Two tracks: I/O language vs engine representation

Rule: every design note, PR, and finding must label Track A or Track B (or both with a split).

Track A — I/O language (still text)

What it is

Glyph alphabet, STATE-IR, tables, deltas, content ids
glyph_codec.py
Prompt cache + result cache + routers

What it optimizes

Tokens in the API prompt
Human/agent orchestrator memory files

What it cannot do

Change KV layout inside a closed model API
Create a non-text neural medium

When to use

Building agents on OpenAI/Anthropic/etc. APIs
Logging / traces / session state

Kill test

≥20–25% tokens_in vs full history chat; ≤10% task quality drop.

Track B — Engine representation (not text)

What it is

AOC / Manifold KV: store r\* anchors of attention outputs A(q)
Hybrid exact local window + compressed far past
r_Σ gate; A0 mode; optional residual-based ideas (KV-Direct class)
Research code under Research'/maths/

What it optimizes

KV memory and attention compute over long m
Fidelity measured on A(q), then e2e generation

What it needs

Model hooks or custom inference
Per-head/layer empirics

When to use

You control runtime (local vLLM, HF, llama.cpp forks, research harness)

Kill test

Mem ↓ at fixed A-error / logit KL; eventually needle/PPL hold; fallback when r_Σ high.

How they compose


User / tools

    │

    ▼

Track A encode (frame, tables)  ── optional, always available

    │

    ▼

Tokenizer → Embed → Layers

    │

    ▼

Track B AOC on eligible heads   ── if engine owned

    │

    ▼

Logits → tokens → Track A update state

Common confusion (project history)

Mistake	Fix
Selling glyph as “new representation vs text”	Call it denser text packing
Only doing Track A when user wants arch magic	Push AOC docs + MVP
Only doing heavy matroids when prototypes work	Harvested MVP: kmeans/greedy first

Decision tree


Do you own the attention implementation?

  NO  → Track A only (+ provider prompt cache)

  YES → Track B AOC with r_Σ gate

         + Track A for agent logs / tools still recommended