Six steps, and one of them is a guard.
Most research tools stop at step four. The reason a brief from here can be acted on is step five: the citations are checked against the pages that were really retrieved, and what fails the check is removed rather than repaired.
You ask a question
A real question, not a keyword. Optionally narrow it: only sources published in the last N days, only these domains, or everything except those. Filters are applied upstream at the search, not as an afterthought on the results.
The question becomes angles
Above the quick rung, one cheap model call decomposes the question into distinct search angles — sub-questions, opposing views, primary sources — and the original question always stays in the mix, so a decomposition that drifts still gets checked against what you actually asked.
Pages are fetched and read
Each angle runs a live web search that returns page text, not just links. Duplicates across angles collapse to the fullest read of each page, and the reading budget for the depth you chose decides how much of each one is actually taken in.
The brief is written with numbers attached
The sources go to the synthesis as a numbered list, and it is asked for claims that each carry the numbers they came from. Nothing about this step is trusted — it is arranged so the next step's job is small.
Every citation is checked
Each number is looked up in the set of pages actually fetched. Numbers that point at nothing are discarded. Claims left holding only discarded numbers are discarded too. Survivors are renumbered so the first citation you are asked to check is [1].
You get a brief, or your credits back
If anything survived, you get key findings, prose, numbered sources with verbatim quotes, a confidence figure, the open questions, and the counts of what the guard removed. If nothing survived, the brief is marked thin and every credit is refunded.
Step five, in detail
Three claims arrive from the synthesis. Watch what reaches the page and, more importantly, what does not.
What each depth actually does
The numbers below are read straight out of the configuration the runner obeys, so this table cannot promise work the run does not do.
| Quick | Standard | Deep | |
|---|---|---|---|
| Price | 15 cr | 35 cr | 75 cr |
| Search angles | 1 | 3 | 6 |
| Pages requested | 6 | 24 | 120 |
| Sources kept | up to 6 | up to 14 | up to 24 |
| Characters read per source | 1,200 | 3,000 | 6,000 |
| Question decomposed first | no — searched verbatim | yes | yes |
| Second pass over its own gaps | no | no | yes |
| Typically takes | ~25s | ~60s | ~150s |
Try it on a question you already know the answer to.
It is the fastest way to find out whether a research tool is worth trusting with the questions you do not.