{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,19]],"date-time":"2026-10-19T15:40:27Z","timestamp":1792424427510,"version":"4.5.0"},"posted":{"date-parts":[[2026]]},"group-title":"SSRN","reference-count":0,"publisher":"Elsevier BV","license":[{"start":{"date-parts":[[2026,1,1]],"date-time":"2026-01-01T00:00:00Z","timestamp":1767225600000},"content-version":"unspecified","delay-in-days":0,"URL":"https:\/\/www.uspto.gov\/ip-policy\/copyright-policy\/copyright-basics"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<jats:p>When a constrained optimizer selects a portfolio, managers, auditors and regulators ask why. Large language models (LLMs) are increasingly used to answer in plain language, but a fluent explanation can cite reasons that did not drive the decision. In a linear program, the effect of every constraint on the optimum is known from the solver's certificate and can be confirmed by relaxing the constraint and resolving. Each claim in an explanation (\"this limit drove the allocation and is worth about x basis points\") can therefore be checked against ground truth rather than against another model or a human judge. We apply this test to 196 bond-portfolio problems built from public FRED data on US Treasury yields and ICE BofA credit spreads (May to October 2026), scoring explanations written by Claude Sonnet 4.5 and Gemini 2.5 Flash without access to the certificate. Across 149 cases in which a limit had exactly zero effect on the optimum, the models named it as a driver in 89% (Claude) and 77% (Gemini) of cases, while naming binding limits in 98% and 90% of cases. Giving the models the portfolio facts that show which limits are tight left Claude's rate essentially unchanged (86%) and reduced Gemini's to 48%, no better than a fixed template that always names the same two limits (50%). The models' estimates of what each binding limit costs were within \u00b135% of the true value in 7% of 1,703 claims, with median errors ranging from 0.29\u00d7 to 8.43\u00d7 the true value and in the same direction for both vendors, even though the holdings shown to the models determine the true values. A third model, GPT-4o-mini, copied the numbers in the prompt's format example and is reported separately. Explanations generated from the certificate avoid both errors by construction.<\/jats:p>","DOI":"10.2139\/ssrn.7625878","type":"posted-content","created":{"date-parts":[[2026,10,19]],"date-time":"2026-10-19T14:59:23Z","timestamp":1792421963000},"source":"Crossref","is-referenced-by-count":0,"title":["Faithful by Construction? An Exact Test of LLM Explanations for Constrained Portfolio Optimization Decisions"],"prefix":"10.2139","author":[{"ORCID":"https:\/\/orcid.org\/0009-0000-0359-3268","authenticated-orcid":true,"given":"Prateek","family":"Bhatnagar","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,10,19]],"date-time":"2026-10-19T14:59:23Z","timestamp":1792421963000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7625878"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7625878","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}