mirror of
https://github.com/allaunthefox/Research-Stack.git
synced 2026-08-09 06:45:46 +00:00
82 lines
3 KiB
Text
82 lines
3 KiB
Text
created: 20260507000000000
|
|
modified: 20260507000000000
|
|
tags: ResearchStack LLM Compression Metaprobe Logogram RaTeX Tang9K
|
|
title: Math Logogram Surface Compiler
|
|
type: text/vnd.tiddlywiki
|
|
|
|
! Math Logogram Surface Compiler
|
|
|
|
The Math Logogram Surface Compiler is Surface-1 for the compression-first LLM pipeline. It turns LaTeX, symbolic logogram strings, and chemistry-style math strings into deterministic canonical cells, fixed-width glyph payloads, substitution receipts, and SFT curriculum rows.
|
|
|
|
Durable source: `4-Infrastructure/shim/math_logogram_surface_builder.py`
|
|
|
|
Receipt: `4-Infrastructure/shim/math_logogram_surface_receipt.json`
|
|
|
|
Curriculum: `4-Infrastructure/shim/math_logogram_surface_curriculum.jsonl`
|
|
|
|
!! Pipeline
|
|
|
|
```
|
|
LaTeX / logogram source
|
|
-> deterministic token cells
|
|
-> display-list-like canonical cell hash
|
|
-> 16-byte Tang-compatible glyph payload
|
|
-> substitution receipt
|
|
-> semantic topology regime
|
|
```
|
|
|
|
!! RaTeX Role
|
|
|
|
RaTeX is the preferred future canonicalizer because it lowers LaTeX/math/chemistry syntax into a Rust-native parse/layout/display-list pipeline. Surface-1 currently uses a deterministic Python canonicalizer so the stack can run today without cloning an external dependency.
|
|
|
|
!! Regime Output
|
|
|
|
The compiler tags each sample as `beautiful_topological_folding`, `ugly_asymmetric_pruning`, or `horrible_manifold_tearing` so destructive compression does not get smoothed into prose.
|
|
|
|
!! RRC Projection Bridge
|
|
|
|
[[RRC Logogram Projection Bridge]] maps compiled logogram cells into the
|
|
[[Rainbow Raccoon Compiler]] as `LogogramProjection` objects.
|
|
|
|
```
|
|
canonical cell hash
|
|
-> bounded glyph payload
|
|
-> substitution receipt
|
|
-> semantic regime guard
|
|
-> RRC type witness
|
|
```
|
|
|
|
The bridge produces five `LogogramProjection` type candidates. Four are ordinary
|
|
merge-admissible projections. The `semantic_tear` sample is repaired into a
|
|
quarantine projection with detached semantic mass, so it is preserved without
|
|
being merged into ordinary tokenbook space.
|
|
|
|
!! Typst Omindirection Role
|
|
|
|
[[Typst Omindirection Plugin Surface]] is the preferred document-side renderer for compiled logograms. It gives each symbolic atom explicit direction, chirality, and tone metadata before dense visual flow.
|
|
|
|
```
|
|
compiled logogram cell
|
|
-> omindirection atom
|
|
-> direction field
|
|
-> chirality field
|
|
-> tone/status field
|
|
-> compiled Typst receipt
|
|
```
|
|
|
|
[[Epigenetic Go-Tile Meta-Manifold]] supplies the next computation model: each compiled logogram becomes a placed tile whose mutable expression marks can alter local liberties, captures, and semantic territory without changing the underlying symbol payload.
|
|
|
|
!! Claim Boundary
|
|
|
|
The compiler canonicalizes and compresses symbolic strings. It does not prove math, chemistry, or PDE claims.
|
|
|
|
!! Links
|
|
|
|
* [[LLM Compression Architecture Priors]]
|
|
* [[Semantic Topology Compression Regimes]]
|
|
* [[Rainbow Raccoon Compiler]]
|
|
* [[RRC Logogram Projection Bridge]]
|
|
* [[N-Space LLM Pipeline Tuning]]
|
|
* [[Tang9K 1-Wire Bitpacking Bridge]]
|
|
* [[Typst Omindirection Plugin Surface]]
|
|
* [[Epigenetic Go-Tile Meta-Manifold]]
|