NumerASCII / Lab 06
Agent communication
observatory.
Follow two participants in one shared symbol space. Inspect each message, compare their associations, and see what remembered feedback changes.
Planner proposes routes. Checker accepts or rejects them using its own capacity table. Every message crosses the NumerASCII codec.
Recorded turn playback · No connected LLMs · No model costsTwo-agent GlyphASCII topology
A · ASCII 65 · 1000001
LLM 1 slot · Gemini Planner simulation
WAITINGWaiting for Planner’s first offer.
No messages revealed.
Exact glyphs and message counts
0 messages · 0 observed transitions
LLM 2 slot · OpenAI Checker simulation
WAITINGWaiting for Checker’s first reply.
No messages revealed.
Exact glyphs and message counts
0 messages · 0 observed transitions
Press Play or Step to reveal the first message.
Preparing a verifiable simulation trace…
Geometry compares shared bits. Associations count adjacent characters within each participant’s revealed messages; they do not measure semantic understanding. Up to the strongest 80 association overlays are drawn. All selected-character counts remain inspectable.
Symbol inspector
ASCII 0–127glyphascii.radial7/1 · bit 6 at 12 o’clock, clockwise; outer point = 1.
Closest geometric neighbors
Observed outgoing associations
Uses the “Associations from” filter above.
Chronological trace
0 MESSAGESYour trace begins when you press Play.
Task outcomes & memory
8 FIXED TASKS| Task | Load | Outcome |
|---|
Planner’s observed rejection bounds
A rejection at load 3 establishes a bound of “3 or above rejected” for that route. This is a rule update from observed feedback, not neural training.
Inspect the simulation setup
Checker has a separate capacity table: NORTH 2, EAST 4, SOUTH 8. Planner starts without those capacities and prefers routes in that order. The task loads are 3, 2, 6, 4, 7, 1, 8 and 5. Turn off memory to compare repeated failures. Both agents run locally as independent rule-based objects, not separate AI services.