A $4.2M order, four tiers deep, on evidence that isn't there

01The decision, the people, the deadline

The long tail. Dana has a contract with VoltCore and with nobody else. The modules contain cells from a tier-2 contract manufacturer, whose cathode active material comes from a tier-3 processor, whose cobalt comes from a tier-4 refiner Dana cannot name. Four tiers, one contractual relationship. Every claim about the bottom two tiers arrives as a claim about a claim.

02Three ways the evidence fails

Seven facts have to hold for the passport to be compliance-grade. Here they are on 7 July, as they actually stand.

Hallucinated — the paragraph that nearly shipped the order

On 5 July, Meridian's sourcing copilot — an LLM with access to the supplier document store — produced this in its pre-award summary:

// sourcing-copilot/v4 · pre-award supplier summary · 2026-07-05T09:41Z
"VoltCore's LFP-54 modules carry third-party verified 16% recycled content
 under ISO 14067 conformance, certificate CFV-2026-1204, and disclose BMS
 state-of-health parameters per Regulation (EU) 2023/1542 Article 14."

It is fluent. It is correctly formatted. It cites a regulation that exists, an article that exists, and a certificate number that looks exactly like the real one. It is wrong in three separate ways:

Why this is the dangerous one Sparse and stale data announce themselves — a careful reader sees the gap. A hallucination does the opposite: it closes the gap in the reader's mind. Dana's file, read quickly, now appears to contain third-party verification and a compliant BMS disclosure. Nothing in the summary looks uncertain. This is precisely the failure mode where a person exercising ordinary care still ends up releasing the order.

03Day 0 — the receipt that says no

7 July 2026, 14:33 UTC. The seven fields go into the engine. Nothing is skipped.

The receipt does not say "risky." It says which field, why, and what would clear it. That converts directly into an evidence request:

// Evidence request ER-4417-01 · to VoltCore Energy GmbH · 7 Jul 2026 · due 17 Jul 2026
Referencing receipt sha256:539c1503…

1. soh_algorithm_disclosure   — BMS state-of-health parameter set + firmware
                                interface spec. Clears at VERIFIED on
                                independent confirmation against shipped firmware.
2. renewable_energy_share_    — Grid supply attestation for the Leipzig line,
   manufacturing                with guarantees of origin. Clears at VERIFIED.
3. material_origin_cobalt     — Physical traceability documentation to the
                                refiner, if it exists. Currently modeled (0.72)
                                on a book-and-claim allocation.

Note: certificate CFV-2026-1204, referenced in our pre-award summary, does not
resolve in the issuer register. Please confirm CFV-2026-1188 scope.

04Hold, fold, or go: how the branch is chosen

WITHHELD is not automatically "walk away." The receipt tells you the verdict; obtainability tells you the action.

05Day 9 — the receipt that says yes, with conditions

16 July 2026. VoltCore delivers two of the three items.

Decision: go forward, with conditions. PO-2026-4417 released 17 July — four days of float remaining, November slot held. The CAUTION verdict is not a warning that gets filed away; it is the source of four specific commercial terms:

Cost of the nine-day hold: $18,000 in re-planned inbound logistics. No expedite premium — the hold fit inside the float.

06What the other path would have cost

Suppose the copilot's paragraph had gone unchallenged and the PO released on 7 July.

Against $18,000 for the hold. The ratio is roughly 18:1 — but the ratio is the smaller point. The larger one is that the $18,000 was a decision, made deliberately, on a documented basis, and the $325,500 would have been an accident.

07The evidentiary test: six questions a reviewer asks

Whether the reviewer is an internal auditor, a customer's compliance team, a notified body, an insurer, or opposing counsel, the questions converge.

08Day 180 — the backtest that changes the rules

3 January 2027. Meridian has 46 sealed decisions in this class. Every one can be re-run exactly as issued. Now the outcomes are known.

Two properties make this legitimate rather than self-serving. First, every earlier receipt still reproduces exactly as issued — the new weights carry a new version number and apply only to decisions made after the change, so nobody can claim the past was reinterpreted to look better. Second, a feature is only promoted if it survives a false-discovery-rate control applied across every hypothesis ever tested — which is what stops a team from fishing through forty variables until something looks predictive. Both are covered by the patent's fourth independent claim.

The mechanism, in ordinary language, is on the Better Decisions through Backtests and Recursive Improvement page.