adaptive-memory-multi-model-router 2.16.2 → 2.16.3

package/src/server/modelMapper.ts

@@ -72,6 +72,23 @@ export function resolveModel(modelName: string, prompt?: string): ModelMapping |
     return lookupProviderModel(route.primary_model, available);
   }
 
+  // 1b. "jev-auto" → System One single-pass routing (Jev interface pattern).
+  // Lazy + guarded: falls back to the heuristic router if the jev module
+  // or its weights are unavailable.
+  if (modelName === "jev-auto") {
+    const query = prompt || "";
+    let route: ReturnType<typeof routeQuery>;
+    try {
+      // eslint-disable-next-line @typescript-eslint/no-var-requires
+      const { jevRoute } = require("../routing/jev/jevRouter");
+      route = jevRoute(query);
+    } catch {
+      route = routeQuery(query);
+    }
+    if (!route.primary_model) return null;
+    return lookupProviderModel(route.primary_model, available);
+  }
+
   // 2. "provider/model" format → direct lookup
   if (modelName.includes("/")) {
     const mapping = lookupProviderModel(modelName, available);

package/tests/routing/jev.test.ts

@@ -0,0 +1,110 @@
+/**
+ * Jev router tests — engine contract, routing shape, fallback behavior.
+ */
+import { describe, it, expect, beforeAll } from "vitest";
+import { scoreOptions, scoreScalar, scoreNoul } from "../../src/routing/jev/optionAttention";
+import { jevRoute, loadWeights, resetWeightsCache, MIN_CONFIDENCE } from "../../src/routing/jev/jevRouter";
+import type { OptionAttentionWeights } from "../../src/routing/jev/optionAttention";
+
+function toyWeights(): OptionAttentionWeights {
+  const D = 16;
+  const rand = (s: number) => {
+    // deterministic LCG
+    let x = s;
+    return () => {
+      x = (x * 1103515245 + 12345) % 2147483648;
+      return (x / 2147483648 - 0.5) * 0.2;
+    };
+  };
+  const r = rand(42);
+  const fill = (rows: number, cols: number) => Array.from({ length: rows }, () => Array.from({ length: cols }, () => r()));
+  return {
+    emb: fill(2048, D),
+    wq: fill(D, D),
+    bq: new Array(D).fill(0),
+    w: Array.from({ length: D }, () => r()),
+    b: [0],
+    ws: Array.from({ length: D }, () => r()),
+    bs: [0],
+    temperature: 1.0,
+    dim: D,
+  };
+}
+
+describe("optionAttention engine", () => {
+  const W = toyWeights();
+
+  it("choice probabilities sum to 1 and match option count", () => {
+    const r = scoreOptions("write a python function", ["groq fast cheap", "openai premium quality", "ollama local"], W);
+    expect(r.probs).toHaveLength(3);
+    const sum = r.probs.reduce((a, b) => a + b, 0);
+    expect(sum).toBeCloseTo(1.0, 5);
+    expect(r.probs.every((p) => p >= 0 && p <= 1)).toBe(true);
+  });
+
+  it("handles unseen options (dynamic option sets)", () => {
+    const r = scoreOptions("math proof", ["brand-new-provider strengths:math quality:0.9", "another"], W);
+    expect(r.probs).toHaveLength(2);
+    expect(r.top[0].prob).toBeGreaterThan(0);
+  });
+
+  it("is deterministic", () => {
+    const a = scoreOptions("same ctx", ["x", "y"], W);
+    const b = scoreOptions("same ctx", ["x", "y"], W);
+    expect(a.probs).toEqual(b.probs);
+  });
+
+  it("empty context and empty options do not throw", () => {
+    expect(() => scoreOptions("", ["only"], W)).not.toThrow();
+    const r = scoreOptions("ctx", [], W);
+    expect(r.probs).toHaveLength(0);
+  });
+
+  it("score head returns value in [0,1]", () => {
+    const s = scoreScalar("design a distributed system", W);
+    expect(s).toBeGreaterThanOrEqual(0);
+    expect(s).toBeLessThanOrEqual(1);
+  });
+
+  it("noul returns calibrated boolean", () => {
+    const n = scoreNoul("write sql", "requires code generation", W);
+    expect(typeof n.value).toBe("boolean");
+    expect(n.prob).toBeGreaterThanOrEqual(0.5);
+    expect(n.prob).toBeLessThanOrEqual(1);
+  });
+
+  it("100 options in < 150ms (regression guard; prod path is ~32 opts + query cache ≈ 2ms)", () => {
+    const opts = Array.from({ length: 100 }, (_, i) => `provider-${i} strengths:code quality:0.${i % 10}`);
+    const t0 = Date.now();
+    scoreOptions("some long context here ".repeat(10), opts, W);
+    expect(Date.now() - t0).toBeLessThan(150);
+  });
+});
+
+describe("jevRouter", () => {
+  beforeAll(() => resetWeightsCache());
+
+  it("falls back to System 2 when no weights file", () => {
+    // weights file either exists (trained) or not; both paths must return a decision
+    const d = jevRoute("write a haiku");
+    expect(d).toBeTruthy();
+    expect(typeof d.primary_model === "string" || d.primary_model === null).toBe(true);
+    expect(d.reasoning).toBeTruthy();
+  });
+
+  it("returns a full RouteDecision shape", () => {
+    const d = jevRoute("implement a rate limiter using token bucket");
+    expect(d).toHaveProperty("fallback_models");
+    expect(d).toHaveProperty("confidence");
+    expect(d).toHaveProperty("reasoning");
+    expect(d).toHaveProperty("estimated_cost");
+    if (d.primary_model) {
+      expect(d.confidence).toBeGreaterThan(0);
+    }
+  });
+
+  it("low-confidence path delegates to heuristics", () => {
+    const d = jevRoute("zzz qwerty obscure gibberish prompt");
+    expect(d.reasoning).toBeTruthy();
+  });
+});

