Say the Core ML encoder packages download on first use, not once published

[alaineid]

Fixes the two stale claims that the Core ML encoder packages are not yet published. Docs only: no Swift code changes.

All six packages have been published releases of Algorythm-Canada/openjev-models since 2026-10-01, and every manifest OpenJevEncoders embeds has packageDownloadsEnabled: true. Both rows of docs/10-other-models.md now say what happens: the packages are downloaded on first use and checked by SHA-256.

Changed lines

Line	Said	Says now	Evidence
docs/10-other-models.md:17 (Verdict, Swift feasibility)	The package, tokenizer and calibrator are downloaded on first use, once the package is published, and checked by SHA-256 (D-033).	The package, tokenizer and calibrator are downloaded on first use and checked by SHA-256 (D-033).	EncoderPackageManifest+Verdict.swift:19 has packageDownloadsEnabled: true for verdict-m18-fp16; its release verdict-m18-fp16-v1 was published on 2026-10-01 with 3 assets.
docs/10-other-models.md:29 (Laya, Swift feasibility)	Once the packages are published, the Mac's is downloaded on first use; an iPhone downloads a per-length package only when the app asks with prefetch(lengths:), ...	The Mac's package is downloaded on first use; an iPhone downloads a per-length package only when the app asks with prefetch(lengths:), ...	EncoderPackageManifest+Laya.swift:19, 69, 114, 159, 204 have packageDownloadsEnabled: true for laya-m18-fp16 and laya-f18-b1s128-fp16 to laya-f18-b1s1024-fp16; their five releases were published on 2026-10-01 with 3 assets each.

The rest of each row is unchanged. Line 29 keeps "refused with EncoderLoadError.noPackage until then", where "then" is the app's prefetch(lengths:) call, and "The tokenizer and rl_agent_config.json come from the checkpoint at any time", which still holds: with the iPhone's per-length set, LayaBackend.load(from:packageSet:functionCapacity:) fetches them while the device holds no package (LayaBackend.swift:359 to 360).

The releases and the manifests

ghp api 'repos/Algorythm-Canada/openjev-models/releases?per_page=20' on 2026-10-06 lists six releases, none a draft or a prerelease:

Release	Published (UTC)	Assets
verdict-m18-fp16-v1	2026-10-01 16:17:10	3
laya-m18-fp16-v1	2026-10-01 18:29:24	3
laya-f18-b1s128-fp16-v1	2026-10-01 18:30:45	3
laya-f18-b1s256-fp16-v1	2026-10-01 18:31:56	3
laya-f18-b1s512-fp16-v1	2026-10-01 18:33:44	3
laya-f18-b1s1024-fp16-v1	2026-10-01 18:35:21	3

Each holds Manifest.json, Data--com.apple.CoreML--model.mlmodel and Data--com.apple.CoreML--weights--weight.bin. The 18 package files the two manifests list (release tag, asset name, bytes, SHA-256) are exactly the 18 uploaded assets the API lists, taking each asset's digest as its SHA-256: none is missing on either side and none differs.

Manifest line (in Sources/OpenJevEncoders/Store/)	Package	packageDownloadsEnabled
EncoderPackageManifest+Verdict.swift:19	verdict-m18-fp16	true
EncoderPackageManifest+Laya.swift:19	laya-m18-fp16	true
EncoderPackageManifest+Laya.swift:69	laya-f18-b1s128-fp16	true
EncoderPackageManifest+Laya.swift:114	laya-f18-b1s256-fp16	true
EncoderPackageManifest+Laya.swift:159	laya-f18-b1s512-fp16	true
EncoderPackageManifest+Laya.swift:204	laya-f18-b1s1024-fp16	true

Other copies

I searched all 649 tracked files twice: as they are, and with comment markers (///, //, #, *, <!--) dropped at line starts, so a phrase split across a line break or across comment lines still matches. The patterns were "once the package(s) is/are published", "not yet published" and "until the package(s) is/are published", plus rewordings: "unpublished", "not (yet) published, uploaded, hosted or released", "yet to be published", "once", "until", "when" or "before" followed by a package, release or asset being published, uploaded, hosted or available, "downloads are off", and "until" or "before its release exists". I then read every sentence that mentions openjev-models, D-033, packageDownloadsEnabled, "first use", OPENJEV_ENCODER_MODELS, prefetch(lengths:) or "until then", and every Markdown sentence that names Verdict, Laya, an encoder or a package together with a download, release or hosting word. Only the two rows above stated the current state wrongly.

Left as written, because they describe a manifest whose downloads are off, a condition rather than the current state:

• Sources/OpenJevEncoders/Store/EncoderPackageStore.swift: the packageDownloadsUnavailable error text (line 563), the case's doc comment (line 535) and the doc comment at lines 175 to 176.
• Tests/OpenJevEncodersTests/EncoderPackageStoreTests.swift (downloadsOff) and LayaPackageStoreTests.swift (tokenizerWithoutPackage), which clear the flag on a copy of a manifest.
• docs/05-architecture.md:554 to 555 ("a manifest whose downloads are off, as a new package's is before its release exists"), and Tools/encoders/README.md:65 to 66 and Tools/encoders/manifest.py, which already say all six are published and describe the step for a new package.

Also left as written: docs/06-decisions.md, since a decision records what was true when it was written, and the spike reports in docs/spikes/.

Already current: README.md:31, docs/deployment.md:56 and 79, the DocC articles (GettingStarted.md, RunningTheServer.md, ReadingVerdictAndLaya.md), Tools/jevbench/README.md:157 and docs/quality.md:566 describe the downloads as they happen now.

The open pull requests #137 and #140 do not touch docs/10-other-models.md.

Checks

• make lint passes.
• The full swift test was not run: no Swift code changed, and on this Mac it loads 33 GB of models.

[Copilot]

Copilot review overview

🟢 Approval recommended

The documentation now accurately matches the published releases and enabled manifests.

Review effort: Balanced
Findings: None

What changed in this PR

Updates encoder documentation to reflect published Core ML packages.

Changes:

• Removes stale publication caveats for Verdict and Laya.
• Clarifies first-use and prefetch download behavior.

File	Description
docs/​10-other-models.md	Documents current Core ML package availability and downloads.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.