Goals and scope

Problem

Modern LLM usage is dominated by text in, text out. For multi-step agents and long context that is expensive and misaligned with how the network computes.

We want a New AI Language in the broad sense: a stack of media (not slogans) that:

1. Cuts tokens / memory / FLOPs without quality collapse.
2. Improves structured multi-step reliability.
3. Prefers no weight training (prompting, routing, external memory, tools, frozen inference structure).

Goals

G1 — Understand the machine

Document, at engineer level:

• Tokenization → embedding → residual
• Attention, QKV, softmax, output A(q)
• KV cache cost
• What training gradients do vs inference

G2 — Name the right bottleneck

Separate:

• Prompt tax (BPE / prose)
• Engine tax (KV m·d, attention matmul)

G3 — Track A: denser I/O language

Invent and run structured text codecs (palette + frame / STATE-IR) for agent memory.

Label clearly: still text.

G4 — Track B: architecture-native representation

Pursue Attention Output Cover (AOC): store a small cover of attention outputs (and hybrid exact window), not full keys and not chat novels — from the maths research (derivation 17+).

G5 — Ground in research

Keep theorems, empirics (r_Σ, A0, hybrid), and brutal honesty in-tree under maths/ and findings/.

G6 — Falsifiability

Every keep:

• Token or memory metric
• Quality / fidelity metric
• Kill condition

Out of scope (unless marked)

• Soft prompts / activation steering that need training
• New trained attention (e.g. tropical attention from scratch) as default
• Claims of infinite free context

Success criteria (project level)

Relationship to monorepo

This folder is the product narrative + knowledge base.

Heavy code and raw experiments stay in Research'/maths/ and memory/.

Lab discovery rules remain in root AGENTS.md.