New era

The cold chain is becoming a connected machine.

Robotics moves the product. Software coordinates the work. Compute processes the evidence. Electricity supports the system.

Understanding the connections means examining the dependencies.

coldchainOS is investigating how physical operations, digital systems, energy and confidential records intersect. It is not a deployed cybersecurity platform, an autonomous refrigeration controller, a certified engineering service or an incident-response provider.

The connected system

Five layers. Information is not control.

The connected cold chain in five layersFive stacked layers: Intelligence, Evidence, Coordination, Automation and Physical. A solid bracket on the left marks equipment-control authority, which sits with operational technology controllers and authorized operators across the Automation and Physical layers. A dashed upward arrow on the right shows information flowing from the physical layers up to Evidence and Intelligence. There is no control path from Intelligence down to the equipment.No control path down from IntelligenceIntelligenceExtraction · retrieval · reviewEvidenceRecords · bills · inventories · logsCoordinationWMS · WES · TMS · vendorsAutomationAS/RS · conveyors · robotsPhysicalProduct · refrigeration · docks · powerControl authorityOT controllers + operatorsInformation flow (read-only)
Solid line: equipment-control authority. Dashed line: information flow. Illustrative, simplified.
  • Physical. Products, refrigeration, buildings, docks, vehicles and power.
  • Automation. Automated storage and retrieval, conveyors, pallet handling, mobile robots and other equipment actually relevant to the operation.
  • Coordination. Warehouse management and execution, transport management, enterprise systems, scheduling and vendor support.
  • Evidence. Service records, utility bills, equipment inventories, contracts, authorized sensor records and operating logs.
  • Intelligence. Document extraction, retrieval, deterministic calculation, model-assisted review and human interpretation.

  • Software down is not refrigeration down

    A refrigeration system may continue operating during an enterprise-software outage while receiving, picking or dispatch stops.

  • Comms down is not cooling failed

    A communications outage is not automatically a refrigeration failure. Check the controller and the temperature records.

  • A reading is not a record

    A sensor reading is not automatically a verified product-temperature record.

AI and robotics

Every benefit comes with a new question.

We promise no productivity, labor savings or elimination of errors. Not every robot uses a language model.

  • Conventional automation: Fixed logic and control. Not a language model.
  • Optimization: Math that schedules or routes within stated rules.
  • Machine vision: Models that read images, such as inspection or pallet detection.
  • Generative AI: Models that produce text or code. Useful for drafting; needs evidence checks.
Potential applications, the question each raises, and why
Potential applicationThe question to askWhy
Extract equipment details from documentsWhat evidence supports the output?A model can misread a nameplate or invent a field. Every extracted value should point to the passage it came from.
Compare conflicting recordsWhat happens when data is missing or stale?A comparison is only as good as the records and their dates. Missing evidence should be reported as missing.
Explain calculation inputsCan a person check each input against its source?Arithmetic should be deterministic and reproducible; a model can explain it but should not be the calculator.
Summarize maintenance historyWhat evidence supports the output?A summary is not a service record. Gaps and conflicts should stay visible.
Support inventory planning and exception reviewWho can approve an action?Output should inform an accountable person. It should not trigger a consequential action by itself.
Assist warehouse automation and inspection, where actually deployedWhat happens during network or power loss? Can people operate safely without the software? Who maintains it and restores its configuration?Continuity, safe manual operation and ownership of maintenance must be settled before reliance.

The six questions behind every application

  1. What evidence supports the output?
  2. What happens when data is missing or stale?
  3. Who can approve an action?
  4. What happens during network or power loss?
  5. Can people operate safely when software is unavailable?
  6. Who maintains the system and restores its configuration?

Compute is physical infrastructure

Models run on electricity, cooling and people.

  • Cloud training

    Large, occasional jobs on rented hardware.

  • Cloud inference

    Convenient and elastic. Data leaves the building, so permissions matter.

  • Local inference

    Keeps processing in-house where feasible. Still needs power, cooling, storage, updates, replacement planning and administration.

Bigger is not automatically better

For a bounded task, evaluate accuracy, latency, memory, energy and total operating cost, not model size alone.

