@adaptic/utils 0.0.1016 → 0.0.1018

package/dist/index.mjs

@@ -63898,7 +63898,7 @@ async function executeChain(alias, execution) {
 
 var schema_version = 1;
 var policy_source = "docs/llm-provider-migration.md#3-routing-policy (v1.1)";
-var revised = "2026-09-10";
+var revised = "2026-09-11";
 var defaults = {
 	request_timeout_ms: {
 		"hot-path": 30000,
@@ -63953,13 +63953,13 @@ var providers = {
 		base_url_env: "DEEPINFRA_BASE_URL",
 		api_key_env: "DEEPINFRA_API_KEY",
 		secret_path: "llm/deepinfra/apiKey",
-		account_status: "pending-onboarding",
+		account_status: "live",
 		lumic_provider: null,
 		published_rpm: null,
 		published_tpm: null,
 		limits_source: null,
 		docs_url: "https://deepinfra.com/docs",
-		notes: "Rate limits and exact model ids are transcribed from the provider console at W3-02, never guessed."
+		notes: "Sole open-weight host. The routing policy originally spread these models across Z.ai first-party, DeepInfra and Groq; DeepInfra's catalogue carries all of them, so consolidating removes three onboardings, three keys and three terms filings at the cost of roughly 28% on llm.reason's input price versus the first-party GLM anchor. Fewer credentials and fewer trust boundaries is worth more than that margin on one alias."
 	},
 	fireworks: {
 		display_name: "Fireworks AI",
@@ -63969,13 +63969,13 @@ var providers = {
 		base_url_env: "FIREWORKS_BASE_URL",
 		api_key_env: "FIREWORKS_API_KEY",
 		secret_path: "llm/fireworks/apiKey",
-		account_status: "pending-onboarding",
+		account_status: "not-in-scope",
 		lumic_provider: null,
 		published_rpm: null,
 		published_tpm: null,
 		limits_source: null,
 		docs_url: "https://docs.fireworks.ai",
-		notes: "Onboarded as a capacity backstop. No alias routes to it until its published RPM ceiling is recorded in provider-limits (W4-04)."
+		notes: "Superseded by the DeepInfra consolidation: the models this host was chosen for are served from DeepInfra under one account. Retained in the registry so a future re-split needs a config change rather than a new onboarding."
 	},
 	zai: {
 		display_name: "Z.ai",
@@ -63985,13 +63985,13 @@ var providers = {
 		base_url_env: "ZAI_BASE_URL",
 		api_key_env: "ZAI_API_KEY",
 		secret_path: "llm/zai/apiKey",
-		account_status: "pending-onboarding",
+		account_status: "not-in-scope",
 		lumic_provider: null,
 		published_rpm: null,
 		published_tpm: null,
 		limits_source: null,
 		docs_url: "https://docs.z.ai",
-		notes: "First-party GLM host. Section 3 anchors first-party GLM below the converged multi-host price, so it is the primary leg for llm.reason."
+		notes: "Superseded by the DeepInfra consolidation: the models this host was chosen for are served from DeepInfra under one account. Retained in the registry so a future re-split needs a config change rather than a new onboarding."
 	},
 	deepseek: {
 		display_name: "DeepSeek",
@@ -64001,13 +64001,13 @@ var providers = {
 		base_url_env: "DEEPSEEK_BASE_URL",
 		api_key_env: "DEEPSEEK_API_KEY",
 		secret_path: "llm/deepseek/apiKey",
-		account_status: "live",
+		account_status: "not-in-scope",
 		lumic_provider: "deepseek",
 		published_rpm: null,
 		published_tpm: null,
 		limits_source: null,
 		docs_url: "https://api-docs.deepseek.com",
-		notes: "Already a registered lumic provider, so its legs can also be served by the degraded direct transport. Off-peak windows are captured in gateway config at W3-05 for batch scheduling."
