Every question we asked, and what the answer cost.
Dossier runs deep research across several backends at once, keeps the receipts, and publishes them. This is the whole corpus — including the runs that failed and the money they cost anyway.
This collection is what one person and one laptop produced between — and — — about — runs a day. It is heavily technical: infrastructure, agent tooling, hardware and developer economics. It is not a survey of any field, it was never intended to be balanced, and every question in it was asked because someone had a decision to make that week.
— of these questions were put to several backends at once. Those are the interesting ones: you can read where the models disagreed instead of taking one of them at its word.
Agreement is not corroboration. Where a claim is supported, support is counted in independent registrable domains, never in how many backends said the same thing.
The library
One row per question. The cost is what Dossier committed for it, and a $0.00 row is a question answered on a subscription that was already being paid for — not a free lunch.
Dossier's own committed and settled figures. Your provider's console is the billing record and may differ. — questions are held back from this list: some are client work or this business's own internals, and some make critical claims about a named third party that were never checked first-party — this engine has taken vendor facts from aggregators and got them wrong before. All of them are still counted in every total above, because a total that quietly drops rows is not a total.
What did not work
— of — runs failed and — were cancelled — — of everything attempted. A cancelled run keeps the estimate already committed; it is not refunded. These are in the corpus the search reads, so you can ask what breaks.
How a question becomes a report
Plan it, free
Dossier engineers the brief, picks the archetype, and returns a cost band and a contract fingerprint. Nothing has spent yet. Every band on this page came from here.
Run the panel
The same brief goes to several backends at once — the API ones that bill, and the CLI subscriptions you already hold, which do not. It returns a handle immediately and keeps going for anywhere from four to sixty minutes.
Read it, then check it
Reports come back with their sources. Dereference every cited URL, profile the source mix, and diff what the backends claimed against each other before anyone acts on it.
A four-to-sixty-minute job that reports a percentage is lying about what it knows. Progress here is elapsed time and a cost band, and a member that dies says why.
Two ways to run it
Dossier drives the CLI subscriptions you already pay for. — of the runs in this library cost nothing this way. You supply the provider keys; the budget gate is yours.
Read what it producedA subscription buys cloud execution and nothing else: the same engine, running when your laptop is shut. Not yet on sale, and this page will not quote a price it cannot honour.
Same house, one engine, different jobs
dossier.
Deep research your AI can be held to. Runs the panel, keeps the budget, dereferences the citations, and puts the report at a shareable URL. This page is its output.
Browse the corpusmargin‸
Living team pages. Where a report like these stops being a document and becomes something a team argues with — comment by typing or talking, and an agent that applies the feedback and commits it. Margin runs Dossier underneath for its own research.
See Margin