Energy economics

Use the site’s actual bills and tariffs. We do not attribute every utility increase to AI, and we do not assume spare compute or recovered heat produces net savings. See the power-bill section.

The IEA projects data-centre electricity demand to roughly double by 2030 [International Energy Agency]. A projection, not a statement about any one tariff.

The edge appliance

An owner-controlled appliance. A design goal, not a product.

Edge appliance trust boundaries (design goal)A design goal, not an implemented system. Three stacked zones. At the top, operational technology with refrigeration controllers, PLCs, robots and safety systems. In the middle, an owner-controlled zone holding the appliance. At the bottom, external services. Data moves from the operational technology zone down to the owner-controlled zone only through a read-only export. There is no control path from the appliance up to operational technology. Connections to external services are off by default and need explicit approval.Operational technology (OT)Refrigeration controllersPLCs · robots · safety systemsOutside the appliance’s authorityread-only exportno control pathOwner-controlled zone · managed ITAuthorized documentsLocal modelsDeterministic calculationsSource retrievalEvidence-linked brief (human review)gated · explicit approvalExternal services · off by defaultCloud models · vendors · remote supportUpdates and telemetry reviewed
Design goal, not implemented. Trust boundaries: OT data arrives by read-only export; nothing flows from the appliance to OT.
Intended capabilities and their status
CapabilityStatusWhat it means today
Store authorized documentsPlannedNot built.
Run suitable local modelsPlannedNot built.
Run deterministic calculations offlinePrototypeBrowser calculators on this site are tested; they are not an appliance.
Retrieve source passagesPlannedNot built.
Prepare evidence-linked briefsPlannedNot built.
Continue supported functions without an external model connectionPlannedNot built.
Evaluation schema, 12 draft cases and checksImplemented and verifiedRuns in this repository; labels are unreviewed drafts.
Planned

Architecture goals

  • Sit in a managed IT environment.
  • Use approved read-only interfaces or controlled exports for OT data.
  • Do not bridge IT and OT networks by default.
  • Do not directly control refrigeration, PLCs, robots or safety systems.
  • Do not require public inbound access.

Safety stance

The appliance must not become a single point of failure for cooling. Safety controls and authorized operating procedures stay independent of language-model availability or output.

No performance measurements or support promises are published, because none exist.

Confidentiality and the NDA

Control where the evidence goes.

Local processing can reduce external transfers. It does not automatically ensure confidentiality, cybersecurity, contract compliance or regulatory compliance. We do not claim every cloud upload violates an NDA, and we do not claim local deployment makes an agreement irrelevant.

  • Who may access and process records
  • Whether external processing is permitted
  • Storage, embeddings, logs, temporary files and backups
  • Retention and deletion
  • Vendor telemetry and update traffic
  • Remote support
  • Encryption and key management
  • Access permissions and project separation

No confidential customer material belongs in a public dataset without authorization and rights. This public website accepts no confidential uploads; the dev room is a local-only prototype that sends nothing.

Cybersecurity: the questions that matter

Ask first. Change carefully.

We are not a cybersecurity or incident-response provider. These are review questions for qualified professionals.

Operational

  • Unknown devices and undocumented remote connections

    Start with an asset inventory, including cellular modems added by operators, vendors or integrators.

  • Internet-exposed industrial controllers

    Verify exposure through the owner’s authorized review. Do not assume it either way.

  • Shared or default credentials

    Replace defaults; avoid shared accounts.

  • Excessive access permissions

    Least privilege, with project separation.

  • Vendor remote access

    Monitored and approved, not standing and invisible.

  • Unpatched systems and maintenance constraints

    Changes need approved procedures that account for reliability and safety. Do not patch or scan live industrial equipment blindly.

  • Ransomware affecting business systems

    Know which business functions stop, and whether refrigeration continues.

  • Unverified or unrestorable backups

    Run restoration exercises and record the result.

  • Software supply-chain compromise

    Know what runs on an appliance and where it came from.

  • False, missing or delayed telemetry

    A sensor reading is not automatically a verified product-temperature record.

