Sorted every model list by what a model is made for (#984)

app/Tests/Models/Corpus/CapabilitySnapshot.txt

@@ -20,10 +20,7 @@
 # what each of them must answer, but it states capabilities alone.
 #
 ALIBABA_CLOUD | qvq-max | ALWAYS_REASONING, CHAT_COMPLETION_API, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
-ALIBABA_CLOUD | qwen-max-latest | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 ALIBABA_CLOUD | qwen-mt-plus | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
-ALIBABA_CLOUD | qwen-plus-latest | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
-ALIBABA_CLOUD | qwen-turbo-latest | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 ALIBABA_CLOUD | qwen-vl-max | CHAT_COMPLETION_API, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 ALIBABA_CLOUD | qwen2.5-14b-instruct-1m | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 ALIBABA_CLOUD | qwen2.5-72b-instruct | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
@@ -67,7 +64,7 @@ FIREWORKS | accounts/fireworks/models/qwen3-235b-a22b | CHAT_COMPLETION_API, FUN
 FIREWORKS | whisper-v3 | SPEECH_INPUT, TEXT_OUTPUT | TRANSCRIPTION | (unknown) | (unknown) | (unknown)
 GOOGLE | gemini-1.0-pro-vision | CHAT_COMPLETION_API, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 GOOGLE | gemini-2.0-flash | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, SPEECH_INPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | (unknown) | 3600 per request | PROVIDER_API countTokens
-GOOGLE | gemini-2.0-flash-live-001 | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, SPEECH_INPUT, SPEECH_OUTPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | (unknown) | (unknown) | (unknown)
+GOOGLE | gemini-2.0-flash-live-001 | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, SPEECH_INPUT, SPEECH_OUTPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | REALTIME | (unknown) | (unknown) | (unknown)
 GOOGLE | gemini-2.5-flash | ALWAYS_REASONING, AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, SPEECH_INPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | 1048576 | 3600 per request | PROVIDER_API countTokens
 GOOGLE | gemini-2.5-flash-image | CHAT_COMPLETION_API, IMAGE_OUTPUT, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | IMAGE_GENERATION | (unknown) | (unknown) | (unknown)
 GOOGLE | gemini-2.5-flash-lite | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, OPTIONAL_REASONING, SPEECH_INPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | 1048576 | 3600 per request | PROVIDER_API countTokens
@@ -115,7 +112,7 @@ LITE_LLM | anthropic/claude-sonnet-5 | CHAT_COMPLETION_API, FUNCTION_CALLING, MU
 LITE_LLM | azure/gpt-5.6 | CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, REASONING_BY_DEFAULT, TEXT_INPUT, TEXT_OUTPUT, WEB_SEARCH | CHAT | 1050000 | (unknown) | TIKTOKEN o200k_base
 LITE_LLM | bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0 | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 LITE_LLM | the-fast-one | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
-MISTRAL | codestral-2508 | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
+MISTRAL | codestral-2508 | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | TEXT_COMPLETION | (unknown) | (unknown) | (unknown)
 MISTRAL | magistral-medium-2506 | ALWAYS_REASONING, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 MISTRAL | ministral-14b-2512 | CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
 MISTRAL | ministral-3b-latest | CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)

app/Tests/Models/Corpus/CorpusOrigin.cs

@@ -17,8 +17,8 @@ public enum CorpusOrigin
     NAMED_BY_A_RULE,
 
     /// <summary>
-    /// The app carries this model in a built-in list, such as the one Alibaba Cloud models are
-    /// picked from when the provider serves no catalog.
+    /// The app carries this model in a built-in list, such as the aliases Anthropic answers to but
+    /// does not list, or the transcription model GWDG serves without naming it.
     /// </summary>
     BUILT_INTO_THE_APP,
 

app/Tests/Models/Corpus/ModelCorpus.cs

@@ -134,23 +134,28 @@ public static class ModelCorpus
     ];
 
