Vela 3.0 (vela3-fast) on Core ML: decisions, PII and unsupported-claim spans

Package.swift

@@ -85,6 +85,7 @@ let package = Package(
         .executableTarget(
             name: "IssueTriageDemo", dependencies: ["FluidUse"], exclude: ["README.md", "fetch-issues.sh"]),
         .executableTarget(name: "Vela2Check", dependencies: ["FluidUse"]),
+        .executableTarget(name: "Vela3Check", dependencies: ["FluidUse"]),
         .executableTarget(name: "EvokeCheck", dependencies: ["FluidUse"]),
         .executableTarget(
             name: "EvokeSearchDemo", dependencies: ["FluidUse"], exclude: ["README.md", "fetch-posts.py", "demo.sh"]),

README.md

@@ -312,6 +312,35 @@ PII-masked before sending, replies checked against a source document) and
 [Sources/FrontDoorDemo](Sources/FrontDoorDemo/README.md) (1,000 synthetic chatbot messages screened and routed at ~90
 messages per second: guard on the Neural Engine, route + PII on the GPU).
 
+## Vela 3.0 decisions and spans
+
+`Vela3Manager` runs [vela3-fast](https://huggingface.co/vllm-sr/vela3-fast) (Vela 3.0, vLLM Semantic Router × KR Labs;
+UltraBERT-380M encoder, Apache-2.0) on Core ML: decisions over options you name (choice or yes / no about a state), and
+the two label sets of the release's `span_sets` adapter, personal information (17 types) and the words of an answer its
+context does not support. One encoder pass per question on the GPU (~10 ms for a decision, ~18 ms for a PII pass over a
+paragraph; the Neural Engine is slower for this encoder at every length). Rendering, readout, thresholds and the span
+decoder reproduce the release's own code: against the Python runtime on the same encoders, 0 token mismatches on 23
+Unicode-heavy texts, 0 / 120 decisions and 0 / 16 span sets differ, including windowed inputs
+(`swift run -c release Vela3Check parity …`). One pass holds 1,024 tokens: a longer decision throws `tooLong`; for spans the
+context of an unsupported-claim check is cut first, then long text is read in overlapping windows (the release's rule).
+The pinned snapshot downloads from
+[FluidInference/vela-3.0-coreml](https://huggingface.co/FluidInference/vela-3.0-coreml) on first use (~1.48 GB: two
+multifunction packages, macOS 15 / iOS 18).
+
+```swift
+let vela = try await Vela3Manager.load(from: try await Vela3ModelStore.ensure())
+let r = try await vela.decide(
+    state: "My card was charged twice and I need this fixed today.",
+    questions: [
+        .choice(id: "intent", instructions: "What does the customer want?",
+                options: [("refund", "Money back for the duplicate charge"), ("cancel", "Cancel the order")]),
+        .yesNo(id: "urgent", instructions: "The customer needs a fast response."),
+    ])
+print(r["intent"]!.answer, r["urgent"]!.yes!)
+print(try await vela.pii("Hi, I'm Dana Whitfield, dana@example.com.").spans.map { "\($0.label): \($0.text)" })
+print(try await vela.unsupported(context: "Adults may take up to 4 g a day.", answer: "Adults may take up to 6 g a day.").spans.map(\.text))
+```
+
 ## d1-omni-600M decisions and moderation
 
 `D1OmniManager` runs the text path of Liquid AI's [d1-omni-600M](https://huggingface.co/LiquidAI/d1-omni-600M)

Sources/FluidUse/Vela2/Vela2Manager.swift

@@ -514,12 +514,29 @@ struct Vela2Heads: Sendable {
                 w[name] = Array(f[offset..<(offset + count)])
             }
         }
-        guard let n = w["norm.weight"], let p = w["w_o.bias"] else { throw Vela2Error.invalidAsset("heads.bin") }
-        self.hidden = n.count
-        self.proj = p.count
+        try self.init(weights: w)
+    }
+
+    /// Heads from named fp32 tensors (Vela 3 ships them as safetensors). `span: false` is a decision-only head set
+    /// (no `w_t` / `w_e` / `log_tau_span`), which is never asked for span logits.
+    init(weights w: [String: [Float]], span: Bool = true) throws {
+        guard let n = w["norm.weight"], let p = w["w_o.bias"], let logTau = w["log_tau"]?.first else {
+            throw Vela2Error.invalidAsset("readout heads are missing norm / w_o / log_tau")
+        }
+        let h = n.count, d = p.count
+        var shapes: [(String, Int)] = [
+            ("norm.bias", h), ("w_o.weight", d * h), ("w_q.weight", d * h), ("w_q.bias", d), ("w_p.weight", d * h), ("w_p.bias", d),
+            ("cls_mlp.0.weight", 2 * h * h), ("cls_mlp.0.bias", 2 * h), ("cls_mlp.3.weight", 2 * h), ("cls_mlp.3.bias", 1),
+        ]
+        if span { shapes += [("w_t.weight", d * h), ("w_t.bias", d), ("w_e.weight", d * h), ("w_e.bias", d), ("log_tau_span", 1)] }
+        for (name, count) in shapes where w[name]?.count != count {
+            throw Vela2Error.invalidAsset("readout head \(name) is missing or has \(w[name]?.count ?? 0) values, expected \(count)")
+        }
+        self.hidden = h
+        self.proj = d
         self.w = w
-        self.tau = exp(w["log_tau"]![0])
-        self.tauSpan = exp(w["log_tau_span"]![0])
+        self.tau = exp(logTau)
+        self.tauSpan = span ? exp(w["log_tau_span"]![0]) : 1
     }
 
     private func layerNorm(_ x: ArraySlice<Float>) -> [Float] {