+		notes: "First-party DeepSeek API, no longer routed to. Its G1 filing found no commitment not to train on API inputs, no stated retention period, no sub-processor list, and storage under PRC jurisdiction — a materially weaker data posture than any other provider reviewed, and not one to send trading prompts through. The DeepSeek MODELS are still used and are a good fit; they are now served as open weights from DeepInfra under its Zero Data Retention commitment. The weights and the vendor's hosted API are separable, and only the API carried the exposure. Retained in the registry so the distinction stays visible rather than being rediscovered."
 	},
 	groq: {
 		display_name: "Groq",
@@ -64023,7 +64023,7 @@ var providers = {
 		published_tpm: null,
 		limits_source: null,
 		docs_url: "https://console.groq.com/docs",
-		notes: "Section 3 names Groq or Cerebras as the llm.fast primary, but the W3 onboarding list does not include either. See open item OI-01: until an owner resolves it, llm.fast's primary leg is unreachable and the alias serves from its secondary and closed legs."
+		notes: "Superseded by the DeepInfra consolidation: the models this host was chosen for are served from DeepInfra under one account. Retained in the registry so a future re-split needs a config change rather than a new onboarding."
 	},
 	openrouter: {
 		display_name: "OpenRouter",
@@ -64061,51 +64061,55 @@ var aliases = {
 		routes: [
 			{
 				role: "primary",
-				provider: "zai",
-				model_id: null,
-				model_id_status: "pending-provider-confirmation",
-				model_id_source: null,
-				model_family: "GLM-5.3",
-				lumic_model: null,
+				provider: "deepinfra",
+				model_id: "deepseek-ai/DeepSeek-V4-Pro",
+				model_id_status: "confirmed",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "DeepSeek V4 Pro",
+				lumic_model: "deepseek-ai/DeepSeek-V4-Pro",
 				params: {
 					temperature: null,
 					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: false,
-					supports_vision: false
+					supports_cache_control: true,
+					supports_vision: false,
+					context_window: 1048576
 				},
 				price_per_mtok: {
-					input: 1.09,
-					output: 3.43,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 first-party GLM-5.3 anchor; re-verify at W2-02"
-				}
+					input: 1.3,
+					output: 2.6,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
+				},
+				notes: "1.6T MoE, 49B active, 1M context. Its model card cites advanced reasoning and long-running agent tasks, and it is the cheapest model in the flagship tier by a wide margin — a quarter of the incumbent's input price and a tenth of its output price."
 			},
 			{
 				role: "secondary",
 				provider: "deepinfra",
-				model_id: null,
-				model_id_status: "pending-provider-confirmation",
-				model_id_source: null,
-				model_family: "DeepSeek V4 Pro",
-				lumic_model: null,
+				model_id: "zai-org/GLM-5.3",
+				model_id_status: "confirmed",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "GLM-5.3",
+				lumic_model: "zai-org/GLM-5.3",
 				params: {
 					temperature: null,
 					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: false,
-					supports_vision: false
+					supports_cache_control: true,
+					supports_vision: false,
+					context_window: 1048576
 				},
 				price_per_mtok: {
-					input: 1.3,
-					output: 2.6,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 DeepInfra anchor; re-verify at W2-02"
-				}
+					input: 1.2,
+					output: 4,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
+				},
+				notes: "A separate lineage from the primary rather than a smaller sibling, so a DeepSeek-family regression cannot take both legs of this chain at once."