     /// <summary>
-    /// Alibaba Cloud. Everything below the two dozen models the app carries is one Qwen tier per
-    /// entry, because each tier answers differently about thinking and vision.
+    /// Alibaba Cloud. One Qwen tier per entry, because each tier answers differently about thinking
+    /// and vision.
     /// </summary>
+    /// <remarks>
+    /// The app carried two dozen of these names in a list of its own until the catalog became the
+    /// only source. Three went with it and are not replaced: qwen-max-latest, qwen-plus-latest and
+    /// qwen-turbo-latest are a naming convention Alibaba has left behind -- its rolling names carry
+    /// no suffix now, and the tier once called turbo is called flash. The ones which stayed are
+    /// here for the other reason: no rule spells any of them out, so they say what becomes of a
+    /// name the rules were not written for.
+    /// </remarks>
     private static readonly CorpusEntry[] ALIBABA_ENTRIES =
     [
-        new(ALIBABA_CLOUD, "qwq-plus", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen-max-latest", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen-plus-latest", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen-turbo-latest", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qvq-max", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen-vl-max", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen-mt-plus", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen2.5-72b-instruct", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen2.5-14b-instruct-1m", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen2.5-omni-7b", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "qwen2.5-vl-72b-instruct", BUILT_INTO_THE_APP),
-        new(ALIBABA_CLOUD, "text-embedding-v3", BUILT_INTO_THE_APP),
+        new(ALIBABA_CLOUD, "qwq-plus", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "qvq-max", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "qwen-vl-max", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "qwen-mt-plus", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "qwen2.5-72b-instruct", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "qwen2.5-14b-instruct-1m", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "qwen2.5-omni-7b", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "qwen2.5-vl-72b-instruct", NAMED_BY_NO_RULE),
+        new(ALIBABA_CLOUD, "text-embedding-v3", NAMED_BY_NO_RULE),
         new(ALIBABA_CLOUD, "qwen3-omni-flash", NAMED_BY_A_RULE),
         new(ALIBABA_CLOUD, "qwen3-vl-plus", NAMED_BY_A_RULE),
         new(ALIBABA_CLOUD, "qwen3-235b-a22b", NAMED_BY_A_RULE),

app/Tests/Models/Corpus/ModelKindCorpus.cs

@@ -33,6 +33,39 @@ public static class ModelKindCorpus
 
         // The one name whose only marker used to be the organization it was published under:
         new(SELF_HOSTED, "sentence-transformers/all-MiniLM-L6-v2", EMBEDDING),
+
+        // Mistral's embedding checkpoint for code. It carries the name of a family which is a text
+        // completion model on this very provider, and stays an embedding model regardless: what a
+        // model is for is said by the word which says it, not by the family it was built from.
+        new(MISTRAL, "codestral-embed", EMBEDDING),
+
+        // Alibaba names its own by the prefix the provider used to filter the catalog by. The rule
+        // says it now, so the prefix is free to go:
+        new(ALIBABA_CLOUD, "text-embedding-v3", EMBEDDING),
+        new(ALIBABA_CLOUD, "text-embedding-v4", EMBEDDING),
+
+        //
+        // What a local Ollama installation serves, taken off its models endpoint rather than
+        // written from memory. The last two are the ones worth having: neither name carries the
+        // word "embed", so both were lost by the phrase the self-hosted provider used to filter
+        // with, and turned up among the chat models instead. Here they are answered by "bge" and
+        // "minilm", which is what those words are written for.
+        //
+        new(SELF_HOSTED, "qwen3-embedding:0.6b", EMBEDDING),
+        new(SELF_HOSTED, "qwen3-embedding:latest", EMBEDDING),
+        new(SELF_HOSTED, "nomic-embed-text:latest", EMBEDDING),
+        new(SELF_HOSTED, "bge-m3:latest", EMBEDDING),
+        new(SELF_HOSTED, "all-minilm:latest", EMBEDDING),
+
+        //
+        // The four whose names say nothing about embedding at all. They are here because the list
+        // above answers them through a word they happen to carry, and these carry none: without a
+        // rule of their own they would count as chat models, which is where they stood.
+        //
+        new(SELF_HOSTED, "stella_en_400M_v5", EMBEDDING),
+        new(SELF_HOSTED, "LaBSE", EMBEDDING),
+        new(SELF_HOSTED, "instructor-xl", EMBEDDING),
+        new(SELF_HOSTED, "gtr-t5-large", EMBEDDING),
     ];
 
