An independent compatibility kit for Liquid LFMs.
Droplet checks whether an LFM, its chat template, its runtime, and its client fit together. It can diagnose a route, apply one narrow history repair on a loopback bridge, and run a reproducible compatibility suite.
Not affiliated with, endorsed by, or connected to Liquid AI.
From the repository root on Linux or macOS:
python3 -m venv .venv && .venv/bin/python -m pip install ./cli && .venv/bin/droplet versionThe package installs from this repository. No PyPI package is assumed. See the
CLI guide for platform details and the three working
commands: doctor, bridge, and verify.
- The same model can look broken when its chat template hides its past tool calls.
- On seeded 1.2B chains, direct scored ordered 0/6 and exact 0/6.
- With history restored, the same chains scored ordered 6/6 and exact 0/6 because the model added an unrequested call.
- In live OpenCode runs, direct scored ordered 0/3 and exact 0/3; bridged scored ordered 3/3 and exact 2/3.
- On 24B six-step chains, direct scored ordered 0/3 and exact 0/3; repaired scored ordered 3/3 and exact 3/3.
- The 8B route needed no repair: 18/18 task and protocol probes were unchanged, with byte parity proven separately by golden hash tests.
These are small seeded A/B reproductions, not a large-n leaderboard. Droplet restores missing history. It does not make a model smarter or more obedient, and its strict suite still fails when the model adds an extra call.
- CLI, installation, support, and command reference
- Evidence, per-run ledgers, and reproduction guide
- Recorded direct and bridged demo
- Upstream source pack and contribution index
- How to contribute
- v0.1.0 release notes
Graphometer's original CLI and reproduction code are licensed under Apache-2.0. Evidence data is licensed under CC BY 4.0. Modified Liquid AI template files remain subject to the LFM Open License v1.0; see the license map for exact file boundaries.
| License | Files |
|---|---|
| Apache-2.0 | Root documentation and license_map_check.py; all of cli/; public/evidence/assert_wire_consistency.py and public/evidence/summarize_ledgers.py; all of upstream/ except the Liquid-derived files and their NOTICE sidecars listed below |
| CC BY 4.0 | LICENSE_CC_BY_4.0.txt and all other files in public/, including evidence data and recorded demo material |
| LFM Open License v1.0 | LICENSE_LFM_OPEN.txt; each upstream/liquid-*-template/NOTICE; and each such folder's fixed_chat_template.jinja, template.diff, render_before.txt, and render_after.txt |
The three license texts are LICENSE_APACHE-2.0.txt,
LICENSE_CC_BY_4.0.txt, and LICENSE_LFM_OPEN.txt. Each template folder's
NOTICE identifies its Liquid AI repository and revision and describes the
2026-08-01 modification. The sidecars keep the proposed upstream template and
diff payloads byte-identical.
Built by a human steward with heavy AI assistance; every public claim was checked against recorded runs.