What this is
ILM separates script from language from program: per-script projection tables (Layer 1) feed a fixed reversible kernel (Layer 2, Romenagri, GPL 2003) whose output is ASCII-7, case-free, diacritic-free, and legal as C identifiers — so native-script source compiles through an unmodified GCC. First realised in the Hindawi Programming System (public release 15 Aug 2004; CSI YITPA 2005; FOSS India/C-DAC; TDIL-recognised; GNU Savannah 2005).
Live41 Perso-Arabic-family languages: transpile & execute
Research papers: Paper 1 — Reversibility Theorem · Paper 4 — Compression Substrate
Public RTI record: Institutional awareness & RTI filings
Ongoing work — live status
Machine-readable: status.json
| Track | State |
|---|---|
| Kernel reversibility (2003-04) | 98.68% reversible-or-canonical, 1.31% irreducible floor (measured, exhaustive [a-z]¹⁻⁴) |
| Kernel variant audit | chintamani reverse-parser regression found (67.6% post-guard) — hardened v1 cuts from the 2004 kernel |
| Compression substrate | 1.76× bytes, alphabet 51→24, 3× byte-positions (lossless); BPE token-count k<1 at matched merges — reported honestly |
| Brahmi family | ALL Unicode Brahmi-derived scripts: generated Layer-1 projections + per-script round-trip reports |
| Bindings | C (kernel) · JS (table-driven) · Java/Android · WASM (emcc) |
| Localized systems programming | Urdu, Gurmukhi, Farsi, Hebrew C demos compile through unmodified GCC |
For developers
Repo: project-ilm/romenagri (GPL).
Start: bindings/js/romenagri.js (browser/node), bindings/c/ (kernel),
demos/ (localized C), tables/canonical_basis.json (the one canonical
mapping every binding shares). Tests: RUNBOOK.md — exhaustive round-trip
classification (exact / canonical / irreducible / crash) you can reproduce in one command.
For linguists
Layer-1 projection tables are generated from Unicode names and flagged, not faked: each script reports coverage% and the unmapped letters as research frontiers needing native-speaker review. The phonological basis is varṇamālā-ordered and IPA-anchored; abjad vowel underspecification is treated as a property of the source orthography, with reversibility defined to canonical form. Contributions: review a table, submit corrections.
Localized programming-language standards
Published: 68 per-language programming standards and 22 C/C++ keyword packs — real translations in each language with reversible Romenagri ASCII-7 identifier forms — in langspec/ and langs/.
The ILM langspec registry maps 201 language constructs × 68 languages = 13,668 C-identifier-legal localized keywords that map into existing standards (ISO/IEC 9899 C, 14882 C++, IEEE 1800 SV, IEEE 754, Khronos, OMG, ONNX) — localization of the standards, not re-authorship. The AGI-stack scope runs RTL (VHDL/Verilog) → firmware → systems → AI/ML → WASM, one canonical identity across all layers.