package/tools/calibrate_temp.py

@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+"""
+calibrate_temp.py — post-train temperature calibration + weight rounding.
+
+Binary-searches temperature so mean(top_prob) on the distillation corpus ≈
+TARGET (slightly below train accuracy → honest calibration), then rounds all
+floats to 4 decimals (≈2× smaller JSON) and rewrites the weights file.
+
+Usage: python3 tools/calibrate_temp.py [--target 0.90]
+"""
+
+import argparse
+import json
+import math
+import os
+import sys
+
+import numpy as np
+
+sys.path.insert(0, os.path.dirname(__file__))
+from train_jev import tokenize, softmax  # reuse exact tokenizer
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--data", default="data/jev-distill.jsonl")
+    ap.add_argument("--weights", default="src/routing/jev/weights/jev-router-weights.json")
+    ap.add_argument("--target", type=float, default=0.90)
+    args = ap.parse_args()
+
+    with open(args.weights) as f:
+        W = json.load(f)
+    emb = np.array(W["emb"]); wq = np.array(W["wq"]); bq = np.array(W["bq"])
+    w = np.array(W["w"]); b = np.array(W["b"]); D = W["dim"]
+    scale = 1.0 / math.sqrt(D)
+
+    rows = [json.loads(l) for l in open(args.data) if l.strip()]
+    ctx_tok = [np.array(tokenize(r["context"])) for r in rows]
+    labels = np.array([r["label"] for r in rows])
+
+    # precompute all scores once (temperature-independent)
+    all_scores = []
+    for i, r in enumerate(rows):
+        E = emb[ctx_tok[i]]
+        Ho = np.stack([emb[np.array(tokenize(o))].mean(axis=0) for o in r["options"]])
+        Q = np.tanh(Ho @ wq.T + bq)
+        A = softmax((Q @ E.T) * scale, axis=1)
+        C = A @ E
+        all_scores.append((Q * C) @ w + b[0])
+    acc = float(np.mean([int(np.argmax(s)) == y for s, y in zip(all_scores, labels)]))
+    print(f"train top-1 (T=1): {acc:.3f}")
+
+    def mean_top_prob(T):
+        ps = [softmax(s / T) for s in all_scores]
+        return float(np.mean([p.max() for p in ps]))
+
+    lo, hi = 0.1, 8.0
+    for _ in range(40):
+        mid = (lo + hi) / 2
+        if mean_top_prob(mid) > args.target:
+            lo = mid   # too confident → raise T
+        else:
+            hi = mid
+    T = (lo + hi) / 2
+    print(f"calibrated temperature: {T:.4f} → mean top-prob {mean_top_prob(T):.3f} (target {args.target})")
+
+    # round to shrink file
+    W["emb"] = np.round(emb, 4).tolist()
+    W["wq"] = np.round(wq, 4).tolist()
+    W["bq"] = np.round(bq, 4).tolist()
+    W["w"] = np.round(w, 4).tolist()
+    W["b"] = np.round(b, 4).tolist()
+    W["ws"] = np.round(np.array(W["ws"]), 4).tolist()
+    W["bs"] = np.round(np.array(W["bs"]), 4).tolist()
+    W["temperature"] = round(T, 4)
+
+    with open(args.weights, "w") as f:
+        json.dump(W, f)
+    print(f"rewrote {args.weights} ({os.path.getsize(args.weights)/1024:.0f} KB)")
+
+if __name__ == "__main__":
+    main()