 			},
 			{
 				role: "closed_incumbent",
@@ -64153,54 +64157,80 @@ var aliases = {
 		routes: [
 			{
 				role: "primary",
-				provider: "anthropic",
-				model_id: "claude-opus-4-7",
+				provider: "deepinfra",
+				model_id: "moonshotai/Kimi-K3",
 				model_id_status: "confirmed",
-				model_id_source: "lumic-utils/src/types/openai-types.ts SUPPORTED_MODELS — already routed on in production",
-				model_family: "Fable-class (see policy_note)",
-				lumic_model: "claude-opus-4-7",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "Kimi K3",
+				lumic_model: "moonshotai/Kimi-K3",
 				params: {
 					temperature: null,
-					max_output_tokens: 128000,
+					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
 					supports_cache_control: true,
 					supports_vision: true,
-					context_window: 1000000
+					context_window: 1048576
 				},
 				price_per_mtok: {
-					input: 10,
-					output: 50,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 Fable-class anchor"
-				}
+					input: 2.85,
+					output: 14.25,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
+				},
+				notes: "2.8T open-weight, built explicitly for long-horizon agentic workflows: tool calling, repository navigation, and iterating on logs and test feedback. The most expensive open model selected, which is justified here because this is the hardest workload and the lowest volume."
 			},
 			{
 				role: "secondary",
 				provider: "deepinfra",
-				model_id: null,
-				model_id_status: "pending-provider-confirmation",
-				model_id_source: null,
-				model_family: "Kimi K3",
-				lumic_model: null,
-				shadow_only: true,
+				model_id: "deepseek-ai/DeepSeek-V4-Pro",
+				model_id_status: "confirmed",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "DeepSeek V4 Pro",
+				lumic_model: "deepseek-ai/DeepSeek-V4-Pro",
 				params: {
 					temperature: null,
 					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: false,
-					supports_vision: false
+					supports_cache_control: true,
+					supports_vision: false,
+					context_window: 1048576
 				},
 				price_per_mtok: {
-					input: 2.85,
-					output: 14.25,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 DeepInfra anchor; re-verify at W2-02"
+					input: 1.3,
+					output: 2.6,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
 				},
-				notes: "Shadow-eval only per Section 3. Configured and scored, never served to a caller, so the hardest agentic path keeps its closed primary while the open candidate accrues evidence."
+				notes: "Also cites long-running agent tasks, at a fifth of the primary's output price."
+			},
+			{
+				role: "closed_incumbent",
+				provider: "anthropic",
+				model_id: "claude-opus-4-7",
+				model_id_status: "confirmed",
+				model_id_source: "lumic-utils/src/types/openai-types.ts SUPPORTED_MODELS — already routed on in production",
+				model_family: "Fable-class (see policy_note)",
+				lumic_model: "claude-opus-4-7",
+				params: {
+					temperature: null,
+					max_output_tokens: 128000,
+					supports_temperature: true,
+					supports_json_schema: true,
+					supports_tools: true,
+					supports_cache_control: true,
+					supports_vision: true,
+					context_window: 1000000
+				},
+				price_per_mtok: {
+					input: 10,
+					output: 50,
+					as_of: "2026-09-10",
+					source: "docs/llm-provider-migration.md#3 Fable-class anchor"
+				}
 			}
 		]
 	},
@@ -64223,52 +64253,55 @@ var aliases = {
 		routes: [
 			{
 				role: "primary",
-				provider: "groq",
-				model_id: null,
-				model_id_status: "pending-provider-confirmation",
-				model_id_source: null,
-				model_family: "gpt-oss-120B",
-				lumic_model: null,
+				provider: "deepinfra",
+				model_id: "deepseek-ai/DeepSeek-V4-Flash",
+				model_id_status: "confirmed",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "DeepSeek V4 Flash",
+				lumic_model: "deepseek-ai/DeepSeek-V4-Flash",
 				params: {
 					temperature: null,
 					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: false,
-					supports_vision: false
+					supports_cache_control: true,
+					supports_vision: false,
+					context_window: 1048576
 				},
 				price_per_mtok: {
-					input: 0.039,
-					output: 0.19,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 gpt-oss-120B anchor (DeepInfra price; Groq price unconfirmed pending OI-01)"
+					input: 0.09,
+					output: 0.18,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
 				},
-				notes: "Unreachable until OI-01 is resolved — Groq is not on the W3 onboarding list."
