The pattern catalog and edge-first AI cascade runtime.

Pattern-lang composes two layers:

1. A lawful synthesizer — 690 named primitives, 41 target languages, deterministic codegen. Intent maps to a typed DAG of patterns; the DAG lowers to Cortex IR; the IR emits in any target. Same input, same output. Picojoule-budgeted by construction.
2. A verifier-first cascade runtime — L0 cache → L1 lawful → L2 small models (Prism MLGRU, Liquid CfC, MRL, DeBERTa) → L3 medium models (Gemma 4, LMM) → L4 frontier. Each tier wrapped in an independent verifier (recompute, EBM constraint, Lean proof, k-of-n vote). Accept on agreement, escalate on reject.

Capability per joule, not capability per parameter.

This is the public release surface. Source is private under openIE-dev/pattern-lang-core. Use the binaries.

This repository ships release binaries and documentation, not source. Pattern-lang is free to use software, not an open-source project. See LICENSE for the Business Source License 1.1 terms — converts to Apache-2.0 four years after each binary's release date.

Releases are attached to GitHub Releases. The marketing surface is at pattern-lang.ai.

• pattern-lang — CLI for synthesis, composition, and the cascade
• joule-edge-server — Anthropic-compatible HTTP server (POST /v1/messages) with joules_spent and tier_used in usage
• joule-edge-cli — Plan + Execute + Diagnose + Compose wired into a single binary

Target platforms:

Once a release is published:

# macOS / Linux — grab the latest release for your platform
curl -fsSL https://github.com/openIE-dev/pattern-lang/releases/latest/download/install.sh | sh
# Verify
pattern-lang --version

For air-gapped installs, download the binary tarball from the Releases page and place it on $PATH.

# Engage the catalog interactively
pattern-lang engage
# Compose a typed DAG for an intent
pattern-lang compose "sort a list and remove duplicates"
# Emit in your target language
pattern-lang emit --target rust
pattern-lang emit --target python
pattern-lang emit --target go

# Start the local edge server
joule-edge-server --bind 127.0.0.1:7777
# Hit it with any Anthropic-format client
curl http://127.0.0.1:7777/v1/messages \
 -H "Content-Type: application/json" \
 -d '{
 "model": "pattern-lang",
 "messages": [{"role": "user", "content": "gcd of 12 and 8"}]
 }'
# Response includes usage.joules_spent and usage.tier_used

For Rust users, the public crate ships as binary releases only; embedded library integration is in openIE-dev/pattern-lang-core (private; contact for access).

query → L0 cache (~nJ) miss → continue
 → L1 lawful synth (~μJ) accept if catalog matches
 ↓ verify (recompute / Lean / EBM)
 → L2 Prism / Liquid (~μJ–mJ)
 ↓ verify (k-of-n)
 → L3 Gemma 4 / LMM (~mJ)
 ↓ verify
 → L4 frontier (~10s mJ) RPC, deferred

Each tier reports its joule estimate. The router walks tiers in cost order; the verifier gates each step. On reject, escalate. On accept, return — with a receipt.

Live verification of the cost wedge:

# After install:
pattern-lang bench cascade
# Shows: verify_first is ×ばつ cheaper than always-frontier
# at the same 100% correctness on the in-tree benchmark.

• docs/getting-started.md — install, first query, configuration
• docs/cli.md — CLI reference
• docs/server.md — joule-edge-server and the /v1/messages API
• docs/cascade.md — tier model, routing policies, verifier types
• docs/license-faq.md — BSL 1.1 in plain English

The deeper architecture story lives at synthesis.openie.dev — the strategic doctrine that locates pattern-lang in the seven-axis map of information synthesis.

Business Source License 1.1. Licensed Work is the binaries attached to releases of this repository. Free for non-commercial use, internal use by organisations under 1ドルM ARR, and security / academic research. Change Date is four years after each binary's release date; on that date the binary converts to Apache-2.0.

Issues and questions: open an issue here. The security policy covers responsible disclosure.

For commercial-licensing inquiries (outside the BSL Additional Use Grant), see pattern-lang.ai or reach OpenIE via openie.dev.

Not affiliated with, or endorsed by, any AI vendor named in the documentation.