     /// <summary>
@@ -57,6 +90,12 @@ public static class ModelKindCorpus
         new(GOOGLE, "gemini-3-pro-image", IMAGE_GENERATION),
 
         new(GOOGLE, "imagen-4.0-generate-001", IMAGE_GENERATION, AnsweredTodayAs: CHAT, Reason: "The markers never knew the name; the family ported in the Google step states it. Nobody noticed because the Google provider shows only names beginning with gemini."),
+
+        new(ALIBABA_CLOUD, "qwen-image-edit", IMAGE_GENERATION),
+
+        // Google ships one image model under a codename instead of a description. Nothing in it
+        // says drawing, and the provider only ever saw it because its catalog was read whole.
+        new(GOOGLE, "nano-banana-pro-preview", IMAGE_GENERATION),
     ];
 
     /// <summary>
@@ -82,6 +121,29 @@ public static class ModelKindCorpus
         new(SELF_HOSTED, "parakeet-tdt-0.6b-v2", TRANSCRIPTION),
         new(SELF_HOSTED, "wav2vec2-large-xlsr-53", TRANSCRIPTION),
         new(MISTRAL, "voxtral-mini-latest", TRANSCRIPTION),
+
+        // The rest of what Mistral actually serves, off its own catalog. The app offered the one
+        // name above alone, because that is the one the documentation names; these three were there
+        // the whole time. Both sizes come as a rolling name and as a dated snapshot.
+        new(MISTRAL, "voxtral-small-latest", TRANSCRIPTION),
+        new(MISTRAL, "voxtral-mini-2602", TRANSCRIPTION),
+        new(MISTRAL, "voxtral-small-2507", TRANSCRIPTION),
+
+        // NVIDIA's other speech line, once plain and once as the hub names it. The second one is
+        // what makes the rule a segment worth keeping: the organization comes off before any rule
+        // sees the name, so what is left has to carry the word on its own.
+        new(SELF_HOSTED, "canary-1b-flash", TRANSCRIPTION),
+        new(SELF_HOSTED, "nvidia/canary-180m-flash", TRANSCRIPTION),
+
+        //
+        // Alibaba's speech line. Every one of these begins with the letter the Alibaba Cloud
+        // provider kept its whole chat list by, so all of them stood among the models somebody
+        // talks to. The last one carries two words at once, and the one which decides is not the
+        // longer one.
+        //
+        new(ALIBABA_CLOUD, "qwen3-asr-flash", TRANSCRIPTION),
+        new(ALIBABA_CLOUD, "qwen3-asr-1.7b", TRANSCRIPTION),
+        new(ALIBABA_CLOUD, "qwen-audio-3.0-asr-flash-streaming", TRANSCRIPTION),
     ];
 
     /// <summary>
@@ -97,6 +159,13 @@ public static class ModelKindCorpus
         // The one name which glues the word to something else, and the reason the three words are
         // not loosened into substrings:
         new(SELF_HOSTED, "xtts-v2", SPEECH_SYNTHESIS),
+
+        new(ALIBABA_CLOUD, "qwen-tts", SPEECH_SYNTHESIS),
+
+        // The Voxtral which speaks instead of listening. It stands here to hold the other half of
+        // Mistral's transcription list: the catalog is asked now, so what is not a transcription
+        // model has to be kept out by what it is, not by the list having been short.
+        new(MISTRAL, "voxtral-mini-tts-latest", SPEECH_SYNTHESIS),
     ];
 
     /// <summary>
@@ -111,6 +180,75 @@ public static class ModelKindCorpus
         new(OPEN_AI, "gpt-realtime-whisper", REALTIME),
 