package/tools/distill.mjs

@@ -0,0 +1,143 @@
+#!/usr/bin/env node
+/**
+ * distill.mjs — System 2 → System 1 distillation corpus generator.
+ *
+ * Runs advancedRouter.routeQuery (the heuristic "System 2" router) over a
+ * synthetic prompt corpus spanning A3M's routing domains, and emits JSONL:
+ *
+ *   {"context": "<prompt>", "label": <provider idx>, "complexity": 0.42,
+ *    "needs_code": true, "options": ["<model option text>", ...]}
+ *
+ * Train with tools/train_jev.py afterwards.
+ *
+ * Usage:  npm run build && node tools/distill.mjs > data/jev-distill.jsonl
+ */
+
+import { routeQuery, MODEL_PROFILES } from "../dist/routing/advancedRouter.js";
+import { writeFileSync, mkdirSync, readFileSync, existsSync } from "fs";
+
+// ---------------------------------------------------------------------------
+// Synthetic prompt corpus — domain × complexity × intent coverage
+// ---------------------------------------------------------------------------
+
+const TEMPLATES = [
+  // trivial / short chat
+  ["hi", "hello there", "thanks", "ok got it", "good morning", "hey quick question"],
+  ["what is 2+2", "define entropy", "what does HTTP stand for", "name a primary color", "capital of France"],
+  // simple code
+  ["write a python function to reverse a linked list", "add two numbers in javascript", "css center a div",
+   "sql select where clause", "regex match an email", "bash list files by size", "git undo last commit",
+   "python read a csv file", "js fetch with async await", "html form with validation"],
+  // moderate code
+  ["implement a LRU cache in typescript with O(1) operations", "write a rate limiter using token bucket",
+   "refactor this express app into clean architecture layers", "design a postgres schema for multi-tenant SaaS",
+   "write unit tests for a react hook", "dockerize a fastapi service with healthchecks",
+   "explain this stack trace and fix the race condition", "optimize this slow SQL query with proper indexes"],
+  // hard / architecture
+  ["design a distributed rate limiter for 10k rps across regions", "architecture review: should we move from REST to gRPC",
+   "implement consensus for a 5-node kv store", "migrate a monolith to event-driven microservices with zero downtime",
+   "design a vector database index for billion-scale embeddings", "review this kubernetes manifest for production readiness"],
+  // math / reasoning
+  ["prove the integral of 1/x is ln|x|", "solve this differential equation numerically", "bayes theorem applied to medical testing",
+   "calculate the eigenvalues of this matrix", "explain the central limit theorem with a proof", "monte carlo simulation in python"],
+  // domains
+  ["diagnose chest pain differential", "explain insulin resistance pathophysiology", "clinical trial design for a new oncology drug",
+   "contract clause review for liability limits", "GDPR compliance checklist for a saas", "case law on fair use for AI training data",
+   "portfolio optimization with markowitz", "explain this 10-K cash flow statement", "hedge delta-neutral options strategy",
+   "fine-tune bert for sentiment classification", "compare adam vs sgd optimizer convergence", "interpret this shapley value plot",
+   "promote soil health in organic farming", "structural load calculations for a bridge", "chemical synthesis route planning"],
+  // translation / multilingual
+  ["translate this sentence to japanese: good morning", "traduce esta frase al español",
+   " übersetze diesen Satz ins Deutsche", "translate and localize this UI string to hindi"],
+  // creative
+  ["write a haiku about the sea", "short story about a robot learning to paint",
+   "marketing copy for an eco water bottle", "write a limerick about databases", "brainstorm names for a coffee startup"],
