When you ask Pilot5 one question, here is exactly what fires.

Before any AI model reasons, the companion interviews you, one exchange at a time: the decision, the situation, your constraints, the facts and figures, and what a useful answer looks like. It also infers what you did not spell out: perimeter, certainty, biases, and preferred output format. When it has enough, it asks whether to start and shows you the brief: the exact text the panel will receive. You pick the mode from three cards that open from the message box. None is recommended. The interview uses no credits, and from your AI assistant the same interview runs through the Pilot5 connector.

A moderation service screens the question and, in the web app, the attached documents and the brief; Pilot5's own rules screen the question too. Attempts to override the models' instructions are detected in eight languages and refused. A blocked request is refused and its credits come back. Credits are reserved from an estimate, and at the end the run is charged at actual use, never above the mode's cap; the unused part of the reservation is returned.

Your privacy settings then govern the whole run. In legal runs, every model call goes to provider endpoints that keep no data. With Zero Data Retention (Settings › Privacy), the question, the brief, the analyses and the result are kept out of Pilot5's database. A Private Session deletes its runs and conversations when you end the session, or automatically about an hour after you stop. Pilot5's model-routing account is set so that providers may not train on your prompts.

Before any model gives a view, Pilot5 gathers what the panel will work from, so the answer rests on retrieved sources and not only on what the models remember. Your documents come first: in the panel modes, documents up to about 100,000 characters are given whole, and longer ones are searched for their most relevant passages. Material imported from a connected tool arrives as a document, and your own connected systems (MCP servers) are queried during the run.

Around that, several systems run at once: a frame analyst reframes your question to surface the real decision behind what was literally asked; a persona gap analyzer asks each perspective what specific data it needs and sends targeted searches to fill those gaps; and a hybrid retrieval layer draws on 550+ data sources and integrations across 21 knowledge domains: 400+ verified institutional sources (IMF, EUR-Lex, PubMed, WHO, OECD, SEC, NICE, ClinicalTrials.gov, arXiv, Cochrane, Eurostat, INSEE, BOFIP, EPO and more) plus 160+ specialist API adapters for real-time queries. Stored legal texts are labelled by country, and official legal sources are fetched live, such as the enacted text of US provisions the panel cites. Three independent web search services run in parallel, and links that do not respond are removed from the source list.

A dedicated names model, run by Pilot5, finds the names of people and organisations and ID numbers in the material you supplied, and replaces each one with a placeholder such as [PERSONNE_1] or [SOCIETE_1]; the same name always gets the same placeholder. In the panel modes this covers your documents, the pasted context and conversation, saved memories, connected tools and earlier results, and your typed question is passed as written. In The Expert it covers the question too. The table linking placeholders to names stays with Pilot5 and puts the names back at the end. If masking cannot finish, the run stops before any model receives your material, and your credits come back.

Then five seats are filled, each with its own role, so the analyses do not overlap. Each seat gets the model ranked best for that role and that type of question, with one provider family per seat; in legal runs, only models with a zero-retention endpoint can be seated. One seat also acts as the arbiter, chosen by the question's domain.

• The Architect reasons about structure, system dependencies, benchmarks, and operational rigor.
• The Strategist reads macro trends, market signals, competitive dynamics, and timing.
• The Engineer examines technical feasibility, math and code accuracy, and execution constraints.
• The Counsel surfaces ethics, nuanced risk, regulatory exposure, and clarity of obligation.
• The Contrarian argues against your own position, objection by objection, and says what would answer each. When you state none, it runs a premortem on its own best answer.

The five models answer separately and at the same time, so no model sees another's view before forming its own. Each receives the brief, the evidence file and the sources most relevant to its role, and Pilot5 records a fingerprint of the shared input, so the run shows that all five started from the same material. Each writes its position, its claims with their sources, its risks and its confidence. Five complete analyses. Five distinct angles. Not five versions of the same answer. The Round 1 checks then run on the five analyses (Step 9).

Each analysis is reduced to a structured summary and labelled only as ANALYSIS_1 through ANALYSIS_5, in shuffled order, so no model can tell which provider or role wrote which. Each model critiques all five, with the results of the Round 1 checks in hand, and attacks ideas on merit, not on author.

After Round 2, Pilot5 decides whether the debate needs another round before the verdict. The most common is a duel: on up to two sharp disagreements, two models from different providers argue it out, and a referee rules only when it reaches the same ruling with the arguments in both orders; otherwise the point is marked unresolved. When the panel converges suspiciously fast, a Devil's Advocate round stress-tests the consensus before it hardens; focused research and calibration rounds can be added too. Premature consensus is treated as a failure signal, not a green light. Positions that lose the debate are kept and passed to the arbiter as the dissent. A run has at most six rounds, and every Dream Team run gets at least one of these extra rounds.

The A-Team asks you at most once, after Round 1, and only for a fact that could change the answer and that only you have; if you do not answer, it carries on and records that it went ahead without it. The Dream Team pauses after every round, and once before the write-up, so you can redirect it, add facts, or let it continue; after three minutes it continues on its own.

The arbiter seat drafts the verdict from the panel's results: each model's verdict, the points of tension, the claims nobody refuted, and the dissent. The verdict is GO / PIVOT / NO GO — or INSUFFICIENT BASIS when grounding is too thin to recommend either direction. A dedicated writer model then turns the draft and the full texts of Rounds 1 and 2 into the memo you read.

Each claim in the memo is labelled Sourced from data, Cited from web, or AI reasoning, verify independently, so you always know what is grounded and what is reasoning. Pilot5 checks the label instead of trusting it:

• On Round 1. Up to five sourced claims are checked against a fresh web search; names that do not appear in your documents, and figures with no grounding, are flagged; US statutory claims are checked against the enacted text.
• On the arbiter's draft. It must contain its required sections, including the dissenting view.
• On the memo. Every sourced claim must match a source retrieved in the run, and a separate AI model reads the passage behind each one, up to 40 per memo; a claim it cannot confirm is shown as AI reasoning. The verdict, the prose and the arithmetic are also checked by rule.
• Against your document. In legal memos, when your document reached the panel whole, up to 200 statements are checked against it, and confirmed contradictions appear under “Checked against your document”.

The Expert runs one check on its answer: every sourced claim must match a source retrieved in the run.

The result arrives in the conversation as a chat message, in your language, with the full memo one click away, and the names from Step 4 are put back. The memo gives the verdict, the recommendation, the strongest opposing view, and the unknowns that could change the answer, with every source numbered as a footnote. When the question weighs options, a decision matrix grades each one against the panel's criteria, and testable predictions are extracted so you can record later what actually happened. Each panel memo shows how many claims are sourced from data, cited from the web, or AI reasoning, and an Evidence check out of 10 with a reason for every point. Every round's analyses stay attached to the run, which can be reopened, exported as a PDF, or retrieved from an AI assistant. Credits are settled at actual use, within the mode's cap.

All of it (the brief, the research stack, the cross-critique, the checks) for one question. Your question.