diff --git a/src/content/blog/2026-08-24-phonological-layer-shipped.adoc b/src/content/blog/2026-08-24-phonological-layer-shipped.adoc
index da9a945..7bdc9c3 100644
--- a/src/content/blog/2026-08-24-phonological-layer-shipped.adoc
+++ b/src/content/blog/2026-08-24-phonological-layer-shipped.adoc
@@ -1,4 +1,4 @@
-= The phonological layer: ten models, three wrong turns, and one correction
+= The phonological layer: eleven models, three wrong turns, and one correction
Ronald Tse
v1.0, 2026-08-24
:doctype: book
@@ -13,7 +13,7 @@ haraqat a scribe left out. Turning unwritten Thai into phonemes.
Reading Urdu aloud. Those are learned — and for two years they lived
in a research branch, promising and unshipped.
-This August we shipped the phonological layer: ten neural models under
+This August we shipped the phonological layer: eleven neural models under
the same discipline as the maps — one artifact format, checksums
verified on every load, identical output from Ruby, Python, and
TypeScript, and a measured number next to every claim. This post is
@@ -32,11 +32,13 @@ The catalog now spans four languages and two tiers:
|urd-g2p-1.0 |grapheme→phoneme |CER 14.77 |fp32, 1.3 GiB
|urd-diac-1.0 |diacritization |CER 3.74 |fp32, 1.3 GiB
|heb-diac-1.0 |diacritization |DER 29.0 greedy / 17.5 beam-4 |fp32, parts
-|tha-g2p-base-1.0 |grapheme→phoneme |PER 9.19 (teacher 4.43) |fp32, parts
-|fas-g2p-1.0 |grapheme→phoneme |CER ≈1.6 — above published SOTA on SentenceBench |fp32, parts
+|tha-g2p-base-1.0 |grapheme→phoneme |PER 3.53 greedy (teacher 1.25) |fp32, parts
+|ara-diac-1.0 |diacritization |DER 2.58 — best dedicated on SadeedDiac-25 |fp32, parts
+|fas-g2p-1.0 |grapheme→phoneme |CER ≈1.6 — above the published best on the SentenceBench homograph benchmark |fp32, parts
|tha-g2p-small-1.0 |grapheme→phoneme |PER 2.85 greedy |int8, 246 MiB
|tha-g2p-small-1.0-int4 |same student, 4-bit |byte-identical decode |int4, 193 MiB
|heb-diac-small-1.0 |diacritization |DER 30.37 (teacher 24.79) |fp32, 1.3 GiB
+|ara-diac-small-1.0 |diacritization |DER 8.26 full-set (see update below) |fp32, 1.3 GiB
|===
Every student was distilled from a frozen, independently evaluated
@@ -131,3 +133,31 @@ Try it: `pip install secryst` —
Python, the same in TypeScript. The catalog lives at
https://github.com/interscript/interscript-ml[interscript-ml]; the
measured story behind every number lives in its results log.
+
+== Update, 2026-08-26: two Arabic releases and one subset lesson
+
+The catalog grew to eleven: `ara-diac-1.0` (server) and
+`ara-diac-small-1.0` (client) shipped for Arabic haraqat restoration.
+
+*ara-diac-1.0* is the r6 teacher — and on the full SadeedDiac-25
+benchmark, under the benchmark's own Misraj evaluator protocol, it is
+the best dedicated model measured: 2.5793% DER, ahead of GLM-5.2
+(2.6911), Gemini Flash (3.1926), GPT-4 (3.8645), and Sadeed's own
+1.5B model (7.2915); only Claude 3.7 Sonnet's published 1.3941 sits
+above it. At 580M parameters, runnable locally from a checksummed zip.
+
+*ara-diac-small-1.0* taught us a measurement lesson the hard way. Its
+first published number, 3.66% DER, came from the benchmark's first 300
+paragraphs; the full 1,200-paragraph run scores 8.26%. The teacher
+reproduces its documented value on the same run (2.5815 vs 2.5793), so
+the harness was sound — the subset simply sat in the student's
+training-domain neighborhood and hid a domain-generalization gap. The
+subset figure is withdrawn, the full-set number ships in the metadata,
+and two standing rules follow: student-tier numbers publish from full
+benchmark sets only, and a subset figure is labeled as such at first
+publication.
+
+Also corrected above: the Thai server-tier row now carries the greedy
+numbers (3.53, teacher 1.25) measured under the runtime protocol — the
+decode section earlier in this post explains why the beam-4 figures
+overstated the error.
diff --git a/src/pages/ml.astro b/src/pages/ml.astro
index c1c03fb..99e9c3b 100644
--- a/src/pages/ml.astro
+++ b/src/pages/ml.astro
@@ -38,7 +38,15 @@ const serverModels = [
id: "tha-g2p-base-1.0",
task: "g2p",
pair: "Thai → IPA",
- metric: "PER 9.19 (teacher 4.43)",
+ metric: "PER 3.53 greedy (teacher 1.25)",
+ artifact: "fp32 · 2.6 GiB · parts",
+ status: "released",
+ },
+ {
+ id: "ara-diac-1.0",
+ task: "diacritization",
+ pair: "Arabic → haraqat",
+ metric: "DER 2.58 — best dedicated model on SadeedDiac-25",
artifact: "fp32 · 2.6 GiB · parts",
status: "released",
},
@@ -73,10 +81,18 @@ const clientModels = [
id: "fas-g2p-1.0",
task: "g2p",
pair: "Persian → IPA",
- metric: "CER ≈1.6 · homograph 77.34%",
+ metric: "CER ≈1.6 · SentenceBench homograph 77.34 (published best 76.89)",
artifact: "fp32 · 2.6 GiB · parts",
status: "released",
},
+ {
+ id: "ara-diac-small-1.0",
+ task: "diacritization",
+ pair: "Arabic → haraqat",
+ metric: "DER 8.26 full-set (300-para subset had read 3.66)",
+ artifact: "fp32 · 1.3 GiB",
+ status: "released",
+ },
]
---
@@ -94,12 +110,11 @@ const clientModels = [
Where maps end,
models begin.
