Replication package — The Institutional Window: How Contract Law Bounds Liability Signaling of Human Fallback Capability under Generative AI — (v1.3.1.1)

Abstract

Replication package for "The Institutional Window: How Contract Law Bounds Liability Signaling of Human Fallback Capability under Generative AI" (Bauer 2026). The paper asks when a liability commitment can still certify a provider's preserved human fallback capability once generative AI makes the output itself uninformative. Four legal primitives map a posted cap and agreed-damages term into retained exposure: litigation viability, the penalty doctrine, displacement to the state or an indemnity pool, and mandatory limits on contractual caps. Where the positive image is connected and sub-threshold commitments remain admissible, retained exposure has the message set {0} u [F, C]: low types pool at zero, intermediate types separate on a schedule anchored at F, high types may pool at the ceiling. Where law removes the zero-exposure region, separation begins at the bottom type. Contents. The model core and the exact message-topology construction; occupation and jurisdiction calibration vectors with their legal and empirical provenance; the claim-robustness engine over 88 occupation and 46 jurisdiction perturbations; a Saltelli global sensitivity analysis; figure code; the interactive occupation-by-jurisdiction simulator with its model core and a JavaScript/Python cross-check; the manuscript and the mathematical companion with their LaTeX sources; and a script-to-table map (SCRIPT_TO_TABLE.md) naming the producer of every reported object. Verification. The acceptance harness (code/verify_all.py) re-derives the full grid of five occupations by twelve configurations and exits 0 only if every reported quantity reproduces: equilibrium and boundary conditions including the attained-boundary tie break, the separation of full, interior, ceiling-pooling and structural-zero regimes, the Austrian band openings, robustness coverage reported claim-by-claim as 187 of 200 claim-cells, the thirteen English-rule cells left unsupported because a hole meets the candidate path and the three proved separately, the disclosed tipping values of the prevailing probability, and the sensitivity block against both its base seed and a second seed. The simulator's JavaScript core is an independent implementation and reproduces the package on 65 quantities: the separating range of all sixty central cells and h_min for the five occupations (sim/crosscheck_js.mjs). MANIFEST.sha256 covers the deposit; code/MANIFEST.sha256 covers the thirteen code files. Requirements: Python 3.10+ with NumPy; Matplotlib to regenerate figures; Node 18+ for the cross-check; pdflatex for the manuscript. Randomness is seeded throughout (Saltelli base seed 20260826, second seed 7). Scope. Parameters are placed ordinally against published evidence; they are not estimated. Jurisdiction vectors are stylized benchmarks, not measured regimes. Figures 4, 5 and 6 are shipped as PDFs in tex/ but are not regenerated by this package. Version 1.3.1 supersedes 1.2.0. The prevailing probability in enforcement participation is now an exposed module parameter (default 0.50, central numbers unchanged), disclosed in Appendix B and stress-tested as its own robustness family with tipping values reported; the sensitivity paragraph is restated with seed bands rather than a point ranking; the Austrian case authority is corrected; and the simulator was brought onto the current construction. The code in this deposit is released under the MIT licence. The manuscript and companion PDFs and their LaTeX sources are included for convenience and remain under the licence of the corresponding arXiv posting.