+  // security
+  ["penetration test plan for a web app", " OWASP top 10 mitigations", "secure secret storage in kubernetes",
+   "threat model an oauth flow", "analyze this suspicious payload"],
+  // devops / infra
+  ["terraform module for vpc with private subnets", "kubernetes hpa based on custom metrics",
+   "ci pipeline with canary deploys", "postgres replication lag troubleshooting", "nginx rate limiting config"],
+  // multimodal
+  ["describe this image and extract the text", "transcribe this audio file to text",
+   "analyze this chart and summarize trends", "generate alt text for this screenshot"],
+  // long context
+  ["summarize this 50 page document about climate policy: " + "policy considerations ".repeat(80),
+   "review this long transcript and extract action items: " + "the team discussed ".repeat(100)],
+];
+
+const LENGTHS = ["", "be brief", "be detailed and thorough", "step by step",
+  "explain like I am five", "with code examples", "in a table", "as a checklist"];
+
+const PREFIXES = ["", "please ", "I need you to ", "can you ", "urgent: ", ""];
+
+function corpus() {
+  const prompts = [];
+  for (const group of TEMPLATES) {
+    for (const p of group) {
+      prompts.push(p);
+      for (let k = 0; k < 3; k++) {
+        const pre = PREFIXES[Math.floor(Math.random() * PREFIXES.length)];
+        const suf = LENGTHS[Math.floor(Math.random() * LENGTHS.length)];
+        prompts.push(pre + p + (suf ? " — " + suf : ""));
+      }
+    }
+  }
+  return prompts;
+}
+
+// ---------------------------------------------------------------------------
+// Run
+// ---------------------------------------------------------------------------
+
+const OUT_DIR = new URL("../data/", import.meta.url).pathname;
+mkdirSync(OUT_DIR, { recursive: true });
+
+// warm profile cache
+routeQuery("warmup");
+const profiles = MODEL_PROFILES;
+const modelNames = Object.keys(profiles);
+if (modelNames.length === 0) {
+  console.error("no model profiles — check provider config");
+  process.exit(1);
+}
+
+// option text must EXACTLY match jevRouter.optionTextForModel
+function optionTextForModel(model) {
+  const p = profiles[model];
+  const strengths = (p.strengths || []).slice(0, 6).join(",");
+  const cost = p.cost_per_1k_input != null ? ((p.cost_per_1k_input + (p.cost_per_1k_output || 0)) / 2).toFixed(4) : "?";
+  return `${model} provider:${p.providerName || "?"} type:${p.type || "api"} strengths:${strengths} quality:${p.quality_score ?? "?"} cost:${cost}`;
+}
+
+const options = modelNames.map(optionTextForModel);
+
+// merge benchmark prompts if present
+let bench = [];
+const benchPath = new URL("../eval/benchmark_dataset.jsonl", import.meta.url).pathname;
+if (existsSync(benchPath)) {
+  bench = readFileSync(benchPath, "utf8").trim().split("\n").map((l) => JSON.parse(l).prompt);
+}
+
+const all = [...new Set([...corpus(), ...bench])];
+const rows = [];
+for (const prompt of all) {
+  const d = routeQuery(prompt);
+  if (!d.primary_model) continue;
+  const label = modelNames.indexOf(d.primary_model);
+  if (label < 0) continue;
+  rows.push({
+    context: prompt,
+    label,
+    complexity: d.features?.complexity ?? 0.3,
+    needs_code: d.features?.has_code ?? false,
+    options,
+  });
+}
+
+const outPath = OUT_DIR + "jev-distill.jsonl";
+writeFileSync(outPath, rows.map((r) => JSON.stringify(r)).join("\n"));
+console.error(`wrote ${rows.length} rows → ${outPath} (${modelNames.length} candidate models)`);
+const dist = {};
+for (const r of rows) dist[modelNames[r.label]] = (dist[modelNames[r.label]] || 0) + 1;
+console.error("label distribution:", JSON.stringify(dist, null, 0).slice(0, 400));

