Governance & accountability

“The AI told me so”
is not a defense.

When an AI-assisted decision is challenged (by a board, a regulator, a counterparty, or a court), a conversation log with three confident paragraphs is not a defensible record. A Pilot5 deliberation is.

Standard AI — 12 months later, board review

“On what basis was that recommendation made? What risks were flagged? What alternatives were considered?”

“I have a conversation log. Three paragraphs. The model said it looked standard. There is no further record.”

Pilot5 deliberation — same question, same board

“On what basis?”

“Five independent analyses. Cross-critique. Evidence check 6/10, with the reasons on record: what the sources supported, and the open questions that could still change the answer. The Contrarian identified the jurisdiction risk. Recommendation: proceed with clause modification. Full record here.”

Deliberation ID on record

Every deliberation is timestamped, archived, and retrievable by ID. The record exists before you need it.

5 independent positions documented

Not a unified AI answer. Five separate analytical positions — reviewable individually, with dissent explicitly preserved.

Evidence check on record

The Evidence check and its reasons, the dissenting position, and the information gaps are all part of the archived record — not added after the fact.

What survives the room

Independence is enforced at the architecture level — not through prompting or instructions that could be ignored. In Round 1, the five AI models receive the brief via parallel async calls (asyncio.gather in pipeline.py). No perspective sees another’s response until all five have completed.

Every A-Team and Dream Team deliberation records an R1_ISOLATION_PROOF event: a SHA-256 hash of the shared Round 1 context, logged at dispatch before any perspective answers and archived alongside the telemetry.

What lands on the record: GO · PIVOT · NO GO, or INSUFFICIENT BASIS when grounding is too thin to recommend either direction, with an Evidence check, decision matrix, action plan (each step keyed by action, owner, deadline), information gaps, falsification conditions, and a preserved Minority Report when one perspective refuses to converge. It is not averaged. It is not diplomatically softened. It is the argument the deliberation could not refute.

What’s actually on the record

“Audit trail” isn’t a single document. It’s a structured event stream + persisted artefacts the orchestrator writes as the deliberation runs. Each line below is a real event type or table row, not marketing language.

event
R1_ISOLATION_PROOF — SHA-256 hash of the shared Round 1 context all five personas received, with the persona count and dispatch time. Emitted once per panel deliberation, before any persona returns.
event
SOURCED_TAGS_DOWNGRADED — emitted when the audit verifier rewrites a fabricated [SOURCED] claim to [INFERRED]. Fields: round, persona, count.
event
FALLBACK_TRIGGERED — emitted when a persona slot routes to its fallback model (rate limit, provider outage, model error). Tells you which model actually answered, not just which was scheduled.
status
awaiting_feedback (Dream Team) — the deliberation halts at this status between rounds for HITL steering. Each pause is part of the record — including the user feedback that resumed it.
table
deliberation_telemetry, response_telemetry — row per round / per persona, with cost, latency, token counts (including prompt-cache reads and writes), fallback chain depth.
event
INPUT_GUARDRAIL_BLOCKED — emitted when the safety guardrails redact content from a round’s input, with the reasons.

How to retrieve a deliberation record

Every deliberation is keyed by a UUID. Two endpoints return the record: one for the structured snapshot, one for the full event timeline.

GET /v1/deliberations/{deliberation_id}
The full structured record — brief, all rounds, all persona responses, synthesis output, Minority Report. UUID key. Ownership-enforced: only the user who launched the deliberation receives the body.
GET /v1/deliberations/{deliberation_id}/events
The full event timeline — every dispatch, every completion, every fallback, every audit downgrade. Ordered chronologically with timestamps. This is what a regulator, auditor, or counterparty actually wants to see.
POST /v1/deliberations/{deliberation_id}/outcome
Submit observed outcomes against testable predictions the deliberation made. The verdict was not the end of the record — the predictions are testable later, and the outcome you log here closes the loop.

The defense exists before you need it. For board-level decisions, M&A diligence, regulatory exposure, contract review, capital allocation, or any commitment that may be challenged later, the deliberation record is the documentation.

How it works →For legal & compliance →