  • Physical theft or compromise of the edge appliance

    Treat the appliance as a sensitive asset: storage, keys, access.

AI-specific

  • Instructions hidden in retrieved documents

    Treat document contents as untrusted data, not instructions.

  • Confidential information appearing in outputs

    Control who can see outputs; keep projects separate.

  • Untrusted tools and excessive agent permissions

    Give models no more reach than the task needs.

  • Poisoned datasets

    Know the provenance and rights of every training and evaluation source.

  • Unsupported conclusions

    Require source links for every claim.

  • Overconfidence under missing evidence

    Reward appropriate abstention.

  • Treat document contents as untrusted data, not instructions.
  • Keep model outputs separate from executable commands.
  • Require explicit authorized review for consequential actions.
  • Reference an asset inventory, controlled access, segmentation, monitored remote access, backup restoration exercises and approved maintenance procedures.
  • Do not blindly patch or scan live industrial equipment. Changes must account for operating reliability and safety requirements.

Dated real-world evidence · research snapshot, November 9, 2026

What the record shows, and what it doesn’t.

A dated snapshot, not a live feed. Each example shows what it establishes and what it does not.

  • Documented public evidence

    CISA: 2026-07-22

    What it documents
    A government update warning that Iranian-affiliated actors targeted programmable logic controllers, adding detection guidance for malicious changes in Rockwell PLC code modules and widening the manufacturer scope.
    Sector and scope
    Multiple critical-infrastructure sectors, including water and wastewater, energy and government facilities. July 22, 2026.
    Why it matters here
    Cold-storage plants also rely on PLCs and HMIs. The mitigations (control network access, validate project files, tell service providers) are transferable questions.
    What it does not establish
    It does not describe a cold-chain incident, give incident counts or name victims.
    A review question
    Who can reach our PLCs, and when did someone last validate the project files?

    [CISA]

  • Documented public evidence

    CISA: 2026-07-30

    What it documents
    A government alert on a significant increase in actors targeting exposed PLCs in the water sector, locking out operators by changing passwords and IP addresses, with boil-water notices and manual operations.
    Sector and scope
    U.S. water and wastewater systems. July 30, 2026.
    Why it matters here
    It shows how a loss of control access forces manual operation, and flags undocumented cellular modems.
    What it does not establish
    It is a water-sector lesson, not a cold-chain attack, and it gives no counts or attribution.
    A review question
    If we lost remote access to a controller tomorrow, could people operate safely by hand, and do we have a known clean backup?

    [CISA]

  • Documented public evidence

    TGW Logistics: 2026-06-09

    What it documents
    A vendor announcement of an automation package (shuttle warehouse, conveyors, palletizing) for a refrigerated and frozen distribution center in Texas.
    Sector and scope
    One facility under construction; go-live anticipated at the end of 2027. June 9, 2026.
    Why it matters here
    It shows automation moving into temperature-controlled buildings, which adds equipment, power and software dependencies.
    What it does not establish
    It is not operational, gives no throughput or energy figures, and does not describe the control software.
    A review question
    What does each automated stage depend on, and how does the site run when one fails?

    [TGW Logistics]

  • Documented public evidence

    International Energy Agency: 2026 (exact date not confirmed)

    What it documents
    An agency report projecting that global data-centre electricity demand roughly doubles by 2030.
    Sector and scope
    Global projection to 2030. Verified through press summaries only.
    Why it matters here
    Compute draws power, and grid capacity is a shared constraint that can matter to any large load, including a cold-storage site.
    What it does not establish
    It does not show that data centers raise any particular local rate, and it says nothing about one utility’s tariff.
    A review question
    What do our actual bills and tariff show, and what does a new load do to demand charges?

    [International Energy Agency]

  • Documented public evidence

    National Institute of Standards and Technology: 2023-09 (final)

    What it documents
    NIST’s final guide to operational technology security (Rev. 3, September 2023), with a Rev. 4 initial public draft open for comment through Nov 30, 2026.
    Sector and scope
    General OT guidance. Final is Rev. 3; Rev. 4 is a draft.
    Why it matters here
    It is a neutral reference for OT security that addresses performance, reliability and safety requirements.
    What it does not establish
    It is not a compliance determination for any site, and the draft is not final.
    A review question
    Which of these practices has a qualified person confirmed for our environment?

