+
+## License
+
+MIT

loadpath-0.1.1/README.md

@@ -0,0 +1,258 @@
+# loadpath
+
+Generative design for 3D printed parts (and cast or milled ones), built to be
+driven by coding agents. Tell your agent what a part has to carry and where it
+mounts; loadpath grows the part with topology optimization, meshes it as a
+watertight solid with the interfaces cut exactly, and checks that mesh with a
+fresh finite element analysis. You get an STL, 3MF or STEP and a report: mass,
+safety factor, deflection, thinnest member, printability, pass or fail.
+
+<table>
+<tr>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/printed-hook.png" width="380"></td>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/printed-hook.png" width="380"></td>
+</tr>
+<tr>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/bell-crank.png" width="380"></td>
+<td align="center" width="50%"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/bell-crank.png" width="380"></td>
+</tr>
+<tr>
+<td align="center" colspan="2"><sub>Left, the part loadpath delivered (blue: load, red: kept bore). Right, its
+stress from a fresh analysis of that mesh, blue (none) to red (the part's 98th percentile), on the shape bent by
+the worst load case. Top: a 300 N wall hook printed upright without supports, 24.2 g, SF 2.73. Bottom: a 500 N
+aluminium bell crank (LPBF), 23.4 g, SF 3.22. More in <a href="#examples">Examples</a>.</sub></td>
+</tr>
+</table>
+
+## Point your agent at it
+
+```bash
+pip install "loadpath[mcp]"
+claude mcp add loadpath -- loadpath mcp   # any MCP client: {"command": "loadpath", "args": ["mcp"]}
+```
+
+Then ask in plain words:
+
+> Design a PETG bracket for a 10 kg shelf. It screws to the wall with two
+> M5 bolts 40 mm apart and the shelf sits 60 mm out. Safety factor 2,
+> printed on an FDM printer.
+
+The agent writes the study, checks it, runs it in the background and reads
+the report back. Its guide is [loadpath/SKILL.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/loadpath/SKILL.md) (the
+`loadpath_guide` tool returns it); the tools and the loop are in
+[docs/agents.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/agents.md). The server runs locally and nothing
+leaves the machine. Units are millimetres, newtons and megapascals. MIT
+licensed.
+
+## What it does
+
+- **Topology optimization** on a voxel grid (SIMP with a density filter,
+  Heaviside projection and a robust minimum-member formulation), searching
+  for the lightest part that meets a safety factor and a deflection limit
+  across several load cases. A stress-constrained mode (MMA) is available.
+- **Verification of the delivered mesh**, not the optimizer's density
+  field: the STL is analysed again, on a tetrahedral mesh when the
+  `crosscheck` extra is installed and on a fine voxel grid otherwise. Keep-out
+  intrusion and lost interface geometry are measured with exact booleans.
+- **Process rules** for FDM (overhangs, bridges, layer bond), SLS and MJF
+  (trapped powder, escape holes, warp-prone panels), resin (supports, peel,
+  suction cups) and powder-bed metal (internal ceilings, powder
+  clean-outs), plus milling and casting constraints.
+- **Field modelling**: a part as a graph of implicit fields, so skins,
+  lattices (TPMS and strut) and fillets can be graded by stress or by
+  distance ([docs/fields.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/fields.md)).
+- **Other analyses**: large deflection and buckling, compliant mechanisms,
+  thermal loads, fatigue and creep estimates, lattice crush and impact,
+  vibration, lattice-Boltzmann flow and conjugate heat for heat exchangers
+  and cold plates, and layer-by-layer warp and residual stress
+  ([docs/BUILDSIM.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/BUILDSIM.md)).
+- **CAD in and out**: a STEP or STL part becomes a draft study with its
+  bores and bolt patterns found ([docs/IMPORT.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/IMPORT.md)); results
+  export as STL, 3MF, and STEP with exact planes and cylinders at the
+  interfaces; a lattice part (strut cells, TPMS, graded, skinned or open)
+  goes to STEP lattice and all, as smooth patches on its true surface ([docs/fields.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/fields.md#the-lattice-itself-in-the-step-lattice-exact)).
+- **Mesh checks without a study** (`loadpath.checks`): wall thickness,
+  overhangs, trapped voids, fluid-network isolation, component fit.
+
+## How much to trust it
+
+Read [docs/VALIDATION.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/VALIDATION.md) before relying on a number.
+In short:
+
+- **The solvers match analytic answers** in the test suite: beam
+  deflection within 3-5%, stress at a hole within about 13%, Lamé pressure
+  vessels within 5-11%, Poiseuille flow exactly.
+- **The structural check is cross-checked**: CalculiX and loadpath's own
+  tetrahedral solver agree within 0.1% on the same mesh. The voxel check
+  reads 24-49% higher stress than the tet check, so it errs safe.
+- **The optimizer is benchmarked**: 1-5% less stiff than top88 and top3d at
+  equal material on three of their textbook problems, stiffer on the fourth.
