Calculations & proofs
Nothing to hide. Here is what is measured, how, and where to reproduce it. Full derivations and the theorem statements live on the academic results page and in the repo; this is the index.
Reversibility (name the variant)
| Variant | Result | Method |
|---|---|---|
| Hardened 2003–04 kernel | 98.68% rev-or-canonical, 1.31% irreducible | Exhaustive sweep over [a-z]¹–⁴ = 475,254 strings; round-trip str → ACII → Romenagri → ACII → str. |
Repo bindings/c (post-guard) | 67.62% rev-or-canonical, 2.70% crash | Same harness; a reverse-parser regression vs the 2003–04 kernel, reported honestly. |
Heuristic totality (alphabet limits)
Roman→Devanagari heuristic: 100% of [a-z]¹–³ map to Devanagari (raw kernel left 29.7% undefined + 2.703% crash); of the heuristic image, 91.56% is kernel-stable. The alphabet ceiling is absorbed by the front-end, not by changing the kernel.
Compression (lossless)
On an Indic corpus: bytes 1.76×, alphabet 51 → 24 (≈2.12×), byte-positions 3×; BPE merge
gain k < 1. Lossless because the transform is reversible (up to the irreducible floor).
Alphabet bound
The canonical inventory (ACII) is |Ω| = 84 primitive symbols — claimed sufficient for 44+ Brahmic scripts; the minimality of 84 is an open question (stated as open, not proven).
Two layers, do not conflate
Keyword localization (alias/#define-style: localized keyword → canonical construct, e.g.
map.html, HindiC++) preserves semantics. Orthographic transliteration (the
Romenagri kernel) is the reversible script layer — that is where the 98.68% lives.