    [National Institute of Standards and Technology]

Why we are building datasets

Teach the system to recognize the gap, and prove when it cannot.

  • Equipment and document consistency
  • Energy calculations and tariff assumptions
  • Operating dependencies and continuity
  • Confidentiality and access boundaries
  • Security review questions
  • Source reliability and stale evidence
  • Unknowns, contradictions and appropriate abstention

Data rules

  • Synthetic examples and appropriately licensed public evidence only.
  • Public availability does not grant redistribution rights.
  • Private client evidence stays separate.
  • A generated label is not verified ground truth. Every case is a draft until reviewed.

What a case contains

case_id · domain · scenario_type · source references and rights · observation date · facts supplied · facts missing · question · expected answer · calculation or rule · acceptable uncertainty · unsupported claims to reject · evidence required · reason to abstain · safe next step · testability · review owner · version. Schema: eval/new-era/schema.json.

An evidence-review card

Illustrative Potential issue

Observation
The offering memorandum says “recently upgraded refrigeration.”
Supporting evidence
A dated invoice for a replacement refrigeration controller.
Missing information
Anything confirming a compressor replacement; equipment schedule; nameplate data.
Potential implication
Compressor age and condition may differ from what the memorandum suggests. Not a finding that equipment failed.
Suggested verification
Request the equipment schedule, nameplate data and service records.
Responsible professional
A qualified refrigeration engineer.
Limitations
Synthetic example. Based only on the documents listed. “Potential issue” marks a supported concern; “confirmed finding” would require evidence that establishes it.

Twelve synthetic draft cases

All are unreviewed drafts. Ten can be checked mechanically; two need expert review.

The twelve draft evaluation cases
CaseScenarioChecked bySplitLabel
NE-001invoice vs claimMechanical checktrainDraft, unreviewed
NE-002nameplate vs consumptionMechanical checktrainDraft, unreviewed
NE-003wholesale vs retailMechanical checktrainDraft, unreviewed
NE-004missing data vs excursionMechanical checkdevDraft, unreviewed
NE-005single point vs routeMechanical checkdevDraft, unreviewed
NE-006vendor claim vs site throughputMechanical checktrainDraft, unreviewed
NE-007untested backupMechanical checkheldoutDraft, unreviewed
NE-008local model vs no external trafficMechanical checktrainDraft, unreviewed
NE-009nda vs processing permissionExpert reviewtrainDraft, unreviewed
NE-010instruction in documentMechanical checkheldoutDraft, unreviewed
NE-011business outage vs refrigerationMechanical checkheldoutDraft, unreviewed
NE-012unknown exposureExpert reviewheldoutDraft, unreviewed

Validating the validator

Planned measures: arithmetic correctness, source support, missing-information recognition, appropriate abstention, unsupported-claim rate, resistance to document instructions, synthetic confidential-data leakage, and held-out performance.

No scores are published because none has been measured. An evaluation does not certify operational safety.

Splits that don’t leak

Cases are split into train, dev and held-out by source family, so related cases never straddle a split. The repository’s checks fail if they do.

Sources

Where every claim comes from.

Checked on 2026-11-09. The same list is docs/new-era-sources.json.

  • CISA, FBI, EPA and partners updated a joint advisory on Iranian-affiliated actors targeting programmable logic controllers across U.S. critical infrastructure, adding guidance to detect malicious changes in reusable code modules in Rockwell Automation PLC programs and expanding the manufacturer scope to include observed targeting of Schneider Electric, Siemens and possibly other PLC manufacturers.

    Published
    2026-07-22
    Period
    Ongoing; original advisory April 2026 (AA26-097A)
    Scope
    Multiple U.S. critical infrastructure sectors named on the page: water and wastewater, energy, government services and facilities
    Kind
    government advisory update
    Qualification
    A multi-sector advisory, not a cold-chain incident. The page gives no count of incidents or organizations, names no victims, and gives no activity dates beyond 'ongoing'. Mitigations listed include controlling network access to PLCs and validating project files for unauthorized changes.
    Verification
    Bulletin page read on 2026-11-09. The fuller advisory AA26-097A was not fetched.

