MTE — Meta-Template Engine

Status: in development, no public artifact yet. The source repo is private with no public demo or package (not on PyPI; no releases tagged). MTE describes itself as a research artifact with a working measurement harness, not a shipping product — its core question (does a structured pipeline beat one-shot LLM adaptation?) is explicitly unproven, though now measurable.

MTE adapts a reusable Python template to a target domain as a bounded, reviewable process — a monotonic loop reviewed at three scales (architecture / module / line), stressed by adversarial tests in a sandbox, and halted by a circuit breaker when it stops improving — instead of a single LLM prompt you can’t inspect.

What it does

• Adapts a template to a domain via a nested optimize/review loop that is monotonic in a join-semilattice (state can only improve).
• Reviews at three scales — architecture (L0), module (L1), and line (L2) — with per-level step budgets and gating thresholds.
• Stress-tests in a sandbox with adversarial, generated pytest cases (Docker with a subprocess fallback; a Go sandbox executor).
• Measures itself — mte compare runs a one-shot baseline and the full pipeline over the same cases and reports the delta, so the benefit is testable rather than asserted.

Who it’s for

1. Researchers of structured, lattice-based LLM orchestration, fractal review, and quiescence detection — the subsystems are real, tested, and composable.
2. Anyone testing the premise — use mte compare / mte eval to check whether pipeline-style adaptation beats one-shot on their own cases.
3. As a teaching codebase — eight cleanly separated subsystems, 380+ tests, property-based CRDT laws, a sandbox with a real threat model.

For a one-off code adaptation, the README recommends asking an LLM directly — that’s the baseline MTE measures itself against. MTE earns its overhead only with many adaptations of the same scaffold, a need for an audit trail, or interest in the orchestration itself.

Current State

Seven subsystems (lattice, templates, optimization, adversarial, review, quiescence, orchestrator) are wired end-to-end through MetaTemplateEngine.adapt — an eighth, training, ships as a separate data-export command — and reachable via four interfaces: CLI, REST API, an MCP server, and a web dashboard, over one service core. The suite is 380+ tests green on Python 3.12/3.13 at ~88% line coverage (CI gate ≥80%), ruff-clean, with the Go sandbox building and wheels passing twine check. The local Lefthook gate passes; hosted GitHub Actions is currently billing-paused.

What’s proven: the subsystems are correct and composable, the sandbox is hardened, and the pipeline runs to completion producing valid Python. What’s not proven yet: that the pipeline beats one-shot adaptation — the instrument to test it (mte compare) is built and plumbing-tested, but a live recorded verdict is the next milestone.

Source is private; architectural detail, a technical deep-dive, and the evaluation methodology are available on request via the contact page.

What’s Needed For This Entry To Tighten

• A public source repository linked via githubUrl, and/or
• A public demo URL or downloadable artifact that exercises the template-adaptation flow.