- The 289 authority-backed maps cover romanization systems that a
- committee has published. Some conversions have no committee: restoring
- the haraqat a scribe left out, turning unwritten Thai into phonemes,
- reading Urdu aloud. Those are learned — and they ship under the same
- discipline as the maps: one artifact format, checksums verified on
- every load, byte-identical output from every runtime.
+ The 289 authority-backed maps cover romanization systems that a committee has published.
+ Some conversions have no committee: restoring the haraqat a scribe left out, turning
+ unwritten Thai into phonemes, reading Urdu aloud. Those are learned — and they ship under
+ the same discipline as the maps: one artifact format, checksums verified on every load,
+ byte-identical output from every runtime.
@@ -113,12 +128,11 @@ const clientModels = [
The catalogue
Every model, with its number.
- No model is published without a measured metric next to its
- teacher's, on the same harness, in the open. Students are distilled
- from frozen teachers and gated at a pre-agreed error budget. Each
- entry resolves from the models.yaml index; artifacts above GitHub's
- 2 GiB cap ship as sha256-verified parts that the runtimes
- reassemble transparently.
+ No model is published without a measured metric next to its teacher's, on the same
+ harness, in the open. Students are distilled from frozen teachers and gated at a
+ pre-agreed error budget. Each entry resolves from the models.yaml index; artifacts above
+ GitHub's 2 GiB cap ship as sha256-verified parts that the runtimes reassemble
+ transparently.
@@ -162,8 +176,11 @@ const clientModels = [
}
- “Releasing” = passing its parity gate now, entering the release
- pipeline. Distillation budgets and per-model provenance: interscript-ml/docs/RESULTS.md.
+ “Releasing” = passing its parity gate now, entering the release pipeline. Distillation
+ budgets and per-model provenance: interscript-ml/docs/RESULTS.md.
@@ -177,12 +194,11 @@ const clientModels = [
The contract
One artifact. Any runtime.
- A model is a zip — the Interscript Model Format, IMF v1. Anything
- that can read a zip, hash a file, and run two ONNX sessions can
- serve it; no Interscript training code required. Every member is
- sha256-verified against metadata.yaml on load; tampering raises
- loudly. The tokenizer is raw UTF-8 bytes — no vocabulary to
- download, no sentencepiece to drift.
+ A model is a zip — the Interscript Model Format, IMF v1. Anything that can read a zip,
+ hash a file, and run two ONNX sessions can serve it; no Interscript training code
+ required. Every member is sha256-verified against metadata.yaml on load; tampering raises
+ loudly. The tokenizer is raw UTF-8 bytes — no vocabulary to download, no sentencepiece to
+ drift.
- Byte tokenizer — token id = byte + 3, trailing EOS
@@ -211,11 +227,10 @@ const clientModels = [
The crystals
Three runtimes, same bytes out.
- The neural layer is served by secryst — the same
- cross-runtime contract as the map layer. Ruby, Python, and
- TypeScript resolve a model id against the index, fetch (or reuse
- the cache), verify every checksum, and decode with the KV-cache
- graph. Golden sets pin the three implementations to each other.
+ The neural layer is served by secryst — the same cross-runtime contract as the map
+ layer. Ruby, Python, and TypeScript resolve a model id against the index, fetch (or reuse the
+ cache), verify every checksum, and decode with the KV-cache graph. Golden sets pin the three
+ implementations to each other.
@@ -230,12 +245,12 @@ npm install secryst
require "secryst"
model = Secryst::Model.load("tha-g2p-small-1.0")
model.translate("สวัสดี")
-# => "sa˨˩.wat̚˨˩.diː˧"
+# => "sa˨˩.wat̚˨˩.diː˧"
- Point deployments at a mirror with SECRYST_INDEX; pin the
- cache with SECRYST_CACHE. Both are read at call time in
- all three runtimes.
+ Point deployments at a mirror with SECRYST_INDEX; pin the cache with SECRYST_CACHE. Both are read at call time in all three runtimes.
@@ -249,10 +264,9 @@ model.translate("สวัสดี")
Provenance
Measured, or it doesn't ship.
- Teachers are frozen before distillation and never LLM-generated —
- language-model teachers hallucinate diacritics. Every number in the
- ledger links to a protocol in the open results log. BSD-3-Clause,
- code and weights.
+ Teachers are frozen before distillation and never LLM-generated — language-model teachers
+ hallucinate diacritics. Every number in the ledger links to a protocol in the open results
+ log. BSD-3-Clause, code and weights.
@@ -277,7 +291,6 @@ github.com/secryst # runtimes + training
.page-headline em {
font-style: italic;
color: var(--color-brand);
- font-variation-settings: "SOFT" 100, "WONK" 1;
}
.hero-lead {
font-size: var(--text-lead);
@@ -318,7 +331,7 @@ github.com/secryst # runtimes + training
color: var(--color-stone);
text-transform: none;
letter-spacing: 0;
- font-family: var(--font-serif);
+ font-family: var(--font-sans);
font-style: italic;
}
.ledger {