+				notes: "284B MoE, 13B active, 1M context. Measured median 1241 ms on a realistic trading prompt, and the only candidate that answered in directly usable form instead of narrating its reasoning first — which on a hot path is the difference between a parseable answer and a parsing problem."
 			},
 			{
 				role: "secondary",
 				provider: "deepinfra",
-				model_id: null,
-				model_id_status: "pending-provider-confirmation",
-				model_id_source: null,
-				model_family: "gpt-oss-120B",
-				lumic_model: null,
+				model_id: "zai-org/GLM-5.3-Flash",
+				model_id_status: "confirmed",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "GLM-5.3 Flash",
+				lumic_model: "zai-org/GLM-5.3-Flash",
 				params: {
 					temperature: null,
 					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: false,
-					supports_vision: false
+					supports_cache_control: true,
+					supports_vision: true,
+					context_window: 1048576
 				},
 				price_per_mtok: {
-					input: 0.039,
-					output: 0.19,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 DeepInfra anchor; re-verify at W2-02"
-				}
+					input: 0.15,
+					output: 0.5,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
+				},
+				notes: "A different lineage from the primary, with a 1M context and JSON mode. Chosen over the faster Nemotron 3 Nano, which measured the quickest median of any candidate but rejects response_format outright with HTTP 405 — and llm.fast is a structured-output alias, so a leg that cannot express the contract is not a fallback at all, only a slower way to fail."
 			},
 			{
 				role: "closed_incumbent",
@@ -64315,52 +64348,55 @@ var aliases = {
 		routes: [
 			{
 				role: "primary",
-				provider: "deepseek",
-				model_id: "deepseek-v4-flash",
+				provider: "deepinfra",
+				model_id: "openai/gpt-oss-20b",
 				model_id_status: "confirmed",
-				model_id_source: "lumic-utils/src/types/openai-types.ts SUPPORTED_MODELS — already routed on in production",
-				model_family: "DeepSeek V4 Flash",
-				lumic_model: "deepseek-v4-flash",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "gpt-oss-20B",
+				lumic_model: "openai/gpt-oss-20b",
 				params: {
 					temperature: null,
-					max_output_tokens: 64000,
+					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: true,
+					supports_cache_control: false,
 					supports_vision: false,
-					context_window: 1000000
+					context_window: 131072
 				},
 				price_per_mtok: {
-					input: 0.14,
-					output: 0.28,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 V4 Flash anchor; re-verify at W2-02"
-				}
+					input: 0.03,
+					output: 0.14,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
+				},
+				notes: "The cheapest tool-capable model in the catalogue at $0.03/$0.14, which is what a high-volume classification and log-analysis workload should be paying."
 			},
 			{
 				role: "secondary",
-				provider: "zai",
-				model_id: null,
-				model_id_status: "pending-provider-confirmation",
-				model_id_source: null,
-				model_family: "GLM 5.3 Flash",
-				lumic_model: null,
+				provider: "deepinfra",
+				model_id: "inclusionAI/Ling-3.0-flash-Fin",
+				model_id_status: "confirmed",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "Ling 3.0 Flash Fin",
+				lumic_model: "inclusionAI/Ling-3.0-flash-Fin",
 				params: {
 					temperature: null,
 					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: false,
-					supports_vision: false
+					supports_cache_control: true,
+					supports_vision: false,
+					context_window: 262144
 				},
 				price_per_mtok: {
-					input: 1.09,
-					output: 3.43,
-					as_of: "2026-09-10",
-					source: "docs/llm-provider-migration.md#3 GLM first-party anchor; Flash-tier price unconfirmed, re-verify at W2-02"
-				}
+					input: 0.06,
+					output: 0.18,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
+				},
+				notes: "Finance-enhanced through continued training on financial data, with a 262k context. A deliberate second opinion on financial text rather than a generic fallback."
