The Panel

Five perspectives that deliberate every question.

Disagree to decide.

The Panel is the five-perspective deliberation council that anchors every Pilot5 deliberation. Five AI models, MECE-designed, with deliberately non-overlapping mandates. The Architect, the Counsel, the Strategist, the Engineer, the Contrarian — each asks a different question, and none of them is allowed to stand in for another.

Models analyse independently in the first round (Round 1 — Divergence, zero cross-consultation, audit-traceable isolation). They critique each other anonymously in the second round, knowing each other only as ANALYSIS_1 through ANALYSIS_5. The arbiter persona — selected by domain — synthesises only what survived attack.

The same five-perspective structure runs every deliberation, every domain, every region. The model behind each persona shifts at runtime (benchmark-driven selection by use case and budget), but the role contract does not.

Structure, benchmarks, operational rigor

The Architect

“What does the data actually say?”

Builds the spreadsheet. Anchors the room in numbers — revenue impact, downside cases, operational dependencies. Refuses to let the conversation drift into vibes when the data is available.

Default model
gpt-5.2
Leads on
general · pricing · logistics · finance · operations

Ethics, second-order effects, reputational and legal risk

The Counsel

“What could go wrong, and for whom?”

Maps the people, contracts, regulators, and norms that the recommendation will collide with. Ethical clarity without moralising. Knows the difference between risk that should stop a decision and risk that should be managed.

Default model
claude-sonnet-4.6
Leads on
legal · risk · compliance · hr · security · medical

Macro trends, competitive dynamics, long-horizon positioning

The Strategist

“What does this mean in context?”

Reads the market signal. Tracks where the category is moving and what staying put would imply. Long-context synthesis: the Strategist ties what's in front of the panel back to where the world is heading.

Default model
gemini-2.0-flash
Leads on
strategy · marketing · product · sales · research

Technical feasibility, code accuracy, mathematical precision

The Engineer

“Does this actually work?”

Stress-tests the implementation. Estimates effort, latency, integration risk, and edge cases. When the recommendation has a number in it, the Engineer is the one who checked it.

Default model
x-ai/grok-4
Leads on
code · architecture · data · tech

Sovereign perspective, anti-convergence by mandate

The Contrarian

“What if everyone here is wrong?”

The structural counterweight. Mandated to challenge whatever the other four agree on, even when the consensus is well-evidenced. Premature agreement is a failure mode, not a success — the Contrarian is the architecture's defence against it.

Default model
mistral-large-latest
Leads on
design— The only domain where dissent itself is the methodology. Across the other twenty domains the Contrarian preserves disagreement rather than resolving it.

How the panel runs

  • Blind divergence. All five analyse the question in parallel with no cross-consultation (asyncio.gather over all 5 personas in pipeline.py). Architecture-level isolation; the orchestrator emits an R1_ISOLATION_PROOF event with a SHA-256 hash of the exact context each persona received, archived in telemetry.
  • Anonymous critique. Round 2 relabels the analyses as ANALYSIS_1 through ANALYSIS_5. Personas attack the weakest reasoning in the others without knowing which persona — or which model — produced it. The critique is on the argument, never on the brand.
  • Arbiter synthesis. A single persona — selected by use-case across 21 domains — produces GO / PIVOT / NO GO (or INSUFFICIENT BASIS when grounding is too thin) plus a confidence score and a Minority Report when one perspective refused to converge. The arbiter map is below.
  • Adaptive rounds. When post-critique agreement crosses the consensus threshold (≥ 0.9 for The A-Team, ≥ 0.8 for The Dream Team), or the panel flags information gaps or hidden assumptions, the orchestrator queues additional rounds before synthesis can run. Five round types, listed below.

Arbiter routing — 21 domains, 5 personas

The synthesis arbiter is not the same persona on every question. The orchestrator picks by domain — pricing routes through the Architect, code through the Engineer, legal through the Counsel. 21 use-cases are configured today.

  • The Architect5 domains
    general · pricing · logistics · finance · operations
  • The Strategist5 domains
    strategy · marketing · product · sales · research
  • The Engineer4 domains
    code · architecture · data · tech
  • The Counsel6 domains
    legal · risk · compliance · hr · security · medical
  • The Contrarian1 domain
    design
    Where dissent itself is the methodology. Preserves disagreement across the other twenty.

When the orchestrator extends

Three rounds is the spine. For deeper questions — or when the panel agrees too quickly — the adaptive orchestrator queues additional rounds, each tied to a specific failure mode it’s designed to catch. These are real round types in the dispatch enum, not marketing names.

  • devil_advocatePost-critique agreement ≥ 0.9 (A-Team) / ≥ 0.8 (Dream Team)

    Forces a structured rebuttal of the emerging consensus. Premature agreement is a failure mode, not a signal. The Dream Team's lower threshold means it gets stress-tested sooner.

  • assumption_surfacingHidden premises flagged in critique

    Pulls implicit assumptions into the open and stress-tests them before they enter synthesis.

  • focused_researchInformation gaps flagged

    Targeted retrieval on a single unresolved question — rather than re-running the whole research wave.

  • calibrationConfidence drift detected

    Re-anchors confidence scores against the evidence that actually survived critique.

  • client_checkinDream Team only — every round, deterministic

    The deliberation pauses for your steering input. Status flips to awaiting_feedback; resumes on submit.

Resilience — when a slot needs a fallback

Every persona has a four-step fallback chain. If the Architect’s primary provider returns an error or is rate-limited, the slot routes through the next persona in chain — not by re-running the Architect, but by mapping the role to the closest substitute. The five-perspective contract holds; the model behind the slot may not.

  • The Architect
    Engineer → Counsel → Strategist → Contrarian
  • The Counsel
    Strategist → Contrarian → Architect → Engineer
  • The Strategist
    Counsel → Contrarian → Architect → Engineer
  • The Engineer
    Architect → Strategist → Counsel → Contrarian
  • The Contrarian
    Strategist → Counsel → Engineer → Architect

Continue

  • How it Works → — the deliberation pipeline end-to-end, including round orchestration and arbiter selection.
  • Regional Panels → — US, Global, EU, APAC, MENA panel configurations and the models that anchor each one.
  • Audit Trail → — how the five-position record is preserved, including the Minority Report.
  • The Three Modes → — when the full panel runs versus when a single smart-routed model is enough.