> ## Documentation Index > Fetch the complete documentation index at: https://docs.serviceplan-agents.com/llms.txt > Use this file to discover all available pages before exploring further. # Quality & Trust > How you can tell whether a coworker's answer is actually solid — not just a black box. A coworker's reply isn't just a paragraph of text you have to take on faith. Every deliverable is built to show its own work — where a number came from, how confident the coworker actually is in it, and what it thinks you should do about it. ## Why this matters You want to know two things before you act on anything a coworker sends you: is this actually true, and how much should I trust it? Rather than asking you to just believe a polished-looking answer, the platform builds the receipts directly into the deliverable itself. ## How it actually works Specific stats and claims in a research deliverable get a numbered marker like `[1]`, `[2]`, with a matching source list at the end. Every hard number traces back to a real, specific source — not a vague "studies show" or "research indicates." Key findings get tagged as **High**, **Medium**, or **Low** confidence, depending on how strong the underlying evidence is — multiple solid, recent sources versus a single older source, or something pieced together from general web results. You can tell at a glance how much weight a specific claim can carry. Every research deliverable ends with the coworker's own read on it — overall confidence, any real limitations in what it found, and a clear recommendation. Not just a pile of data with no interpretation attached. For a lot of research work, a second, independent pass re-checks specific claims against the original source material and flags anything that looks fabricated, weakly sourced, or misattributed — before it ever reaches you. This runs often, though exactly when it applies can vary by task type and channel, so treat it as a strong, frequent extra layer of scrutiny rather than an unconditional guarantee on every single task. For visual work specifically, there's a separate independent check that the numbers actually shown in a chart or dashboard match the real source data before anything ships — so a rounding error or a dropped row doesn't quietly make it into something you present to someone else. ## Concrete examples * A market-sizing figure in a research report shows up as "€2.4B market by 2027 \[3]" with source `[3]` spelled out at the bottom — not just stated as fact. * A competitive analysis flags one specific claim as **Low confidence** because it only found one, slightly dated source — while the rest of the report is High confidence — so you know exactly where to double-check before making a decision. * A dashboard Alex builds gets checked against your original spreadsheet before delivery, catching a case where a filtered-out row would otherwise have skewed a total. ## What this is not This isn't a guarantee that every answer is perfect, and it isn't a black-box "trust score." It's a set of real, visible habits — citing sources, labeling confidence, being upfront about limitations, and checking work independently — that let *you* judge how much weight to put on something, rather than asking you to just trust the polish. ## How this connects This trust layer applies for as long as a task runs — see [Tasks](/en/tasks/overview) for the full picture. It's part of what happens during [execution](/en/tasks/plan-checkpoint), and it's exactly what a [draft review](/en/tasks/draft-review) lets you check before anything is finalized. See [Hannah](/en/coworkers/hannah) for more on how research deliverables get built in the first place. If a number or claim ever looks off, just ask — a coworker can walk you through exactly where it came from and how confident it really is.