 			},
 			{
 				role: "closed_incumbent",
@@ -64396,7 +64432,7 @@ var aliases = {
 		isolation_capable: false,
 		pinned: true,
 		eval_gate: "none-pinned-judge",
-		policy_note: "PD-6: pinned and never auto-swapped, and deliberately carries no fallback. A judge that silently failed over would grade one model's output against another model's standard, which would corrupt every gate decided by it — so it fails loud instead.",
+		policy_note: "PD-6: pinned and never auto-swapped, and deliberately carries no fallback. Moved off the closed incumbent because a judge running on a vendor we are migrating away from cannot impartially grade that migration — it would be scoring the challengers against its own family.",
 		budget: {
 			basis: "provisional-pre-baseline",
 			monthly_usd: 150,
@@ -64409,39 +64445,34 @@ var aliases = {
 		routes: [
 			{
 				role: "primary",
-				provider: "anthropic",
-				model_id: "claude-sonnet-4-6",
+				provider: "deepinfra",
+				model_id: "deepseek-ai/DeepSeek-V4-Pro",
 				model_id_status: "confirmed",
-				model_id_source: "lumic-utils/src/types/openai-types.ts SUPPORTED_MODELS — already routed on in production",
-				model_family: "Sonnet-class",
-				lumic_model: "claude-sonnet-4-6",
+				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
+				model_family: "DeepSeek V4 Pro",
+				lumic_model: "deepseek-ai/DeepSeek-V4-Pro",
 				params: {
-					temperature: 0,
-					max_output_tokens: 64000,
+					temperature: null,
+					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
 					supports_cache_control: true,
-					supports_vision: true,
-					context_window: 1000000
+					supports_vision: false,
+					context_window: 1048576
 				},
 				price_per_mtok: {
-					input: 3,
-					output: 15,
-					as_of: "2026-09-10",
-					source: "lumic-utils/src/functions/llm-config.ts anthropicModelCosts"
-				}
+					input: 1.3,
+					output: 2.6,
+					as_of: "2026-09-11",
+					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
+				},
+				notes: "Pinned with no fallback (PD-6). A judge that failed over would grade one model's output against another model's standard, corrupting every gate decided by it."
 			}
 		]
 	}
 };
 var open_items = [
-	{
-		id: "OI-01",
-		question: "Section 3 makes Groq or Cerebras the llm.fast primary, but neither appears in the W3 onboarding provider list (DeepInfra, Fireworks, Z.ai, DeepSeek, aggregator). Which host serves llm.fast's primary leg, and is it added to W3?",
-		blocks: "llm.fast primary reachability; W5-03 canary for llm.fast",
-		owner: "CEO"
-	},
 	{
 		id: "OI-02",
 		question: "Section 3 names a Fable-class primary for llm.agentic, but the lumic model registry's strongest registered Anthropic model is claude-opus-4-7. Register a Fable-class model in @adaptic/lumic-utils, or accept opus-4-7 as the llm.agentic primary?",
@@ -64599,6 +64630,52 @@ function orderedRoutes(definition) {
 function routeKeyFor(alias, isolated, role) {
     return `${alias}${isolated ? ISOLATED_SUFFIX : ""}#${role}`;
 }
+/**
+ * Whether one route may serve, and if not, why.
+ *
+ * Extracted from the chain resolver rather than left inline so the admission
+ * rules can be exercised directly. Inline, the only way to prove an exclusion
+ * rule works was to point it at the live table and hope the table still
+ * contained something excludable — which made the proof a statement about how
+ * incomplete the routing policy happened to be that week, and turned finishing
+ * the policy into a test failure. A rule worth enforcing has to be provable on
+ * a route constructed to violate it.
+ *
+ * Order matters: shadow-only is checked before model-id confirmation because a
+ * shadow leg is held back by policy regardless of whether its id is confirmed,
+ * and reporting it as "unconfirmed" would misdescribe a deliberate choice as an
+ * unfinished one.