package/tools/train_jev.py

@@ -0,0 +1,213 @@
+#!/usr/bin/env python3
+"""
+train_jev.py — train the OptionAttention System One router (numpy only, vectorized).
+
+Reads data/jev-distill.jsonl (from tools/distill.mjs), trains:
+  - choice head: provider option → calibrated probability   (cross-entropy)
+  - score head:  context → complexity in [0,1]              (MSE + sigmoid)
+  - noul head:   context+"question" → boolean prob          (CE, 2-way)
+
+Architecture (must match src/routing/jev/optionAttention.ts):
+  q_i  = tanh(Wq · meanPool(trigrams(option)) + bq)
+  a_ij = softmax_j(q_i · E_j / sqrt(D))
+  c_i  = Σ_j a_ij E_j
+  s_i  = w · (q_i ⊙ c_i) + b          →  softmax(s/T) = probs
+
+Usage:
+    python3 tools/train_jev.py [--epochs 200] [--dim 64] [--lr 0.5]
+"""
+
+import argparse
+import json
+import math
+import os
+import sys
+import time
+
+import numpy as np
+
+VOCAB = 2048
+CTX_CAP = 160
+
+def trigram_hash(b0, b1, b2):
+    h = 0x811C9DC5
+    for c in (b0 & 0xFF, b1 & 0xFF, b2 & 0xFF):
+        h = ((h ^ c) * 0x01000193) & 0xFFFFFFFF
+    return h % VOCAB
+
+def tokenize(text, cap=CTX_CAP):
+    s = " " + text + " "
+    out = []
+    for i in range(len(s) - 2):
+        out.append(trigram_hash(ord(s[i]), ord(s[i+1]), ord(s[i+2])))
+        if len(out) >= cap:
+            break
+    return out or [trigram_hash(32, 32, 32)]
+
+def softmax(x, axis=-1):
+    x = x - np.max(x, axis=axis, keepdims=True)
+    e = np.exp(x)
+    return e / np.maximum(e.sum(axis=axis, keepdims=True), 1e-9)
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--data", default="data/jev-distill.jsonl")
+    ap.add_argument("--out", default="src/routing/jev/weights/jev-router-weights.json")
+    ap.add_argument("--epochs", type=int, default=200)
+    ap.add_argument("--dim", type=int, default=64)
+    ap.add_argument("--lr", type=float, default=0.5)
+    args = ap.parse_args()
+
+    rows = []
+    with open(args.data) as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                rows.append(json.loads(line))
+    if not rows:
+        sys.exit("no training rows — run tools/distill.mjs first")
+
+    rng = np.random.default_rng(42)
+    D = args.dim
+    lr = args.lr
+    lr_c = args.lr          # choice path
+    lr_s = 0.20             # score path
+
+    emb = rng.normal(0, 0.40, (VOCAB, D))
+    wq = rng.normal(0, 0.40, (D, D))
+    bq = np.zeros(D)
+    w = rng.normal(0, 1.2, D)
+    b = np.zeros(1)
+    ws = rng.normal(0, 0.30, D)
+    bs = np.array([0.0])
+    temperature = 1.0
+
+    NQ = "requires code generation"
+    nq_yes_tok = np.array(tokenize(NQ + " — yes"))
+    nq_no_tok = np.array(tokenize(NQ + " — no"))
+
+    labels = np.array([r["label"] for r in rows])
+    complexity = np.array([r["complexity"] for r in rows])
+    needs_code = np.array([1.0 if r["needs_code"] else 0.0 for r in rows])
+    n = len(rows)
+
+    ctx_tok = [np.array(tokenize(r["context"])) for r in rows]
+    opt_tok = [[np.array(tokenize(o)) for o in r["options"]] for r in rows]
+    opt_lens = np.array([[len(t) for t in orow] for orow in opt_tok], dtype=float)
+
+    scale = 1.0 / math.sqrt(D)
+    M = len(rows[0]["options"])
+
+    def forward_choice(i, opt_tokens):
+        E = emb[ctx_tok[i]]                              # (T, D)
+        Ho = np.stack([emb[ot].mean(axis=0) for ot in opt_tokens])  # (M, D)
+        Q = np.tanh(Ho @ wq.T + bq)                      # (M, D)
+        S = (Q @ E.T) * scale                            # (M, T)
+        A = softmax(S, axis=1)
+        C = A @ E                                        # (M, D)
+        scores = (Q * C) @ w + b[0]
+        return E, Ho, Q, A, C, scores
+
+    print(f"rows={n} options={M} dim={D} epochs={args.epochs} lr={lr}")
+    t_start = time.time()
+    for epoch in range(args.epochs):
+        ce = se = ne = 0.0
+        correct = 0
+        order = rng.permutation(n)
+        for i in order:
+            y = labels[i]
+            E, Ho, Q, A, C, scores = forward_choice(i, opt_tok[i])
