/// The answers the rebuild has to change, one entry per model.
/// </summary>
/// <remarks>
/// Every entry here was found by running the corpus against the rules as they stand and reading
/// what came back. They are kept out of the snapshot so that the rebuild does not copy them: a
/// snapshot says "do not change this", and a wrong answer is the one thing that must change.
///
/// Three kinds of mistake are collected below, and they are the three the new architecture is meant
/// to make impossible rather than fix one by one:
///
/// - A model which is not a chat model at all is answered as if it were one. The app already knows
/// better: it asks its providers for embedding and transcription models through methods of their
/// own. The capability rules never hear about that and hand out tool calling and image input.
/// - The same model gets two different answers depending on which spelling it arrives in. That is
/// the routing graph leaking into the rules, and it is what the explicit hosts are for.
/// - A prefix rule swallows a variant whose name says the opposite. That is priority written by
/// hand, and it is what computed specificity is for.
/// </remarks>
publicstaticclassExpectedChanges
{
/// <summary>
/// Where the app itself states that a model is not a chat model.
/// </summary>
privateconststringTHE_APP_LISTS_IT_AS_AN_EMBEDDING_MODEL="The app asks every provider for its embedding models separately, through IProvider.GetEmbeddingModels.";
Reason:"An image generation model is answered with the Google default for everything which is not a Gemini: it is told it reads images, writes text, and calls functions, and the one thing it does is not said at all.",
Source:"Provider/Google/ProviderGoogle.cs keeps only names beginning with \"gemini-\" in its chat model list, so this model is never a chat model to begin with; Provider/ModelKindExtensions.cs classifies image generation separately."),
Reason:"Claude 3.5 Sonnet loses its image input when it arrives under the Bedrock spelling: the vendor sits behind a dot rather than a slash, so neither the gateway detection nor the reseller check finds it.",
Source:"The same model as \"anthropic/claude-sonnet-4-0\" and the other Claude entries of this corpus, which all report image input."),
Reason:"Mistral Small 3.1 is told that it thinks, and the very same model is told the opposite when it arrives through Mistral's own API. The rules for the open weights answer for the whole 3 and 4 range in one line, and reasoning arrived with 4.",
Source:"The corpus entry \"mistral-small-2503\" is this model at Mistral and reports no reasoning; Mistral names Magistral as the thinking model of that generation."),
Reason:"The descriptive name is recognized as a GPT model and then placed nowhere: every version rule matches the beginning of the name, which here is the list number. The model loses the reasoning it is known for.",
Source:"The GWDG entry \"gpt-5.5\" of this corpus is the same model and does report reasoning by default."),
Reason:"Granite 4.2 thinks unless the request says otherwise, and there is a rule which says so -- for \"granite-4.2\". Ollama glues the version to the family name, so the rule never sees the models anybody runs locally.",
Source:"IBM documents thinking on by default from Granite 4.2; the Ollama library lists the same checkpoint as \"granite4.2\"."),
Reason:"An NVIDIA model is answered by the Llama rules because it carries the name of the weights it was built from, and the Llama block stands above the Nemotron one. It loses the thinking switch, which is one of the two things NVIDIA changed about those weights.",
Source:"The corpus entry \"nemotron-3-49b\" is the generation after it and does report thinking; NVIDIA documents the detailed thinking switch for the Llama-Nemotron models."),
Reason:"The alias for the non-reasoning GPT-5 is claimed by the \"gpt-5-\" prefix rule and is told it always reasons, which is the one thing its name rules out.",
Source:"OpenAI names this alias as the non-reasoning model of the GPT-5 line; the corpus entry \"gpt-5\" next to it is the reasoning one."),
