Work

The Comparator Is Part of the Discovery: Counterfactual Comparator Equalization and Cross-Representation Falsification in Autonomous Engineering Search

Year: 2026
Type: preprint
Abstract: Autonomous and AI-assisted engineering search frameworks routinely optimize high-dimensional physical designs — spanning 3D chip-package microfluidics, power delivery networks, and orbital computing infrastructure — against a fixed baseline model. We show that when an optimizer can vary both its proposed architectural mechanism and secondary baseline parameters (contact geometry, supply-chain cost tables, yield assumptions, or physical representation fidelity), the resulting performance ratio mixes true structural advantage with baseline mismatch. We formalize Counterfactual Comparator Equalization (CCE), a mathematical decomposition of any composite performance gain G(a, b, c) = M(a, b, c)/M(a0, b0, c0) into baseline-defect remediation Δgeom, auxiliary assumption transfer Δaux, and isolated mechanism contribution Δmech, coupled with Exogenous Cost-Floor Feasibility Bounds and Cross-Representation Physical Plausibility Oracles (CRPPO). Evaluating three multi-physics engineering campaigns, we demonstrate that: (1) a 1.9672× (3.0571 vs. 1.5540 TFLOPS/\$) chip-package co-design claim collapses to +0.332% (1.873302 vs. 1.8671 TFLOPS/\$) when contact pitch (0.75 mm), PDN resistance (0.380 mΩ), die/package costs, and yield (86.5%) are equalized, and becomes slightly negative (−2.66%, 1.8142 vs. 1.8637 TFLOPS/\$) under a realistic split-plane interposer architecture; (2) the 5-year IT electricity bill alone (\$4,380.00 at 1,000 W, \$0.10/kWh) caps any zero-hardware-cost packaging multiplier at 2.450× (4.5662 TFLOPS/\$), proving 10× packaging-only targets mathematically impossible; and (3) five unit-tested simulation artifacts that passed 100% of self-checks failed immediately under cross-representation checks (SPICE netlist dissipating 614.4 W vs. 50 W claim; 63.04 mm TO-247 footprint overflowing a 45.0 mm zone; STEP file with 139 points and 0 solid faces; hydraulic solver halving channel velocity 0.397 vs. 0.794 m/s; and single-layer 8 µm RDL routing only 343/1,750 = 19.6% of HBM3e pins). Evidence classification: computational and source-audit only (SIMULATION_AND_AUDIT_VERIFIED). No physical silicon interposer, ceramic hardware, cold plate, or orbital hardware was fabricated or measured. All figures and numbers are deterministically reproduced by reproduce_paper1_cce_crppo.py (standard-library Python 3, under 1 second), included with the figures in the reproducibility bundle. Reproducibility code in the bundle is Apache License 2.0; paper text is CC BY 4.0.
Language: English
Open Access status: green
OpenAlex ID: W7235399147
Cites:
Cited by:
Related to: 10