ColdCall

Keeps medicine shipments honest

{{ n.label }}1

An AI agent for medicine that got too warm

Is this medicine still safe to release?

Medicine travels in refrigerated trucks, and sometimes those trucks get too warm. When that happens, someone has to decide: is the shipment still good, does it need re-testing, or must it be destroyed? ColdCall is an AI agent that investigates each case — it reads the drug's official label, checks the route and the weather, runs the regulated temperature math, has a second, independent calculation verify every number — and then stops and waits for a human to sign the final decision.

See a decision being made

Where the data comes from

{{ s.label }}

{{ s.value }}

What the animation above shows

{{ e.n }}

The three possible outcomes

{{ l.pattern }}

{{ l.n }}

Every outcome has its own pattern as well as its own colour, so the meaning survives even in black-and-white print.

What we are honest about

Every number is calculated twice, by two independent programs. Two programs agreeing is still not absolute proof — but it catches a coding mistake in either one, and we say so on every verdict.

We clearly separate what comes from regulation (the temperature math, the drug label's range) from what is our own configurable policy (how many hours of excursion we allow). The two are labelled everywhere they appear.

No verdict on this site is invented for display — each one is real computed output, and if the calculation fails the screen says so.

{{ r.label }}

{{ r.value }}

Medicine that goes bad in transit is a $35 billion-a-year problem

Vaccines, insulin, antibiotics — a huge share of modern medicine must stay inside a narrow temperature band from the factory to the pharmacy. That journey is called the cold chain. When a shipment warms up, someone has to decide: is it still safe to sell, does it need lab testing, or must it be destroyed? Today that decision is slow, manual, and expensive.

{{ s.value }}

{{ s.src }}

Why one warm truck costs so much

The 45-day investigation, done in minutes — with a human still in charge

ColdCall watches the temperature recording, and the moment a shipment leaves its band it builds the whole case file automatically: where it warmed, the regulated math, the money at risk, the nearest certified cold store. A second, independent program recomputes every number. Then it stops — and a named person makes the one irreversible call.

{{ s.value }}

{{ s.src }}

Where the money comes back

Proved on 206 real shipments — nothing simulated

Claims are cheap, so here is the evidence. We replayed 206 real shipment recordings from five public datasets through the exact same pipeline that handles the live incident — every verdict below comes from a real temperature logger, and every number was computed twice by two independent programs before it counted.

{{ r.label }}

How the 206 shipments were decided

both calculations agreed on all 206

{{ l.n }}

Where the shipments come from

Dataset

Trips

Verdict mix

R · Q · D

{{ ds.legs }}

{{ ds.counts }}

Incident target

VCC-118

Advisory

Temp excursion

Lot

A2231

Units

4,000

Route

Rotterdam DC → Lyon regional hub

Opened

2026-08-29 11:52:32

Phase

{{ roomStatus }}

QUARANTINE / RETEST

verdict of record · VCC-118

{{ gateBandChip }}

Proposed action

Move the shipment to a certified cold store, hold it for lab testing, and file the incident report

If you allow

Executes now, a receipt is written, and the room closes under your name.

If you deny

Nothing is executed — the shipment stays safely held, and your reason goes on the record.

Either way

Appended to the feed; it cannot be edited or removed.

Signature

Deny

Receipt

{{ receiptId }}

Action

quarantine at a qualified store

Signed by

{{ actionTime }}

Standing

executed · room closed

appended to the feed. it cannot be edited or removed.

Conservative fallback

Nothing was executed.

Denied by

{{ actionTime }}

Reason

Fallback

shipment stays quarantined pending QA review

the denial and its reason are appended to the feed.

Stability budget consumed

64.35%

of the 6 h allowance
$50,000 at risk

retest 50%
destroy 100%

35.6 points remain before this becomes destroy.

Independent cross-check

0.000e+00

°C MKT difference
2 of 2 implementations

Primary

quarantine / retest

Independent

quarantine / retest

Agreement is not proof of correctness — it catches a coding error in one path.

The three possible outcomes · measured against the heat allowance

Action

Standing

Value Δ

Basis

Room feed · append-only

{{ f.t }}

{{ f.body }}

Who is waiting on this shipment · 3 named receivers

{{ p.units }}

{{ p.value }}

Certified cold stores nearby · where it can be held

{{ s.dist }}

{{ s.note }}

Audit dossier · the record as it stands

Open the leg evidence →

Print dossier

{{ d.v }}

How the agents work · run

DEMO-0001

same input, same answer

T+ 10.1s
complete 9/11
HALTED AT GATE

The architecture · watch a run

Evidence flows left to right; the four specialists work in parallel; the run pauses at the human gate.

