ATLASPEC 0.2 · RESEARCH CANDIDATE

Cartographic intent, compiled with guardrails.

Agents describe what a map must communicate. Atlaspec validates the semantic contract, rejects unsafe choices, and deterministically emits renderer artifacts.

Inspect the compiler Read the limits

No API key · generated from the repository compiler · no LLM runs in this page

SEMANTIC AUTHORING SURFACE 0.2.0-rc.1

Intent document semantic_type: probability missing_data: explicit

Enforceable policy schema · semantic lint · capability check

Renderer artifact MapLibre · portable Vega-Lite subset

DETERMINISTIC

STATUS Research candidate

LOCAL HOLDOUT 12 unique tasks

V0.2 HOLDOUT Not run

PRODUCTION CLAIM None

01 · COMPILER-GENERATED EXAMPLES

Inspect the part a schema alone cannot guarantee.

These examples are generated from the checked-in Atlaspec compiler. Select a case to inspect its semantic diagnostics, decisions, and renderer artifact.

SEMANTIC LINT

Loading compiler examples…

LOADING

Diagnostics —

Compiler decisions —

02 · EARLY EVIDENCE

Evidence, sized to what was actually tested.

The current results motivate further research. They do not establish universal model generalization, human map-reading benefit, or production readiness.

ONE-TIME LOCAL HOLDOUT · 12 TASKS × 5 REPETITIONS × 2 AGENTS

Direct MapLibre 108 / 120 90% accepted

Atlaspec 120 / 120 100% accepted

+10 pp 95% CI: +3.3 to +18.3 pp

Output tokens per accepted map were 77.3% lower for the tested Codex CLI and 59.4% lower for the tested Claude Code version. Provider token accounting is not comparable across agents.

DEVELOPMENT DIAGNOSTIC

68/72 vs 40/72

Locked v0.2 renderer-health checks favored Atlaspec by 38.89 percentage points. The tasks and two local agents are correlated, and project authors designed the gates.

R&D PILOT · 6 DEV TASKS

Symmetric repair baseline

Direct MapLibre recovered to 6/6 for both agents after one equal repair opportunity. Atlaspec also reached 6/6 with fewer output tokens. The pilot is too small for a release claim.

NOT YET MEASURED

Independent review

No external task designer, blind cartographer panel, or human map-reading experiment has confirmed the result.

Read the local holdout ↗ Read the development diagnostic ↗

03 · PRODUCT BOUNDARY

An opinionated cartographic IR—not a general GIS.

Deterministic compilation trades arbitrary renderer freedom for a smaller, auditable, safety-oriented authoring surface.

ATLASPEC 0.2 DOES

Represent explicit map-reading intent and field semantics
Reject unsupported or misleading cartographic mappings
Compile deterministic MapLibre artifacts
Compile a fail-closed portable Vega-Lite subset
Record diagnostics and inferred decisions

ATLASPEC 0.2 DOES NOT

Perform routing, visibility, shadows, or geoprocessing
Replace GIS, GeoJSON, MapLibre, or Vega-Lite
Support arbitrary renderer-native customization
Guarantee map-reading effectiveness or production safety
Turn the current development result into a sealed holdout

OPEN RESEARCH · REPRODUCIBLE BY DEFAULT

Find the first semantic failure we missed.

Try a map request, inspect the compiler, and report the smallest cartographic counterexample you can produce.

Explore the repository ↗