Finding: Forced result from architecture + maths + cross-pollination
Date: 2026-07-15
Status: synthesis keep
Tracks: A (partial) + B (primary target)
Claim
The highest-leverage “new language” for AI systems is not a fancy human writing system. It is a split media stack:
- Track B: store covers of attention outputs A(q) (hybrid with exact local KV) when r_Σ is low.
- Track A: use palette+frame text only to reduce API/history tax.
Why forced
| Premise | Source |
|---|---|
| Computation is residual + A(q) | Transformer architecture |
| Exact full scores need Θ(md) | Maths Thm0 (17) |
| A(q) often low r_Σ | Empirics Qwen/GPT2 filters |
| Covering size ~ (1/ε)^{r_Σ} independent of m | Thm3 (17) |
| Early pure manifold fails mid layers | Hybrid 18 + brutal analysis |
| Glyph is still BPE text | Direct inspection |
Cross-pollination (compressed)
| Analogy / domain | Transferred axis | Landed as |
|---|---|---|
| Image field / palette | continuum field + indices | Anchors + protos; glyph palette as I/O cousin |
| ε-net / atlas | low-dim cover | r\* anchors |
| Winner-take-all | competitive | A0 mode |
| Diff / WAL | path compression | Hybrid window + Δ |
| Stigmergy | decaying external memory | Paid novelty / absorb |
| Thermo scale | global multiplier | λ / temperature spectrum (advanced) |
| Oriented matroid topes | discrete chambers | Optional locate structure; MVP uses protos |
What we built already
| Artifact | Track |
|---|---|
glyph_codec.py + runs | A |
| Docs in this folder | A+B narrative |
| Research approx pipeline | B (code exists under Research') |
What remains
Track B e2e quality; real tokenizer metrics for A; don’t overclaim matroid ship.
Falsification
- If r_Σ ≈ d on all heads of target models → AOC rarely wins → pivot residual/eviction.
- If glyph fails ≥20% token cut on real agent chats → redesign alphabet / absorb.
failure
null
open_questions
- Residual-only recompute vs AOC: which wins e2e at iso memory?
- Does r_Σ stay low at 32k multi-domain?