+ *
+ * The admitting branch carries the confirmed model id rather than leaving the
+ * caller to re-read it. The id is the very thing admission validated, so
+ * handing it back is what lets the caller use it without a non-null assertion
+ * re-stating a check that already happened.
+ *
+ * @param route The authored route.
+ * @param provider The provider entry the route names.
+ * @returns Admission with the confirmed model id, or the reason for exclusion.
+ */
+function routeAdmission(route, provider) {
+    if (route.shadow_only === true) {
+        return {
+            admit: false,
+            reason: "shadow-only: configured and scored, never served to a caller",
+        };
+    }
+    if (route.model_id_status !== "confirmed" || route.model_id === null) {
+        return {
+            admit: false,
+            reason: `model id unconfirmed (${route.model_family}); it is transcribed from the provider console at onboarding, never guessed`,
+        };
+    }
+    if (provider.account_status !== "live") {
+        return {
+            admit: false,
+            reason: `provider account status is "${provider.account_status}"`,
+        };
+    }
+    return { admit: true, modelId: route.model_id };
+}
 /**
  * Resolve an alias to the ordered chain of legs that can serve it today.
  *
@@ -64626,27 +64703,12 @@ function resolveChain(alias, options = {}) {
     const exclusions = [];
     for (const route of orderedRoutes(definition)) {
         const provider = providerEntry(route.provider);
-        if (route.shadow_only === true) {
-            exclusions.push({
-                role: route.role,
-                provider: route.provider,
-                reason: "shadow-only: configured and scored, never served to a caller",
-            });
-            continue;
-        }
-        if (route.model_id_status !== "confirmed" || route.model_id === null) {
-            exclusions.push({
-                role: route.role,
-                provider: route.provider,
-                reason: `model id unconfirmed (${route.model_family}); it is transcribed from the provider console at onboarding, never guessed`,
-            });
-            continue;
-        }
-        if (provider.account_status !== "live") {
+        const admission = routeAdmission(route, provider);
+        if (!admission.admit) {
             exclusions.push({
                 role: route.role,
                 provider: route.provider,
-                reason: `provider account status is "${provider.account_status}"`,
+                reason: admission.reason,
             });
             continue;
         }
@@ -64656,7 +64718,7 @@ function resolveChain(alias, options = {}) {
             role: route.role,
             providerName: route.provider,
             provider,
-            modelId: route.model_id,
+            modelId: admission.modelId,
             lumicModel: route.lumic_model ?? null,
             params: route.params ?? {},
             routeKey: routeKeyFor(alias, isolated, route.role),
@@ -65231,7 +65293,16 @@ function gatewayFor(chain) {
         gatewayTransport = createGatewayTransport({
             baseUrl,
             apiKeyEnv: config.gatewayApiKeyEnv ?? DEFAULT_GATEWAY_KEY_ENV,
-            modelNameFor: (request) => gatewayModelNameFor(request.route, chain),
+            // The chain is re-resolved from the request's own route rather than
+            // captured from the caller above. The transport is cached for the life of
+            // the process, so a captured chain would bind every later alias to
+            // whichever alias happened to dispatch first: its head leg would never
+            // match, would be addressed as "<alias>.fallback.primary" — a name the
+            // gateway does not register — and the 400 would be absorbed by the
+            // fallback chain. The call still succeeds, from the SECONDARY leg, which
+            // is why this reads as healthy traffic while every alias but one quietly
+            // stops using the model it was chosen for.
+            modelNameFor: (request) => gatewayModelNameFor(request.route, resolveChain(request.route.alias, { isolated: request.route.isolated })),
         });
     }
     return gatewayTransport;
@@ -65319,7 +65390,7 @@ async function callLLMByAlias(content, responseFormat = "text", options) {
             cost: 0,
         });
     }
-    const gateway = gatewayFor(chain);
+    const gateway = gatewayFor();
     const attemptLog = [];
     /**
      * Run the chain, falling back from the gateway to the degraded direct path