+- **Printed lattices match published tests**: against 33 published
+  compression tests of printed lattices, the stiffness check has a median
+  error of 1.23x (held out) and the strength check stays under the measured
+  value on all 33.
+- **Warp prediction is fitted to measured data** (NIST AM-Bench 2018 for
+  LPBF IN625, a published ABS study for FDM) and misses other alloys and
+  materials by 30-60%.
+- **Everything else is simulation only**, including every part mass and
+  safety factor in the case studies. The analysis is linear elastic with
+  datasheet material values; printed parts vary with the printer.
+
+## Case studies
+
+Each states the problem, the setup, the numbers against a named baseline,
+and the limitations ([docs/case-studies/](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/docs/case-studies/)).
+
+| Case study | Result | Baseline |
+| --- | --- | --- |
+| [Eval suite](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/evals.md) | 15 of 15 parts meet their requirements on a fresh analysis of the delivered mesh | Each part's own requirements |
+| [GRIP robot arm](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/grip.md) | A load sweep over the arm's URDF found 3 of 5 links below SF 1; the regrown set meets SF 2 and is 18% smaller | GRIP's committed parts, made by an earlier loadpath with hand-written loads |
+| [Arm link cores](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/arm-link-cores.md) | 42.2 g box section for FDM, 36.8 g hollow link for SLS, both at 0.5 mm and SF 3 | 54.9 g solid web; 54.2 g free-form topology optimization |
+| [Liquid cold plate](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/cold-plate.md) | A die-graded gyroid ties streamline channels; channels filled with a graded gyroid run 7% cooler | Streamline channels and pin fins, equal metal and pressure drop |
+| [TPU finger](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/tpu-finger.md) | up to 8 mm of tip curl for a 5 mm pull; a tendon finger grips 8x harder | A linear-model design (2.5 mm, then the solve fails); a plain bar of equal volume |
+| [Benchmarks](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/case-studies/benchmarks.md) | 1-5% less stiff than top88/top3d at equal material, in 57-60 iterations instead of 130-410 | Ports of top88 and top3d |
+
+## Command line and Python
+
+```bash
+loadpath examples shelf-bracket > bracket.json
+loadpath check bracket.json            # every problem at once, with suggestions
+loadpath run bracket.json --out out/   # optimize, mesh, verify, render
+```
+
+`out/` gets an STL and a render per outcome and `report.json`. Every command
+takes `--quick` for a first look in a minute or two.
+
+Optional extras: `crosscheck` (tetrahedral verification; recommended),
+`render`, `step` (STEP in and out, OpenCascade), `flow` (lattice-Boltzmann,
+numba), `gpu` (PyTorch for large grids on CUDA, Apple Silicon or CPU
+threads).
+
+Studies are validated against a published JSON Schema
+([`schema/`](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/schema/)), and `loadpath check` returns every problem at once
+with its JSON path, so an agent can fix a study in one round trip. Long
+solves run as background jobs the agent polls. Each finished outcome has an
+`attention` list of numbers to look at before trusting it, and renders a
+vision model can inspect.
+
+From Python:
+
+```python
+from loadpath.spec import study_from_dict
+from loadpath.study import run_study
+
+result = run_study(study_from_dict(study))
+best = result.outcomes[result.best]
+best.mesh.export("part.stl")
+print(best.report["mass"], best.report["safetyFactor"])
+```
+
+## Documentation
+
+- [Quickstart](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/quickstart.md): install to a checked part.
+- [Tutorials](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/docs/tutorials/): eleven end-to-end workflows, each run by CI
+  (an FDM bracket to 3MF, an SLS lattice with drain holes, a metal heat
+  exchanger, an impact bumper, a STEP part round trip, LPBF distortion
+  compensation, and more).
+- [Reference](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/loadpath/README.md): the method, every study field, the
+  report, the limits.
+- [Validation](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/VALIDATION.md): what each model has been checked
+  against.
+- [Case studies](https://github.com/connorwhite-online/loadpath/tree/v0.1.1/docs/case-studies/) and [benchmarks](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/BENCHMARKS.md).
+- Topics: [fields](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/fields.md), [materials](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/materials.md),
+  [build simulation](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/BUILDSIM.md), [calibration](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/CALIBRATION.md),
+  [CAD import](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/IMPORT.md), [slicer](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/slicer.md),
+  [multi-material](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/MULTIMATERIAL.md), [impact](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/IMPACT.md),
+  [mechanisms](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/MECHANISMS.md).
+
+<!-- gallery:start -->
+
+## Examples
+
+The eval suite (15/15 passing): robot-arm links, brackets, a bell crank and a hook, each optimized and then checked by a fresh analysis of the delivered mesh. "Under envelope" compares with the design space filled solid, not with a conventionally designed part. These are simulated results; see [docs/VALIDATION.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/VALIDATION.md). Blue arrows are loads, red is keep-out. Every view, stress map and number is in [docs/GALLERY.md](https://github.com/connorwhite-online/loadpath/blob/v0.1.1/docs/GALLERY.md).
