Domain-agnostic high-dimensional whitespace analyzer, finder, visualizer, and explorer.
Built in Rust on rung (ladder!, rung_std::decide, rung_std::llm, rung-agent-core) and wgpu/WGSL shaders.
Live Interactive Demo (GitHub Pages): https://witt3rd.github.io/lacuna/
README.mdis a symlink to this file. Edit onlyAGENTS.md. Custody isAGENTS.md+skills/; no.agent/folder, no inhabit, no formal handoff.
If you embed 100,000 domain entities (Steam games, startups, patents, molecules) into a single vector space and naively pick a point "far away from neighbors," three failure modes immediately occur:
-
Off-Manifold Boundary Drift: In high dimensions (
$D \ge 64$ ), unconstrained distance maximization pushes points onto the exterior boundary of the hypersphere—orthogonal to the entire domain manifold. - Surface vs. Structural Entanglement: Standard text embeddings conflate surface theme/setting (e.g., cyberpunk, medieval, zombies) with structural mechanics (e.g., core gameplay loop, progression, social topology).
- Dead Voids vs. Gold Voids: Many regions of concept space are empty because they are physically impossible, un-fun, or commercially dead.
lacuna solves all three:
-
Four-Layer Universal Input Contract: Decomposes every entity into orthogonal semantic facets (
FacetSpec), unit-normalized per-facet embeddings, and explicit demand + quality signals. -
Interior Delaunay Empty-Sphere Solver: Restricts search to the geodesic interior simplex of diverse, high-demand anchor clusters (
optimize_interior_simplex_center), maximizing the empty-sphere radius $R(x^) = \min_{e \in \text{Corpus}} d(x^, e)$ while guaranteeing$x^*$ remains surrounded by valid domain clusters. -
rung::ladder!Closed-Loop Synthesis &rung_std::decideViability Gate: Abducts concrete opportunity hypotheses (DraftHypothesis), scores them on atomic probability questions viarung_std::decide::Decider(JevDecider/Recorded/CalibratedGeometricDecider), and re-embeds the generated description to prove mathematically that its vector lands inside the void and stays separated from all existing corpus entities (VerifiedOpportunity). -
Continuous Topographic Potential Field (
wgpuWGSL + SVG/WebGPU HTML): Evaluates incumbent saturation density against annular demand gravity to illuminate glowing cyan whitespace basins and bounded Voronoi cells.
- The compiler gates
VerifiedOpportunity.lacuna_rung::synthesize_voidis arung::ladder!. You cannot construct aVerifiedOpportunitywithout passing throughAbducting, therung_std::decide::Deciderviability check, and the closed-loop re-embedding distance proof. - Interior geodesic containment. Candidate void centers are strictly positive convex combinations of anchor facet vectors on the unit hypersphere (
geodesic_simplex_point), preventing off-manifold drift. - Agent-first CLI (
lacuna). Every operation—schema inspection, corpus ingestion, high-D void solving, single-void probing,rungladder synthesis, and topographic field rendering—is exposed as a deterministic CLI subcommand with--jsonoutput so an autonomous agent (rung-agent) can pilot the entire workflow. - Zero-key offline determinism + live
rungscaling. Deterministic locality-sensitive n-gram embeddings (deterministic_embed) andCalibratedGeometricDeciderlet every test and CLI command run with zero network or credentials; passing--live-llmand--jevdispatches throughrung_std::llm::llm_callandrung_std::decide::JevDecider(OpenRouter / System One).
crates/lacuna-core Pure-Rust (wasm32-compatible) schema, multi-facet metric space,
interior Delaunay empty-sphere solver, 2D Sutherland-Hodgman
bounded Voronoi tessellation, potential field grid, and Steam benchmark
crates/lacuna-rung rung::ladder! (SynthesizeVoid), rung_std::decide viability judge,
rung_std::llm abduction, and rung-agent-core turn engine bridge
crates/lacuna-gpu WGSL topographic potential-field shader (field.wgsl), SVG vector
manifold renderer, and self-contained WebGPU/Shader HTML explorer
crates/lacuna-cli The `lacuna` binary: 100% agent-pilotable CLI (--json everywhere)
The crate rustdoc compiles and runs README.md (AGENTS.md). Unlabeled fences are treated as Rust—label non-Rust blocks.
use lacuna_core::{VoidSolverConfig, find_interior_voids, min_distance_to_corpus, steam_seed_snapshot};
fn main() {
let snapshot = steam_seed_snapshot();
let voids = find_interior_voids(&snapshot.schema, &snapshot.entities, &VoidSolverConfig::default());
assert!(!voids.is_empty());
let top = &voids[0];
let weights = snapshot.schema.default_weights();
let min_d = min_distance_to_corpus(&top.center, &snapshot.entities, &weights);
assert!((min_d - top.radius).abs() < 1e-3);
}# 1. Inspect the 5-axis domain schema
cargo run -q -p lacuna-cli -- schema
# 2. Solve & rank top interior Delaunay empty spheres (human table or --json)
cargo run -q -p lacuna-cli -- voids --top-k 5
cargo run -q -p lacuna-cli -- voids --top-k 5 --json
# 3. Probe a specific void's boundary neighbors, Voronoi polygon & facet bridge
cargo run -q -p lacuna-cli -- probe void-01 --json
# 4. Run the rung::ladder! synthesis + rung_std::decide judge + closed-loop proof
cargo run -q -p lacuna-cli -- synthesize void-01 --json
# 5. Run the flagship @SyntheticBeef Steam Games Whitespace Explorer & emit SVG + WebGPU HTML into out/
cargo run -q -p lacuna-cli -- demo steam --count 60 --top-k 5 --svg out/steam_manifold.svg --html out/steam_explorer.htmlcargo fmt --check
cargo clippy --workspace --all-targets --locked -- -D warnings
cargo test --workspace --locked
RUSTDOCFLAGS="-D warnings" cargo doc --workspace --no-deps --locked| Skill | One Line |
|---|---|
lacuna |
Repo caretaker: house git (master iteration, rebase), CI gates, rung tag pinning, release procedure |
pilot-cli |
How an agent pilots lacuna via CLI (ingest → voids → probe → synthesize → field) with --json |
domain-onboarding |
Onboarding any domain (VC, Patents, Pharma, SaaS, Research) onto the 4-layer input contract |
steam-benchmark |
Playbook for the 100,000+ Steam Games Whitespace Explorer (@SyntheticBeef challenge) |