Rule-based systems increasingly issue recommendations in welfare administration, encoding eligibility criteria but omitting the field-level context a reviewer observes. A live deployment of one such system saw officers depart from its recommendations in roughly half of all formally-eligible cases, but field data cannot say what those departures were, or why. In a pre-registered experiment (N = 120; 1,440 decisions), participants reviewed welfare-eligibility cases in which the system's recommendation, the presence of field context, and whether that context reinforced or contradicted the recommendation were crossed independently.

Contextual information raised override by roughly 10 percentage points on aggregate, but that figure is a design artefact: it is a weighted average of two opposite-signed effects whose value depends on the mix of confirming and conflicting cases in the stimulus pool, and it reverses sign under compositions no reviewer would question. The robust, identified result is structural: when field context contradicts a recommendation to approve, override rises from 3% to 53% (interaction OR = 9.48). Departure, moreover, is not one behaviour because reviewers could reverse the recommendation or decline to resolve the case, "override" combines distinct acts, and under evidentiary conflict the dominant departure is escalation (a refusal to resolve) rather than reversal. Reliance on algorithmic advice is therefore not a scalar disposition to trust or distrust; it is conditioned on the alignment between recommendation and evidence.

This paper analyses a consistent divergence between rule-based outputs and operator decisions observed during a field pilot of SARAL, a human-in-the-loop algorithmic triage system deployed for state welfare allocation in Maharashtra, India. Across an analytic cohort of 260 cases where the rule engine returned an eligible-by-rule output, operators departed from the system; rejecting, escalating, or requesting documents, in 52% of cases.

By inductively coding the unstructured field notes generated during the pilot, the paper identifies a taxonomy of "unencoded signals" that operators use where the algorithm compresses multidimensional vulnerability into binary variables: visual asset heuristics, geographic and household proximity proxies, normative cues, and administrative implementation friction. The contribution is descriptive and structural, the analysis documents the divergence and its recurring categories rather than making causal claims, and it frames the pattern as a sociotechnical gap between formal rule representation and situated frontline judgment. The experimental follow-up that isolates these effects causally is reported in the Phase 2 paper above.