         new(OPEN_AI, "gpt-live-1", REALTIME, AnsweredTodayAs: CHAT, Reason: "Found in the chat list while testing. The line which succeeds the realtime models dropped the word, and it is even less of a chat partner: it listens and speaks at once and leaves the thinking to a text model behind it."),
+
+        //
+        // Alibaba builds the word into both of its speech lines, and both are here to hold the
+        // rank the realtime rule carries: a connection AI Studio cannot open stays out of every
+        // list, whether the model would otherwise have spoken or listened.
+        //
+        new(ALIBABA_CLOUD, "qwen-tts-realtime", REALTIME),
+        new(ALIBABA_CLOUD, "qwen3-asr-flash-realtime", REALTIME),
+
+        // The other two of Mistral's Voxtral line. The second one carries "transcribe" and stays
+        // out of the transcription list all the same: whatever it does, it does over a connection
+        // the app cannot open, and that is what the rank on the realtime rule is for.
+        new(MISTRAL, "voxtral-mini-realtime-latest", REALTIME),
+        new(MISTRAL, "voxtral-mini-transcribe-realtime-2602", REALTIME),
+
+        //
+        // Google's whole two-way line, taken off its catalog rather than written from memory. It
+        // uses the other word for the thing OpenAI calls realtime, so none of these was recognized
+        // and all four stood among the models to talk to -- which none of them can be. The last
+        // one translates between two people speaking, which is as far from a chat as it gets.
+        //
+        new(GOOGLE, "gemini-3.8-live", REALTIME),
+        new(GOOGLE, "gemini-3.8-live-extended-thinking", REALTIME),
+        new(GOOGLE, "gemini-3.1-flash-live-preview", REALTIME),
+        new(GOOGLE, "gemini-3.5-live-translate-preview", REALTIME),
+
+        // Google's experimental music model, which is recognized already: it carries the other
+        // word, and this holds that the two rules do not fall out with each other.
+        new(GOOGLE, "lyria-realtime-exp", REALTIME),
+    ];
+
+    /// <summary>
+    /// The models which write music.
+    /// </summary>
+    /// <remarks>
+    /// All four off Google's own catalog. They were never seen because the provider showed only
+    /// names beginning with "gemini" -- the same prefix which kept Gemma out, which is why it had
+    /// to go and why these needed a rule of their own before it could.
+    /// </remarks>
+    private static readonly ModelKindExample[] MUSIC_ENTRIES =
+    [
+        new(GOOGLE, "lyria-3.5", MUSIC_GENERATION),
+        new(GOOGLE, "lyria-3-pro-preview", MUSIC_GENERATION),
+        new(GOOGLE, "lyria-3-clip-preview", MUSIC_GENERATION),
+    ];
+
+    /// <summary>
+    /// The models which are handed a job instead of a message.
+    /// </summary>
+    /// <remarks>
+    /// The last two are the point of binding these rules to Google. Perplexity sells deep research
+    /// as well, and what it sells is a chat model -- so the same two words have to mean different
+    /// things at different providers, which is exactly what a binding is for.
+    /// </remarks>
+    private static readonly ModelKindExample[] AGENT_ENTRIES =
+    [
+        new(GOOGLE, "deep-research-preview-04-2026", AGENT),
+        new(GOOGLE, "deep-research-max-preview-04-2026", AGENT),
+        new(GOOGLE, "deep-research-pro-preview-12-2025", AGENT),
+        new(GOOGLE, "antigravity-preview-05-2026", AGENT),
+        new(GOOGLE, "antigravity-preview-09-2026", AGENT),
+    ];
+
+    /// <summary>
+    /// The model which answers out of what it was handed, and says where the answer came from.
+    /// </summary>
+    private static readonly ModelKindExample[] GROUNDED_ANSWERING_ENTRIES =
+    [
+        new(GOOGLE, "aqa", GROUNDED_ANSWERING),
     ];
 
     /// <summary>
@@ -122,13 +260,23 @@ public static class ModelKindCorpus
     ];
 
     /// <summary>
-    /// The models from before chat completions existed.
+    /// The models which continue a text instead of answering in a conversation.
     /// </summary>
     private static readonly ModelKindExample[] TEXT_COMPLETION_ENTRIES =
     [
         new(HELMHOLTZ, "text-davinci-003", TEXT_COMPLETION),
+
+        // The model Mistral serves for filling a gap in a file. Its name does not begin with the
+        // word the provider used to sort its list by, so it stood among the chat models -- next to
+        // Codestral, which does the same job and was kept out by that very word.
+        new(MISTRAL, "mistral-code-fim-latest", TEXT_COMPLETION),
         new(OPEN_AI, "babbage-002", TEXT_COMPLETION),
         new(OPEN_AI, "gpt-3.5-turbo-instruct", TEXT_COMPLETION),
+
+        // The one of these which is not old: Mistral serves Codestral to fill in the middle of a
+        // file. It says so only for Mistral's own catalog, which is why the open weights of the
+        // same name stay a chat model further down.
+        new(MISTRAL, "codestral-latest", TEXT_COMPLETION),
     ];
 
     /// <summary>
@@ -137,6 +285,7 @@ public static class ModelKindCorpus
     private static readonly ModelKindExample[] OCR_ENTRIES =
     [
         new(MISTRAL, "mistral-ocr-latest", OCR),
+        new(ALIBABA_CLOUD, "qwen-vl-ocr", OCR),
     ];
 