+            probs = softmax(scores / temperature)
+            ce += -math.log(max(probs[y], 1e-9))
+            if int(np.argmax(probs)) == y:
+                correct += 1
+
+            # ---- backward (vectorized over options) ----
+            ds = probs.copy()
+            ds[y] -= 1.0
+            ds /= temperature                              # (M,)
+
+            w -= lr_c * ((Q * C) * ds[:, None]).sum(0)     # scoring vector
+            b -= lr_c * 0.10 * ds.sum()
+
+            dQ = ds[:, None] * (w * C) * (1 - Q * Q)       # (M, D)
+            wq -= lr_c * 0.10 * dQ.T @ Ho
+            bq -= lr_c * 0.10 * dQ.sum(0)
+            dHo = dQ @ wq                                  # (M, D)
+
+            # option-token embedding grads — one batched np.add.at per row
+            flat_idx = np.concatenate([t for t in opt_tok[i]])
+            flat_rep = np.concatenate([np.full(len(t), m) for m, t in enumerate(opt_tok[i])])
+            vals = (dHo[flat_rep] / opt_lens[i][flat_rep, None])
+            np.add.at(emb, flat_idx, -lr_c * 0.10 * vals)
+
+            # context embedding grads: dE = A.T @ dC
+            dC = ds[:, None] * (w * Q)                     # (M, D)
+            dE = A.T @ dC                                  # (T, D)
+            np.add.at(emb, ctx_tok[i], -lr_c * 0.10 * dE)
+
+            # ---- score head ----
+            p = E.mean(axis=0)
+            z = float(ws @ p + bs[0])
+            sig = 1 / (1 + math.exp(-z))
+            se += (sig - complexity[i]) ** 2
+            gz = (sig - complexity[i]) * sig * (1 - sig)
+            ws -= lr_s * gz * p
+            bs -= lr_s * gz * 0.10
+            np.add.at(emb, ctx_tok[i], -lr_s * 0.05 * (gz * ws) / E.shape[0])
+
+            # ---- noul head (every 4th epoch) ----
+            if epoch % 4 == 0:
+                yn = 0 if needs_code[i] > 0.5 else 1
+                ntoks = [nq_yes_tok, nq_no_tok]
+                Hn = np.stack([emb[ot].mean(axis=0) for ot in ntoks])
+                Qn = np.tanh(Hn @ wq.T + bq)
+                Sn = (Qn @ E.T) * scale
+                An = softmax(Sn, axis=1)
+                Cn = An @ E
+                ns = (Qn * Cn) @ w + b[0]
+                npr = softmax(ns)
+                ne += -math.log(max(npr[yn], 1e-9))
+                gn = npr.copy()
+                gn[yn] -= 1.0
+                dQn = gn[:, None] * (w * Cn) * (1 - Qn * Qn)
+                wq -= lr_c * 0.05 * dQn.T @ Hn
+                bq -= lr_c * 0.05 * dQn.sum(0)
+
+        if epoch % 10 == 0 or epoch == args.epochs - 1:
+            el = time.time() - t_start
+            print(f"epoch {epoch:4d}  CE={ce/n:.4f} (chance={math.log(M):.3f})  "
+                  f"SE={se/n:.5f}  NE={ne/n:.4f}  top1={correct/n:.3f}  [{el:.0f}s]")
+
+    # ---------- final eval ----------
+    correct = 0
+    for i in range(n):
+        _, _, _, _, _, scores = forward_choice(i, opt_tok[i])
+        if int(np.argmax(scores)) == labels[i]:
+            correct += 1
+    print(f"FINAL train top-1: {correct/n:.3f}")
+
+    # ---------- temperature calibration on train confidence ----------
+    # scale T so mean top-prob lands near the observed accuracy (calibrated)
+    gaps = []
+    for i in range(0, n, 4):
+        _, _, _, _, _, scores = forward_choice(i, opt_tok[i])
+        srt = np.sort(scores)[::-1]
+        gaps.append(srt[0] - srt[1])
+    mean_gap = float(np.mean(gaps)) if gaps else 1.0
+    temperature = max(0.25, min(4.0, mean_gap))
+    print(f"calibrated temperature: {temperature:.3f}")
+
+    os.makedirs(os.path.dirname(args.out), exist_ok=True)
+    weights = {
+        "emb": emb.tolist(), "wq": wq.tolist(), "bq": bq.tolist(),
+        "w": w.tolist(), "b": b.tolist(),
+        "ws": ws.tolist(), "bs": bs.tolist(),
+        "temperature": temperature, "dim": D,
+    }
+    with open(args.out, "w") as f:
+        json.dump(weights, f)
+    print(f"wrote {args.out} ({os.path.getsize(args.out)/1024:.0f} KB)")
+
+if __name__ == "__main__":
+    main()

