https://ifitsmanu.com/ https://ifitsmanu.com/manu-bhardwaj-headshot.jpg Manu Bhardwaj Manu Bhardwaj. AI engineer and quant. Working AI from the silicon up: chips, kernels, training, runtime, emergent behavior. NYC. https://ifitsmanu.com/papers/serving-frontier/ https://ifitsmanu.com/papers/serving-frontier/figures/cover.png Disaggregated or Colocated? The Cost-Frontier of LLM Serving Under SLO Contracts. First page of Research Paper #1. Disaggregated or Colocated? The Cost-Frontier of LLM Serving Under SLO Contracts. https://ifitsmanu.com/papers/verifier-composition/ https://ifitsmanu.com/papers/verifier-composition/figures/cover.png Calibration Drift Under Verifier Composition. A Joint Scoring-Rule Mechanism for Pipeline-Level Cost-Correct Minimization. First page of the verifier-composition paper. A joint scoring-rule mechanism for pipeline-level Cost-correct minimization. https://ifitsmanu.com/papers/inference-frontier/ https://ifitsmanu.com/papers/inference-frontier/figures/cover.png The Inference-Time Compute Frontier. A Cost-Correct Threshold for Training Versus Test-Time Allocation. First page of the inference-frontier paper. A closed-form threshold for training versus test-time allocation. https://ifitsmanu.com/papers/routing-premium/ https://ifitsmanu.com/papers/routing-premium/figures/cover.png The Routing Premium. An Economic Threshold for Difficulty-Conditional Inference Compute. First page of Research Paper #3. An economic threshold for difficulty-conditional inference compute. https://ifitsmanu.com/papers/routing-premium/figures/fig1_isoclines.png The Routing Premium. An Economic Threshold for Difficulty-Conditional Inference Compute. Isoclines of the routing-premium threshold κ·Δ = γ across classifier calibration κ and workload heterogeneity Δ. Operating points for six deployed systems plotted; every point sits on the positive side of the threshold. https://ifitsmanu.com/papers/routing-premium/figures/fig2_calibration.png The Routing Premium. An Economic Threshold for Difficulty-Conditional Inference Compute. Classifier-calibration sensitivity. The frontier of operating points closest to the threshold shows which systems would flip to negative routing premium under modest disclosure error. https://ifitsmanu.com/papers/routing-premium/figures/fig3_composition.png The Routing Premium. An Economic Threshold for Difficulty-Conditional Inference Compute. Multiplicative composition of the routing-premium threshold with the channel-allocation threshold of Paper #2 (inference-frontier). Frontier reasoning tiers in 2026 satisfy both conditions simultaneously. https://ifitsmanu.com/papers/verifier-procurement/ https://ifitsmanu.com/papers/verifier-procurement/figures/cover.png Verifier Procurement Under Unobservable Quality. A Scoring-Rule Mechanism for Cost-Correct Minimization. First page of the verifier-procurement paper. A scoring-rule mechanism for Cost-correct minimization under unobservable quality. https://ifitsmanu.com/papers/verifier-as-curriculum/ https://ifitsmanu.com/covers/verifier-as-curriculum.png The Verifier as Curriculum. VHG and the Third Role. First page of Field Note #5. The Verifier as Curriculum. https://ifitsmanu.com/papers/the-structural-residual-ceiling/ https://ifitsmanu.com/covers/the-structural-residual-ceiling.png The Structural Residual Ceiling. AI Pre-Decoders for the Surface Code. First page of Field Note #4. AI pre-decoders for the surface code. https://ifitsmanu.com/papers/the-alpha-asymmetry/ https://ifitsmanu.com/covers/the-alpha-asymmetry.png The Alpha Asymmetry. Why Verifiers Can Be Smaller Than Generators. First page of Field Note #3. Why verifiers can be smaller than generators. https://ifitsmanu.com/papers/the-cost-of-being-right/ https://ifitsmanu.com/covers/the-cost-of-being-right.png The Cost of Being Right. Verification Economics in 2026. First page of Field Note #2. Verification economics in 2026. https://ifitsmanu.com/figure_2_pareto_cost_quality.svg The Cost of Being Right. Verification Economics in 2026. ARC-AGI-2 cost-per-task vs accuracy across frontier configurations, December 2025 results analysis. Roughly two orders of magnitude of cost dispersion at near-equivalent accuracy. The cost axis is logarithmic. https://ifitsmanu.com/figure_2_pareto_cost_quality.png The Cost of Being Right. Verification Economics in 2026. ARC-AGI-2 cost-per-task vs accuracy across frontier configurations, December 2025 results analysis. Raster export of the canonical SVG for downstream tooling that does not render SVG. https://ifitsmanu.com/papers/the-inference-stack-2026/ https://ifitsmanu.com/covers/the-inference-stack-2026.png The Inference Stack in 2026. First page of Field Note #1. The inference stack in 2026. https://ifitsmanu.com/figure_1_public_api_price_envelope.svg The Inference Stack in 2026. Selected public API prices, 2023 to 2026, plotted as a 1:1 blended cost per million tokens on a log scale. Roughly two orders of magnitude of compression across the model class envelope.