The team · 10 agents, 1 human

{{ a.note }}

What happened, step by step · nothing can be edited out

{{ e.t }}

{{ e.to }}

{{ e.body }}

The agent stopped itself here. Nothing irreversible has happened yet.

Open the gate →

The recommendation

QUARANTINE / RETEST

MKT
24.54 °C
Label
20–25 °C
Budget
64.35% of 6 h
Out of range
232 / 1,233 min
Cross-check
Δ 0.000e+00

Known limitation, on the record

Two programs agreeing is strong evidence, not absolute proof — they could share the same misreading of the standard. We record this on every verdict.

Proof it works: 206 real shipments

We replayed 206 real recorded shipments — vaccines, mangoes, strawberries, grapes, pharma parcels — from five public research datasets through the exact same pipeline. Nothing here is simulated: every row is a trip a real temperature logger once recorded.

Dataset

Trips

Policy

Verdict mix

R · Q · D

{{ ds.src }}

{{ ds.legs }}

{{ ds.policy }}

{{ ds.counts }}

Shipment register · click any row to see its evidence

every number below was computed from that trip's real recording — none are typed in

Leg

Verdict

MKT °C

Budget

Out of range

Peak °C

Trace

{{ l.id }}

Seven representative trips carry their full temperature chart in this build; every other row still shows the complete computed results for its recording.

← corpus

{{ legId }}

{{ legDsTitle }}

Temperature trace · recorded telemetry

{{ legWindow }}
{{ legChartNote }}

The full temperature chart for this trip is not embedded in this build — the register at right shows everything computed from its {{ legN }} recorded readings. Seven representative trips carry their full chart; open one from the shipment register.

Rationale · budget-grounded

Two implementations agreeing is not proof of correctness — it catches a coding error in one path. Recorded as a limitation on this verdict.

This recording is the open incident VCC-118 — lot A2231, 4,000 units, $50,000 at risk.

Open the response room →

The register

{{ m.v }}

Stability budget

retest 50%
destroy 100%

Policy behind the numbers

Incidents it caught

HOLD

VCC-118 — quarantine / retest

INC-20260829-6D09F2 · lot A2231 · 4,000 units · Rotterdam → Lyon

Timeline · append-only

{{ e.t }}

{{ e.body }}

closed · INC-20260828-5E6958 · quarantine / retest · signed 2026-08-28 · receipt on file

Caught in the recorded datasets · the worst case from each

{{ c.detail }}

{{ c.nums }}

evidence →

Each row is a real temperature excursion found in the public recordings — the single worst trip from each dataset. Click through to see its full evidence: the temperature trace, the math, and the verdict. The complete register of all 206 trips is on the Evidence page.

Where the numbers come from

Every figure on this site traces back to a public source. Below: the five datasets we replayed, the standards the math follows, the market figures on the Overview — each with a link — and an honest note on what is regulation, what is our own policy, and what we cannot claim.

01 · The five datasets · 206 recorded trips, nothing simulated

Dataset
Publisher
Trips
Link

{{ d.trips }}

{{ d.link }} ↗

02 · The standards and services the math follows

{{ d.link }} ↗

03 · The market figures quoted on the Overview

{{ d.name }}

{{ d.link }} ↗

04 · What is regulation, what is ours, and what we cannot claim

{{ p.kind }}

phase · {{ conPhase }}

Excursion

road leg · reefer

Incident target

VCC-118 · lot A2231

4,000 units · amoxicillin 500 mg
detected 2026-08-29 11:52:32

Stability budget consumed · of a 6 h allowance

64.35%

$50,000 at risk
35.6 points to destroy

held {{ dwell }} since detection · budget is fixed, the clock is not

Where it warmed

3 cold stores in range

Rotterdam DC
road leg · 1,233 min
Lyon regional hub

When it warmed

232 min out of range

labelled 20–25 °C

excursion · peak 27.0 °C

08:23
logger DD:33:04:13:34:CD
04:56 +1

Courses of action · budget-grounded

threshold provenance ↗

Action

Standing

Value Δ

Basis

Human decision log

Open the gate

Full transcript

Room feed · append-only

live

{{ f.note }}

Room opened

11:52:32

Incident room created for VCC-118 on the Rotterdam → Lyon leg

INC-20260829-6D09F2

ColdCall · deterministic cold-chain disposition · offline demo
206 legs · 5 datasets · generated 2026-08-29