package/dist/cli/tui.d.ts

@@ -1,6 +0,0 @@
-#!/usr/bin/env node
-/**
- * A3M Router Quick TUI
- * Zero-config experience - no API key needed for free tier
- */
-export {};

package/dist/cli/tui.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"tui.js","sourceRoot":"","sources":["../../src/cli/tui.ts"],"names":[],"mappings":";;AACA;;;GAGG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAEH,mDAAqC;AAErC,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;AAEnC,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;IACtB,OAAO,CAAC,GAAG,CAAC;;;;;;;;;;;;;;CAcb,CAAC,CAAC;IACD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;AAClB,CAAC;AAED,MAAM,OAAO,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;AAExB,oCAAoC;AACpC,IAAI,OAAO,KAAK,QAAQ,EAAE,CAAC;IACzB,OAAO,CAAC,GAAG,CAAC;;;;;;;;;;CAUb,CAAC,CAAC;IAED,MAAM,EAAE,GAAG,QAAQ,CAAC,eAAe,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC;IAEtF,MAAM,WAAW,GAAG;QAClB,EAAE,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,yBAAyB,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE;QACnF,EAAE,QAAQ,EAAE,UAAU,EAAE,KAAK,EAAE,kBAAkB,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE;QACjF,EAAE,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE;QAC3E,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE;KAC1E,CAAC;IAEF,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,SAAS,GAAG;QACV,EAAE,CAAC,QAAQ,CAAC,cAAc,EAAE,CAAC,CAAS,EAAE,EAAE;YACxC,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE,KAAK,MAAM,EAAE,CAAC;gBACrC,OAAO,CAAC,GAAG,CAAC,sFAAsF,CAAC,CAAC;gBACpG,EAAE,CAAC,KAAK,EAAE,CAAC;gBACX,OAAO;YACT,CAAC;YACD,MAAM,CAAC,GAAG,WAAW,CAAC,KAAK,EAAE,GAAG,WAAW,CAAC,MAAM,CAAC,CAAC;YACpD,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC,CAAC,QAAQ,MAAM,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC;YACvD,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC;YAChD,GAAG,EAAE,CAAC;QACR,CAAC,CAAC,CAAC;IACL,CAAC;IACD,GAAG,EAAE,CAAC;IACN,OAAO;AACT,CAAC;AAED,aAAa;AACb,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;IAC1B,OAAO,CAAC,GAAG,CAAC;;;;;;;;;;;;CAYb,CAAC,CAAC;IAED,IAAI,CAAC;QACH,8DAA8D;QAC9D,MAAM,MAAM,GAAG,OAAO,CAAC,0BAA0B,CAAC,CAAC;QACnD,MAAM,IAAI,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,IAAI,CAAC;QAC3C,MAAM,CAAC,iBAAiB,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC;QACnC,OAAO,CAAC,GAAG,CAAC,0CAA0C,IAAI,EAAE,CAAC,CAAC;IAChE,CAAC;IAAC,OAAO,CAAU,EAAE,CAAC;QACpB,MAAM,GAAG,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;QACvD,OAAO,CAAC,GAAG,CAAC,uDAAuD,CAAC,CAAC;QACrE,OAAO,CAAC,KAAK,CAAC,QAAQ,EAAE,GAAG,CAAC,CAAC;IAC/B,CAAC;IACD,OAAO;AACT,CAAC;AAED,aAAa;AACb,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;IAC1B,8DAA8D;IAC9D,MAAM,MAAM,GAAG,OAAO,CAAC,kBAAkB,CAAC,CAAC;IAC3C,MAAM,CAAC,SAAS,EAAE,CAAC;IACnB,OAAO;AACT,CAAC;AAED,mDAAmD;AACnD,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;AAE7B,2CAA2C;AAC3C,IAAI,OAAO,CAAC,GAAG,CAAC,YAAY,EAAE,CAAC;IAC7B,OAAO,CAAC,GAAG,CAAC,iBAAiB,GAAG,KAAK,GAAG,GAAG,CAAC,CAAC;IAC7C,OAAO,CAAC,GAAG,CAAC,gCAAgC,CAAC,CAAC;IAE9C,uBAAuB;IACvB,8DAA8D;IAC9D,MAAM,EAAE,UAAU,EAAE,GAAG,OAAO,CAAC,aAAa,CAAC,CAAC;IAC9C,UAAU,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAiE,EAAE,EAAE;QAC3F,OAAO,CAAC,GAAG,CAAC,kBAAkB,CAAC,CAAC,QAAQ,IAAI,MAAM,EAAE,CAAC,CAAC;QACtD,OAAO,CAAC,GAAG,CAAC,eAAe,CAAC,CAAC,KAAK,IAAI,yBAAyB,EAAE,CAAC,CAAC;QACnE,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,CAAC,cAAc,IAAI,MAAM,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACvE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAU,EAAE,EAAE;QACtB,MAAM,GAAG,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;QACvD,OAAO,CAAC,GAAG,CAAC,cAAc,EAAE,GAAG,CAAC,CAAC;QACjC,OAAO,CAAC,GAAG,CAAC,mDAAmD,CAAC,CAAC;IACnE,CAAC,CAAC,CAAC;IACH,OAAO;AACT,CAAC;AAED,oCAAoC;AACpC,OAAO,CAAC,GAAG,CAAC;;;;;;;;;;;;;;;;;;;;;;;;CAwBX,CAAC,CAAC"}

