orchestrate-input-validation-and-overrideslisted

# Orchestrate: Input Validation and Deterministic Overrides **Source**: *The Engineer's Notebook*, "Getting better at HackerRank Orchestrate" (Shloka Shah) — names input validation *before* model calls as a distinct requirement from output validation *after* generation, and separately names "rule-based overrides for cases where model discretion shouldn't apply" as part of a reliable guardrail design. Reinforced by a first-hand #1-ranked participant case study (Medium, "How I went from 122 to 1"), which describes replacing a binary safety flag with a **conditional, confidence-gated** safety gate: a hard block fires only when the model *cannot cite specific evidence*, rather than automatically downgrading every verdict whenever a low-confidence signal appears — preventing legitimate findings from being silently overridden by noisy metadata. ## Two validation layers, not one `orchestrate-schema-guardrails` (already in this collection) covers validating model *output* against the expected schema. This skill covers the layer before that: validating *input* before it ever reaches a model call. **What input validation catches, specifically**: empty or missing required fields, duplicate IDs across the dataset, malformed file references (an image path that doesn't resolve, a corpus document ID that doesn't exist), and — per the Orchestrate dataset's known design — prompt injection attempts embedded in ticket/claim text. Catching these before a model call means you're not spending