THE RECON-
STRUCTION
The examiner has arrived. The analyst left last spring, the chats were purged on schedule, and the memos cannot speak for themselves. Five claims stand before you with whatever survived around them. Find the claim the record can no longer carry — and once, defend the record that holds.
Press → or swipe to begin. The record starts when you do.
Memory is not a record.
Eighteen months from now, nothing else speaks.
Each examination shows a memo's key claims beside an inventory of what actually survives — extracts, registries, worksheets, retention schedules. Exactly one claim per examination cannot be reproduced from the surviving record: the evidence was purged, was never captured, contradicts itself, or lived on a channel that preserved nothing. Find it. And in Examination Five, be ready to attest a record that holds.
“The work happened. The record didn't survive. Regulators priced that difference at more than two billion dollars.”Module 03 — The Record · on the off-channel communications sweep
One irreproducible claim per examination. The other claims trace cleanly to surviving artifacts. One attempt.
Reproducible means: an independent party, today, could re-derive the claim from preserved artifacts. “The analyst remembers” is not an artifact.
Before lodging, stake your conviction. MEASURED: +1 right, 0 wrong. HIGH CONVICTION: +2 right, −1 wrong. The call is the skill; calibration is the discipline.
Every selection, stake, and call is timestamped and hash-chained in your browser — the append-only invariant used for agent audit trails. Your run is itself evidence.
Group format: read the claims aloud, then the record inventory. Vote on the irreproducible claim, then on the stake. The best debates start when someone says “but we know that's what happened” — that sentence is the whole reason Module 03 exists. Budget five minutes per examination.
The Vanished Chat
Gap: Claim 3. The session was real and is gone. The repair is procedural: AI-assisted analysis that feeds a decision gets exported to the evidence store at creation, inside the memo's retention class, not the chat tool's.
Discussion: what is your chat tool's retention period, what is your credit file's, and which claims in last quarter's memos currently live in the gap between them?
The Version Ghost
Gap: Claim 1. The registry outlived the memo and contradicts it. Repair: score outputs written to the evidence store with version, run ID, and input hash, at scoring time.
Discussion: pick any model-derived number in a live memo and ask the room to name its version and run ID from the record. Time how long it takes. That duration is your exam finding.
The Untraceable Delta
Gap: Claim 2 (the €0.3M delta). Base reproducible, approval reproducible, bridge gone. Repair: a one-page adjustment memo per overlay — basis, quantum, approver — filed in the same retention class as the provision.
Discussion: overlays are where auditors camp for a reason. Ask the room: how many adjusted figures in your last close have a standalone artifact for the adjustment?
The Off-Channel Approval
Gap: Claim 4. The sweep that produced $2B+ in fines was about precisely this: real business, unpreserved channel. Repair is cultural and mechanical — approvals land in the workflow or they don't count, and everyone knows it.
Discussion: ask the room, honestly, where the last urgent approval in their business actually happened — and whether it would survive this examination.
The Intact File
Finding: ATTEST REPRODUCIBLE. This file is the target state: AI work archived at creation, versions pinned, deltas documented, approvals on channel. The drill ends by showing the room what passing looks like.
Discussion: gap analysis, live — which of the five artifact disciplines shown here does your team already do, and which of today's four failure classes is currently sitting in your files?
Complete the five scenarios to receive your rating.
Build work that explains itself.
Five examinations, five artifact disciplines. Carry these forward:
Retention clocks are evidence clocks. AI sessions that feed decisions get archived at creation, in the decision's retention class — the purge runs on schedule, the examiner doesn't.
Pin the version or lose the provenance. Model ID, version, run reference, input hash — or the number is an orphan in eighteen months.
Every delta needs lineage. Overlays are legitimate; undocumented overlays are findings. Each adjustment carries its own basis, quantum, and approver.
Approvals live on preserved channels or not at all. The industry paid $2B+ to learn that the work happening is not the same as the record surviving.
Attest the record that holds. Reproducibility is provable. An evidence culture that can only convict teaches people to stop keeping evidence.
When the examiner arrives, the chat will be gone, the analyst will have moved on, and the memo cannot speak for itself. Build the file that speaks. Run this drill with your first line and your auditors in the same room — they will disagree about Examination Five, and that disagreement is the training.
THE STUDIO — LIVE FIRE 03 · COMPANION TO MODULE 03: THE RECORD