{
 "artifacts": null,
 "category": "architecture",
 "description": "Build a sparse expert cascade whose router uses empirical progress, uncertainty, and compute cost to construct an efficient frontier of experts. Instead of always choosing the expert with the largest immediate gain per FLOP, route different confidence states through an ordered sequence of frontier experts, allowing cheap high-variance experts early and safer or more cost-effective experts near the final decision.",
 "formulas_latex": [
  "$$\\mathsf{a}_i=\\frac{2\\nu_i}{s_i^2},\\qquad CE_i=\\frac{\\nu_i}{c_i}.$$",
  "$$\\mathsf{a}(i_1^*)\u003c\\mathsf{a}(i_2^*)\u003c\\cdots\u003c\\mathsf{a}(i_{\\tilde K}^*),\\qquad CE(i_1^*)\u003cCE(i_2^*)\u003c\\cdots\u003cCE(i_{\\tilde K}^*).$$",
  "$$\\zeta_{myo}=\\{(\\mathsf{n},[\\underline{\\mathcal B},\\overline{\\mathcal B}))\\},\\qquad \\mathsf{n}=\\arg\\max_i CE_i.$$",
  "$$\\hat x_{t+1}=\\hat x_t+\\Delta_i(x_t),\\qquad \\hat\\nu_i=\\mathbb E[\\Delta_i],\\qquad \\hat s_i^2=\\mathbb E[(\\Delta_i-\\hat\\nu_i)^2],\\qquad \\hat{\\mathsf a}_i=\\frac{2\\hat\\nu_i}{\\hat s_i^2+\\epsilon},\\quad \\widehat{CE}_i=\\frac{\\hat\\nu_i}{c_i}.$$"
 ],
 "id": 1495,
 "implementation": "Integrate the method at the inference-time router of a sparse MLP mixture-of-experts or a cascade of classifiers. Let x be a scalar confidence state, preferably the top-1 logit margin, negative entropy, or a calibrated correctness probability. Each expert i receives the current representation and returns an output plus a new confidence state; define Delta_i=x_after_i-x_before. First collect a calibration set and, for every expert and several bins of x, estimate nu_i as the mean of Delta_i, s_i^2 as the variance of Delta_i, and c_i as measured wall-clock latency or FLOPs. Compute a_i=2 nu_i/(s_i^2+epsilon) and CE_i=nu_i/c_i. Remove every expert dominated by another expert with both larger a and larger CE, then sort the remaining experts by a and retain a monotone subsequence in CE. Initialize thresholds S_0\u003c...\u003cS_K over the confidence interval, either as equal quantiles of calibration states or by dynamic programming on a discretized Markov chain whose transition samples are the measured Delta_i. At inference, use the expert assigned to the current interval [S_{j-1},S_j); update x; stop when x exceeds the target confidence boundary, otherwise continue to the next interval. Include a fallback to the highest-accuracy expert if x decreases repeatedly. Pseudocode: estimate statistics; prune frontier; initialize thresholds; for each input set x=x0; while x\u003cT: choose i=router_interval(x); y_i=expert_i(h); x=x+confidence_delta(y_i); h=representation_update(y_i); return y. The paper contributes the statistic definitions, frontier pruning principle, and ordered policy; threshold locations and expert transitions must be estimated empirically because the neural confidence process is not an exact diffusion. Start with a 4-expert MLP cascade on CIFAR-10 or a small language-model MoE on WikiText, comparing a dense strongest expert, an always-best-CE expert, a standard learned top-1 router, and frontier routing. Measure accuracy at fixed average FLOPs, latency at fixed accuracy, number of experts invoked, and calibration. The expected signal is lower average compute than the learned router at matched accuracy, or higher accuracy than the myopic CE router at matched compute; invalidate the idea if frontier routing does not improve the Pareto curve after threshold tuning.",
 "math_summary": "The paper defines each control i by drift nu_i, volatility s_i\u003e0, and cost c_i\u003e0. Its effective-drift statistic is \\(\\mathsf{a}_i=2\\nu_i/s_i^2\\), where larger \\(\\mathsf{a}_i\\) means progress is safer relative to stochastic fluctuation. Its cost-effectiveness statistic is \\(CE_i=\\nu_i/c_i\\), the expected progress per unit cost. The paper's free-boundary policy uses a sequence of selected controls \\(i_1^*,\\ldots,i_{\\tilde K}^*\\) satisfying \\(\\mathsf{a}(i_1^*)\u003c\\cdots\u003c\\mathsf{a}(i_{\\tilde K}^*)\\) and \\(CE(i_1^*)\u003c\\cdots\u003cCE(i_{\\tilde K}^*)\\), with state intervals \\([S^*_{j-1},S^*_j)\\) assigned to control \\(i_j^*\\). We adapt the state to a neural-network confidence or verification margin x. For expert i, \\(\\nu_i\\) is its mean increase in x, \\(s_i^2\\) is the variance of that increase, and \\(c_i\\) is measured latency or FLOPs. The frontier is the nondominated subset in the two-dimensional coordinates \\((\\mathsf{a}_i,CE_i)\\); dominated experts have no reason to be selected under this proxy. The paper contrasts this with the myopic policy \\(\\zeta_{myo}=\\{(\\mathsf{n},[\\underline{\\mathcal B},\\overline{\\mathcal B}))\\}\\), where \\(\\mathsf{n}=\\arg\\max_i CE_i\\). The experiment tests whether state-dependent switching beats this single-expert baseline at equal expected compute.",
 "math_tags": [
  "stochastic-processes",
  "probability",
  "optimization",
  "geometry"
 ],
 "ml_areas": [
  "moe-routing",
  "inference-speedup",
  "mlp"
 ],
 "paper": {
  "arxiv_id": "2607.27563",
  "arxiv_url": "https://arxiv.org/abs/2607.27563",
  "summary_what_math_gives_to_ml": "The paper reduces a multi-action stochastic control problem to two sufficient statistics per action: effective drift, measuring progress relative to diffusion risk, and cost-effectiveness, measuring progress per unit cost. Its key transferable structure is an efficient frontier: most actions can be discarded, and the remaining actions should be deployed in an ordered progression rather than selected by a purely myopic utility score. This suggests a principled adaptive-compute or mixture-of-experts controller in which experts are ranked by empirical improvement, uncertainty, and inference cost. The strongest initial test is a confidence-driven expert cascade with frontier pruning and state-dependent switching thresholds.",
  "title": "Strategies for Milestone-driven Start-ups in Multi-activity Settings",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 6,
  "usefulness": 6
 },
 "solves": [
  "speedup",
  "accuracy",
  "scalability"
 ],
 "title": "Risk-cost efficient frontier routing",
 "url": "https://synthcore.org/idea/1495/risk-cost-efficient-frontier-routing",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)",
  "verification_axes": {
   "benchmark_mechanism": {
    "confirmed": null,
    "tested": false
   },
   "practical_benchmark": {
    "beats_baseline": null,
    "tested": false
   },
   "toy_mechanism_gate": {
    "confirmed": null,
    "tested": false
   }
  }
 }
}