LYCHEETAH / LAMAGUEResearch preview · works offline

Meaning between people and AI

What survives
a handover?

The work moves on. The uncertainty, the objection and the permission boundary should travel with it.

Make a handover for your own task ↓

Constructed example · runs offline · no model calls · nothing you enter is sent anywhere

01

Follow the meaning.

What was said · fictional release review

A human-readable record was authored from these four statements. LAMAGUE did not extract or verify it. You can inspect and download that mapping below.

What the next session receives

These changes are deliberate failure controls. They are not outputs from an AI or measurements of everyday misunderstanding.

02

Give every format the same space.

Same complete source record; lossless round trips using the Python runner.
FormatMessage bytesFields restoredWithin budget?

A byte is not a model token. No record is truncated to make it fit. All required decoder programs are assumed available; their code size is excluded for all formats. Gzip includes its base64 and message-prefix overhead. LAMAGUE L1 uses no dictionary.

03

Inspect it. Reproduce it. Challenge it.

Inspect the original record

Check the five human questions

These answers were supplied by the author. This page has not tested whether another person or model reaches them.

See the encoded messages
What this demonstration establishes

It measures exact preservation of one authored record in three formats and compares twelve deliberately constructed controls. The full-record comparison works equally on JSON and decoded LAMAGUE.

    Run this on your own machine

    From the Framework source checkout, with Python 3.10 or newer:

    python3 examples/meaning_handover/demo.py

    Download the reference and a changed handover above, then compare them locally:

    python3 examples/meaning_handover/demo.py \
      --reference handover-reference.json \
      --candidate handover-candidate.json

    Exit 0 means exact record equality, 1 means a changed record, and 2 means malformed input. Nothing is executed on the basis of the result.

    04

    Make your next handover.

    Write what you mean, then review the record before sharing it with your next AI session or collaborator. This form arranges your words; it does not interpret them or approve an action. Your draft stays in this page until you choose a download. There is no autosave.

    Distinguish what you observed from what someone reported.

    One question or missing check per line. Blank means none provided.

    Use the owner's actual boundary, including any limits.

    A role or pseudonym is enough. Entering it does not authenticate that person.

    Add an objection, affected people or review timing

    One role or group per line.

    If blank, the record says timing was not specified.

    The next test belongs to real work.

    Can someone outside this project hand over a task with fewer corrections, at a fixed message budget? The next study needs independently written source tasks, a strong ordinary summary, structured JSON, LAMAGUE, and independent assessors who do not know which arm produced each answer.

    We will need to count omissions, invented facts, mistaken permission and the time spent repairing them. Using the builder or submitting feedback is a usability observation, not that controlled comparison.