     /// <summary>
@@ -173,6 +322,67 @@ public static class ModelKindCorpus
         new(SELF_HOSTED, "llama3.3:70b", CHAT),
         new(OPEN_AI, "gpt-5.1", CHAT),
 
+        // The open weights of the model Mistral itself serves to fill in the middle of a file.
+        // Whoever runs them runs them behind a chat completion API, so here the name means
+        // something to talk to -- which is what binding that other rule to Mistral protects.
+        new(SELF_HOSTED, "codestral-22b-v0.1", CHAT),
+
+        // The other half of that same Ollama installation, and the pair which makes the point:
+        // qwen3.8 and qwen3-embedding are one family and two answers. A rule written to select
+        // rather than to modify would have to beat the family name to get there.
+        new(SELF_HOSTED, "qwen3.8:latest", CHAT),
+        new(SELF_HOSTED, "gpt-oss:latest", CHAT),
+
+        // Half the chat models of the world carry this word, and one of the embedding names above
+        // is one letter longer than it. A name part is what keeps the two apart:
+        new(SELF_HOSTED, "mistral-7b-instruct", CHAT),
+
+        //
+        // What somebody actually comes to Alibaba Cloud for, held here because the provider is
+        // about to stop keeping its chat list by the letter every one of these begins with. The
+        // last one translates rather than converses, and it does so through the chat completion
+        // API like the others, so this is where it belongs.
+        //
+        new(ALIBABA_CLOUD, "qwen3.8-max", CHAT),
+        new(ALIBABA_CLOUD, "qwq-plus", CHAT),
+        new(ALIBABA_CLOUD, "qvq-max", CHAT),
+        new(ALIBABA_CLOUD, "qwen-mt-turbo", CHAT),
+
+        //
+        // Perplexity's whole catalog, which the app carries as a list because there is no route to
+        // ask. A list somebody picked by hand is not filtered at runtime -- a filter over it could
+        // only ever take a model away, never find one -- so it is held here instead: a rule which
+        // turns one of these into something other than a chat model fails the build rather than
+        // quietly emptying the dropdown.
+        //
+        new(PERPLEXITY, "sonar", CHAT),
+        new(PERPLEXITY, "sonar-pro", CHAT),
+        new(PERPLEXITY, "sonar-reasoning", CHAT),
+        new(PERPLEXITY, "sonar-reasoning-pro", CHAT),
+        new(PERPLEXITY, "sonar-deep-research", CHAT),
+
+        //
+        // The other two deep research models of the world, and the reason Google's rule is bound
+        // to Google and written as a prefix. Perplexity and OpenAI both sell something under that
+        // name which answers over the API the app already speaks, so both stay chat models. Only
+        // Google's own line, whose names start with the words, is handed a job instead.
+        //
+        new(OPEN_AI, "o3-deep-research", CHAT),
+        new(OPEN_AI, "o4-mini-deep-research", CHAT),
+
+        //
+        // The counter-sample to the three rules above, taken off the same three catalogs. Every
+        // one of these carries a word which now means something -- code, live, image -- without
+        // being what that word says, and every one of them has to stay a chat model.
+        //
+        new(MISTRAL, "mistral-code-latest", CHAT),
+        new(MISTRAL, "mistral-vibe-cli-latest", CHAT),
+        new(MISTRAL, "zai-glm-latest", CHAT),
+        new(GOOGLE, "gemma-4-31b-it", CHAT),
+        new(GOOGLE, "gemini-3.8-flash", CHAT),
+        new(X, "grok-4.6", CHAT),
+        new(X, "grok-build-0.1", CHAT),
+
         //
         // Three which were questioned while testing and stay all the same. Grok Build is the coding
         // model behind the xAI CLI and answers like any other Grok. The Groq compound systems are
@@ -199,6 +409,9 @@ public static class ModelKindCorpus
         ..TRANSCRIPTION_ENTRIES,
         ..SPEECH_ENTRIES,
         ..REALTIME_ENTRIES,
+        ..MUSIC_ENTRIES,
+        ..AGENT_ENTRIES,
+        ..GROUNDED_ANSWERING_ENTRIES,
         ..COMPUTER_USE_ENTRIES,
         ..TEXT_COMPLETION_ENTRIES,
         ..OCR_ENTRIES,