+
+<table>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/arm-link.png" width="250"><br><b>arm-link</b><br><sub>55.3 g · 56% under envelope · SF 8.92 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/arm-link-cast.png" width="250"><br><b>arm-link-cast</b><br><sub>67.8 g · 46% under envelope · SF 20.11 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/arm-link-driven.png" width="250"><br><b>arm-link-driven</b><br><sub>48.4 g · 61% under envelope · SF 11.17 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/bell-crank.png" width="250"><br><b>bell-crank</b><br><sub>23.4 g · 84% under envelope · SF 3.22 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-turret.png" width="250"><br><b>grip-turret</b><br><sub>140.4 g · 91% under envelope · SF 2.85 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-upper-arm.png" width="250"><br><b>grip-upper-arm</b><br><sub>207.7 g · 88% under envelope · SF 7.33 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-wrist-link-mjf.png" width="250"><br><b>grip-wrist-link-mjf</b><br><sub>40.2 g · 95% under envelope · SF 5.83 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/grip-wrist-yoke.png" width="250"><br><b>grip-wrist-yoke</b><br><sub>17.3 g · 93% under envelope · SF 3.88 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/mbb-beam.png" width="250"><br><b>mbb-beam</b><br><sub>13.8 g · 79% under envelope · SF 1.51 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/motor-mount.png" width="250"><br><b>motor-mount</b><br><sub>27.6 g · 69% under envelope · SF 3.65 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/printed-hook.png" width="250"><br><b>printed-hook</b><br><sub>24.2 g · 84% under envelope · SF 2.73 · pass</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/wall-bracket.png" width="250"><br><b>wall-bracket</b><br><sub>26.9 g · 94% under envelope · SF 2.85 · pass</sub></td>
+</tr>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/cases/wall-bracket-fasteners.png" width="250"><br><b>wall-bracket-fasteners</b><br><sub>26.5 g · 94% under envelope · SF 2.57 · pass</sub></td>
+</tr>
+</table>
+
+Where the load goes: von Mises stress, blue (none) to red (the part's 98th percentile), on the part bent by its worst load case:
+
+<table>
+<tr>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/bell-crank.png" width="250"><br><sub>bell-crank</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/grip-upper-arm.png" width="250"><br><sub>grip-upper-arm</sub></td>
+<td align="center" valign="top"><img src="https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/stress/printed-hook.png" width="250"><br><sub>printed-hook</sub></td>
+</tr>
+</table>
+
+Two-fluid heat exchangers: two isolated channel networks through one sheet gyroid, with headers that print without support.
+
+![box-crossflow](https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/heatx/box-crossflow.png)
+
+Printed in metal, nothing inside can be supported: the channels end in peaked roofs instead of flat ceilings (red).
+
+![metal ceilings](https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/heatx/lpbf-ceilings.png)
+
+Made another way: the same arm link printed, cast (no undercuts along z), 3-axis milled from both sides with a 6 mm cutter, and with members capped at 6 mm. Each meets SF 3 and 0.5 mm.
+
+![manufacturing constraints](https://raw.githubusercontent.com/connorwhite-online/loadpath/v0.1.1/docs/images/manufacturing/arm-link.png)
+
+<sub>Rendered 2026-11-03 from `6c97b57` by `python docs/gallery.py`.</sub>
+
+<!-- gallery:end -->
+## Limits
+
+- Linear elastic, small deflection in the main check; large deflection and
+  buckling are separate analyses (`nonlinear.py`), with contact (rigid
+  obstacles and self contact, optionally with Coulomb friction) and
+  tendons there only.
+- Printed layers are modelled as transversely isotropic stiffness plus a
+  tension-across-layers strength check. Interlayer shear is not checked
+  separately.
+- Material cards are datasheet values until you calibrate your printer
+  (`loadpath calibrate`).
+- Results depend on the grid: mass at a target safety factor moves about
+  5% with small changes of grid size or starting point.
+- Pure Python with NumPy and SciPy. A 40,000-cell study takes a few minutes
+  on one laptop core; a safety-factor search runs several of those, and the
+  larger eval parts take 15-60 minutes on four cores.
+
+## Tests
+
+```bash
+python tests/run_tests.py && python tests/test_loadpath.py        # core
+python tests/test_tutorials.py                                    # every tutorial, quick mode
+python -m loadpath.evals.run --tier coarse --corpus 6 \
+    --baseline loadpath/evals/baselines/coarse.json --gate regressions   # under 30 min
+python -m loadpath.evals.run --out evals-out/                     # full eval suite, hours
+python -m loadpath.benchmarks.run --out bench-out/                # top88 / top3d benchmarks
+python docs/gallery.py --evals evals-out/                         # refresh the Examples renders
+```
+
+`.github/workflows/tests.yml` lists the full set CI runs.
+
+## License
+
+MIT