Primary case study / Computational design

Constraint-Aware Massing Explorer

A transparent geometric search tool for early option comparison.

Supplied site and programme parameters generate seeded rectangular massing options. Hard feasibility checks are applied before transparent proxy objectives and Pareto ranking.

Open source Read evaluation

Actual local interface / synthetic site inputs

Problem

Early design options must satisfy supplied constraints before environmental, access, area, and form trade-offs are compared.

System

Typed site schema, seeded generators, hard feasibility checks, editable proxy objectives, Pareto ranking, and SVG outputs.

Evidence

864 candidates per method across authored synthetic sites, compared with an unconstrained random baseline.

Boundary

Rectangular schematic envelopes only. No code inference, developed plan, calibrated simulation, structure, or approvable design.

System

Hard checks precede ranking

The tool does not ask a score to compensate for a failed setback or height constraint. Infeasible options are rejected first; proxy objectives only compare the remaining schematic envelopes.

Massing explorer system map showing site inputs, seeded generators, hard constraints, proxy objectives, Pareto ranking, and generated geometry — Generated from checked-in evaluator outputs and typed project metadata. Environmental and access values are labeled proxies.

Measured behavior

Constraint-aware search sharply reduces invalid options

The baseline establishes whether the generator is doing more than sampling rectangles. Results describe the authored synthetic scenarios and simplified geometry, not design quality.

Feasible rate0.977Constraint-aware candidate generation

Baseline feasible rate0.052Unconstrained random option generator

Best GFA error0.19%Mean across authored synthetic sites

Baseline GFA error34.47%Mean best error from the random baseline

Comparison of ranked massing options with plan geometry and proxy measurements — Generated plans and scores expose the trade-offs that produce the ranking.

Local massing explorer interface with plan, isometric view, constraints, and proxy results — Actual local interface using bundled synthetic site parameters.

Engineering decisions

Geometry and assumptions stay inspectable

The project favors explicit schemas and simple geometry over an opaque claim of AI-generated architecture.

Typed project parameters

Site dimensions, setbacks, height, target GFA, access edge, and proxy weights enter through a validated schema rather than free-form prompting.

Seeded option generation

Every option set can be regenerated from its method, parameters, and seed. This makes ranking changes and test failures reproducible.

Hard feasibility gate

Boundary, setback, floor-count, and supplied height rules reject options before multi-objective comparison.

Transparent proxy objectives

Solar exposure, wind openness, access distance, and compactness remain visible heuristics. They are not presented as calibrated environmental analysis.

Retained limitation

A feasible envelope is not a resolved building.

The current schema cannot reason about internal egress, accessibility routes, structure, cores, servicing, or detailed code clauses. These omissions remain outside the score and the integration contract.

Inspect failure analysis

Run locally

Regenerate options and evaluation

The evaluator regenerates the authored synthetic scenarios, compares the constrained method with the baseline, and rewrites the visual evidence deterministically.

python projects/constraint-aware-massing-explorer/evaluate_massing.py python -m pytest tests/test_massing_explorer.py streamlit run projects/constraint-aware-massing-explorer/app.py