Reading what isn’t written

§1

The hidden reading

Many scripts underdetermine pronunciation. Written قدر is consistent with several Arabic words — قَدَر qadar “fate” among them; only the omitted short vowels disambiguate. Hebrew without nikud, Khmer orthography, and unsegmented Thai present the same property in different degrees: the reading exists in the reader’s mind, not on the page.

Recovering it is therefore a constrained generation problem: the output must agree with every glyph that is written, and choose the reading that context implies. That is a job for a sequence-to-sequence model over characters — not for a lookup table, and not for rule systems alone (Khmer proves that), but exactly the shape of task small neural models learn well.

§2

The three tasks

Task	Input → output	Scripts in the index
Diacritization	undiacritized text → same text + short vowels (incl. case endings)	Arabic, Urdu
Vocalization	consonantal text → text + nikud; cantillation passed through	Hebrew
Phonemization (G2P)	orthographic text → phonemes (IPA or romanized)	Thai, Khmer, Persian, Urdu

The same architecture covers all three; only the training corpus differs. That is why one contract and one model format serve the whole family.

§3

How the models work

Secryst models are byte-level encoder–decoders (the ByT5 family) exported to ONNX:

• Byte-level tokenization — no vocabulary files, no out-of-vocabulary tokens, any script works identically. The token table is fixed: byte b maps to id b+3, with a single trailing EOS; pad=0, eos=1, unk=2. Ids are not raw byte values — see the specification.
• Greedy decoding with an optional KV cache — zips that ship decoder-kv.onnx decode incrementally; plain zips recompute. Both must produce identical outputs.
• Small by design — student-tier models are millions of parameters, not billions; they run on CPU in milliseconds per sentence. Teachers (ByT5-base scale) exist for distillation, and the published numbers in Models show what the small ones inherit.

§4

Why local & deterministic

Transliteration and vocalization are infrastructure: they run inside publishing pipelines, archival processing, and embedded runtimes. Secryst’s operating guarantees follow:

• No network at inference — artifacts download once (sha256-verified), then load from cache; every cache hit is re-verified.
• Reproducible — greedy decoding, fixed tokenization, versioned artifacts. Same input, same model version, same output — forever.
• Auditable — the model is a zip you can hash, inspect, and pin; models.yaml is the whole supply chain.

§5

Why we never use LLMs as teachers

Every label in every secryst training corpus is either gold human text or the output of a deterministic rule system. Large language models hallucinate vocalization: they emit fluent, plausible, wrong harakat and nikud. A single systematic hallucination, distilled into a student, becomes a permanent error no eval will localize. We treat LLM-labeled phonology as a poison chain — documented across our Arabic, Hebrew, and Thai experiments — and the constraint is enforced socially in every dataset we accept.

Machine-labeled data is not banned — it is staged: weak labels may pretrain, but only human-verified corpora fine-tune the shipped model (validated: our Hebrew curriculum gained 0.88 DER points this way).

§6

Relation to Interscript

Interscript publishes deterministic transliteration maps — but maps assume vocalized input; undiacritized text cannot be transliterated unambiguously. Secryst is the phonological layer underneath: maps that need it declare a vocalization step and dispatch to the host runtime’s crystal. The crystals themselves have no interscript dependency — a TTS front-end can phonemize Khmer with pip install secryst and nothing else. Interscript defines what a hidden reading is; secryst is how you reveal it, in-process.