agent.md — New AI Language project

Instructions for any human or coding agent working in new_ai_language/.

Mission

Build and document a representation stack for AI that is honest about architecture:

Outer track (I/O): denser-than-prose text media (glyph / STATE-IR) so prompts cost less — still BPE text.
Inner track (engine): architecture-native media for inference — Attention Output Cover (AOC / Manifold KV) and related frozen structures from the maths research — not “shorter English.”

Primary insight: models compute in continuous residual + attention outputs; English is a hostile codec into that geometry. Long-context cost is mostly KV (m·d), not paragraph style.

Goals (priority order)

#	Goal	Success looks like
1	Clarity	Docs explain arch + inference so a new agent never confuses prompt tips with engine representation
2	Forced object	AOC (hybrid exact window + r* anchors of A(q)) is specified and linked to maths proofs/empirics
3	I/O codec	Glyph/STATE-IR runnable on traces (`glyph_codec.py`) for agent memory — labeled as Track A only
4	Math fidelity	Core theorems (Thm0 exact scores, covering of A, A0, hybrid, r_Σ gate) live under `maths/` with source paths
5	Findings	New conclusions from thinking + experiments logged under `findings/` — no silent drops
6	No retrain default	Prefer frozen drop-in / prompting / caching / tools; mark anything that needs training
7	Honesty	Never claim novelty or “magic free codec” without evidence; rediscovery is a valid result

Non-goals

Renaming EMA / RAG / summary as new primitives.
Claiming glyph is a non-text neural representation.
Shipping unvalidated “constant memory for all 7B long-ctx” claims (see brutal analysis).
Training soft prompts / new attention kernels unless explicitly marked needs train.

Folder map


new_ai_language/

  agent.md                 ← you are here (goals + how to work)

  README.md                 ← human entry

  docs/                     ← full explanations (arch, inference, tracks, formats)

  findings/                 ← synthesis + new conclusions

  maths/                    ← distilled theorems + excerpts from Research'

  refs/                     ← pointers to repo files / research paths

How to work (iron rules)

Architecture first — residual, QKV, A(q), KV cost — before prompt slogans.
Two tracks always labeled — I/O tax vs engine tax.
Nothing disappears — dead ends → findings/ as failure or open_question.
Cite sources — Research'/maths/... or memory/... paths.
Falsify — every keep needs a kill test (tokens, A-error, or e2e task).
Prefer implementable MVP — prototypes + kmeans anchors over full chirotope if scaling blocks.

Immediate build targets

Keep docs accurate as code evolves.
~~Wire AOC MVP against harvested approx code~~ → done: new_ai_language/aoc/ (see findings/05_aoc_mvp_wired.md). Next: real raw_kv + HF hooks.
Keep glyph_codec.py as Track A demo only.
Append new empirics to findings/.

Related repo paths (do not duplicate blindly)

Path	Role
`Research'/maths/outputs/`	Full derivation corpus
`Research'/maths/outputs/17_*.txt`	A(q) theorems
`Research'/maths/outputs/HARVESTED_COHERENT_SPEC.md`	Practical MVP
`memory/FORCED_RESULT_WHY_TEXT_FAILS.md`	Arch synthesis
`glyph_codec.py`	Track A encoder
`codec_tesla_loop.py`	Signature transfer search
`AGENTS.md` (repo root)	Lab discovery iron rules

Response style for agents

Mechanism-first, no hype.
Say when something is still text.
Say when something needs engine hooks.
Prefer tables and kill tests over poetry.