  • CISA said it was observing a significant increase in cyber threat actors targeting programmable logic controllers in the Water and Wastewater Systems sector, with actors modifying passwords to lock out operators and disconnecting PLCs by changing their IP addresses, resulting in boil water notices and sustained manual operations.

    Published
    2026-07-30
    Period
    Described as currently observed; no dates given
    Scope
    U.S. Water and Wastewater Systems sector, entities of all sizes
    Kind
    government alert
    Qualification
    A water-sector lesson, not a cold-chain attack. No incident count, victims or attribution are given. Mitigations listed: disconnect PLCs from the internet and route remote access through a VPN or gateway, enable password protection and change defaults, allowlist remote-access IPs, and keep a known clean PLC backup. The alert notes cellular modems installed by operators, vendors or integrators may not be documented.
    Verification
    Bulletin page read on 2026-11-09.

  • TGW Logistics announced it is delivering an automation solution (a shuttle warehouse, energy-efficient pallet and carton conveyors, and automated palletizing and depalletizing stations) for a Lineage refrigerated and frozen distribution center under construction in Hutchins, Texas. Go-live is anticipated at the end of 2027.

    Published
    2026-06-09
    Period
    Announced June 2026; under construction; go-live anticipated end of 2027
    Scope
    One facility, Dallas-Fort Worth area, Texas
    Kind
    vendor press release
    Qualification
    Not operational. A vendor press release with forward-looking statements; it gives no pallet-position, square-footage or throughput figures, and states no energy figures. TGW's -30 C rating is a general product-line claim, not a project specification. The release does not describe warehouse control or management software for this project.
    Verification
    Press release page read on 2026-11-09.

  • The IEA's 'Key questions on energy and AI' projects global data-centre electricity demand to roughly double by 2030.

    Source
    International Energy Agency: Key questions on energy and AI: Executive summary
    Published
    2026 (exact date not confirmed)
    Period
    Projection to 2030
    Scope
    Global
    Kind
    projection
    Qualification
    A projection, not a measurement, and not a statement about any single utility or tariff. Press summaries cite roughly 485 TWh in 2025 rising to about 950 TWh by 2030 (around 3% of global electricity demand); we do not publish those numbers because we could not confirm them on the primary page. Does not show that data centers cause any particular local rate increase.
    Verification
    SECONDARY ONLY. The IEA page returned HTTP 403 to our checker. The doubling projection was confirmed only through press summaries. Confirm against the report before relying on any figure.

  • NIST SP 800-82 Rev. 3, 'Guide to Operational Technology (OT) Security', is the current final guidance, published September 2023. It provides guidance on securing OT while addressing its performance, reliability and safety requirements.

    Source
    National Institute of Standards and Technology: SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
    Published
    2023-09 (final)
    Period
    Not a data series
    Scope
    OT security guidance, general
    Kind
    final guidance
    Qualification
    General guidance, not a compliance determination for any facility. It supersedes Rev. 2. A Revision 4 draft exists (see nist-800-82r4-ipd) and is not final.
    Verification
    Publication page read on 2026-11-09.

  • NIST published an initial public draft of SP 800-82 Rev. 4 on 2026-09-21, aligned with Cybersecurity Framework 2.0, with comments open through 2026-11-30.

    Published
    2026-09-21 (draft)
    Period
    Comment period through 2026-11-30
    Scope
    OT security guidance, general
    Kind
    DRAFT guidance
    Qualification
    A draft, not final. Rev. 3 remains the current final guidance until a revision is finalized. Cited only to show the guidance is being revised.
    Verification
    Publication page read on 2026-11-09.

Review the cases. Break the assumptions.

Operators, OT engineers, counsel and builders: tell us which cases are wrong, which are missing, and what you would need to trust a tool with your evidence.

Explore the projectBring a real problem