package/dist/routing/shadowSampler.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"shadowSampler.d.ts","sourceRoot":"","sources":["../../src/routing/shadowSampler.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,EAAc,oBAAoB,EAAE,aAAa,EAAE,MAAM,kBAAkB,CAAC;AAInF,MAAM,WAAW,mBAAmB;IAClC;;;;;OAKG;IACH,aAAa,CAAC,EAAE,MAAM,CAAC;IAEvB;;;;OAIG;IACH,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAE9B;;;;;;OAMG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAE3B;;;;;OAKG;IACH,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAE3B;;;OAGG;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;IAExB;;OAEG;IACH,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAE9B;;;;OAIG;IACH,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAE9B;;;OAGG;IACH,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,cAAe,SAAQ,aAAa;IACnD,2DAA2D;IAC3D,SAAS,EAAE,OAAO,CAAC;IACnB,sDAAsD;IACtD,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,gDAAgD;IAChD,iBAAiB,EAAE,MAAM,CAAC;IAC1B,wDAAwD;IACxD,QAAQ,EAAE,UAAU,CAAC,OAAO,oBAAoB,CAAC,CAAC;IAClD,4CAA4C;IAC5C,eAAe,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,MAAM,EAAE,SAAS,GAAG,QAAQ,GAAG,KAAK,GAAG,WAAW,CAAC;IACnD,eAAe,EAAE,MAAM,CAAC;CACzB;AAED;;;;;;;;;;;;;GAaG;AACH,qBAAa,aAAa;IACxB,OAAO,CAAC,MAAM,CAAgG;IAG9G,OAAO,CAAC,eAAe,CAAuB;IAG9C,OAAO,CAAC,YAAY,CAAK;IACzB,OAAO,CAAC,WAAW,CAAK;IAExB,YAAY,MAAM,GAAE,mBAAwB,EAW3C;IAED;;;OAGG;IACH,OAAO,CAAC,oBAAoB;IAiC5B;;;;;;;;;;OAUG;IACH,OAAO,CAAC,kBAAkB;IAoB1B;;;;;;;;;;;OAWG;IACH,OAAO,CAAC,wBAAwB;IA4DhC;;;;;;OAMG;IACH,eAAe,CACb,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE;QAAE,gBAAgB,CAAC,EAAE,MAAM,EAAE,CAAC;QAAC,iBAAiB,CAAC,EAAE,MAAM,CAAA;KAAE,GACpE,cAAc,CA+BhB;IAED;;;;;;;OAOG;IACG,cAAc,CAClB,aAAa,EAAE,MAAM,EACrB,YAAY,EAAE,MAAM,GAAG,IAAI,GAC1B,OAAO,CAAC,gBAAgB,CAAC,CAgD3B;IAED;;OAEG;IACH,QAAQ,IAAI;QAAE,WAAW,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAA;KAAE,CAM1E;IAED;;OAEG;IACH,UAAU,IAAI,IAAI,CAGjB;IAED;;OAEG;IACH,SAAS,CAAC,MAAM,EAAE,OAAO,CAAC,mBAAmB,CAAC,GAAG,IAAI,CAEpD;CACF"}