Noon-report cleaning
A skill for validating vessel noon report / consumption data against CII, EU MRV, and FuelEU Maritime rules. Used today on the ship-operator side.
Validation at scale — an agent team approach
Class NK already operates ZETA at scale — the GHG / CII / EU MRV / FuelEU compliance platform with 5,000+ ships and NAPA Fleet Intelligence integration since July 2025. Most of the validation work that flows through ZETA, and through several other Class NK pipelines, follows a recognisable pattern: a client uploads data · someone checks it against rules · the result goes back to the client.
This page documents a pattern for how an agent team can handle the routine parts of that work, while the surveyor stays at every judgment seat. It is a pattern to explore, not a proposal to adopt.
What the data looks like today
A constant stream of operational data flows from owners into Class NK:
- Noon reports — daily vessel performance, fuel, position, weather
- Fuel and emissions logs — for CII, EU MRV, FuelEU compliance
- Certificate evidence — survey photos, inspection records, type-approval data
- Defect rectification proof — closing the loop on raised non-conformities
- Technical files — design approval submissions, equipment specifications
Each submission needs a human to look at it, check it against the relevant rule, decide if it passes, and either approve it or send it back. That work is necessarily slow and necessarily expert — and the surveyor who reviews it is doing two distinct kinds of work at once: routine cross-checking, and expert judgment.
What an agent team approach looks like
An agent team can sit between the client upload and the surveyor’s review. Each agent has one job. Each handoff is logged. The surveyor only enters the loop where judgment is needed.
Client upload │ ▼┌──────────────┐ ┌──────────────┐ ┌──────────────┐│ Intake │ ───▶ │ Validator │ ───▶ │ Anomaly ││ agent │ │ agent │ │ detector │└──────────────┘ └──────────────┘ └──────────────┘ │ │ │ ▼ ▼ ▼ Ingest, clean, Check against Flag entries normalize data. SOLAS, MARPOL, that look wrong, Detect format Class NK rules. with reasons. problems early. Cite the rule. │ ▼ ┌──────────────┐ │ Casefile │ │ agent │ └──────────────┘ │ ▼ Open or update the casefile. Attach evidence. │ ▼ ┌──────────────┐ ┌──────────────┐ │ Escalation │ ──▶ ─── │ Surveyor │ │ router │ │ (human) │ └──────────────┘ └──────────────┘ │ │ ▼ ▼ Send edge cases Reviews flagged to the right items. Approves or surveyor. requests change. │ │ ▼ ▼ ┌──────────────┐ ┌──────────────┐ │ Audit │ │ Client │ │ logger │ │ notified │ └──────────────┘ │ with detail │ └──────────────┘Every step is recorded. Every decision has a citation. Every escalation has a reason.
The capabilities behind this
A few discrete skills, each runnable on its own, compose into the team shown above:
Vessel-data validation
A skill that cross-checks fuel logs, emissions data, and operational records against regulatory rule sets.
Veracity validation
A skill that exercises DNV Veracity’s OVD sandbox — built originally for ship-operator submissions to DNV.
Casefile system
A durable folder structure that gathers evidence, threads decisions, and preserves provenance.
The orchestration substrate is what makes these work as a team — handoffs, memory, audit, a single human at the judgment seat.
What changes — and what doesn’t
The routine work changes shape. The expert work doesn’t.
| Before | With an agent team | |
|---|---|---|
| Routine cross-checking | Done by surveyor, line-by-line | Done by agent, with citations |
| Expert judgment | Done by surveyor | Still done by surveyor |
| Audit trail | Reconstructable from email + files | Single linked casefile, machine-readable |
| Client visibility | Silence until decision | Live progress with rule-by-rule reasoning |
| Surveyor time per submission | Hours | Minutes for routine; full time for judgment calls |
The surveyor’s role becomes more specialised, not less needed. The routine work that doesn’t require their expertise stops absorbing their time.
What this is not
A few things worth being clear about:
- Not full automation. The surveyor stays at every judgment point.
- Not a black box. Every agent decision has a citation. Every escalation has a reason. The audit trail is human-readable.
- Not external. If Class NK ever ran this pattern, it would run on Class NK’s own infrastructure. Class NK’s data wouldn’t leave Class NK’s perimeter.
- Not all-or-nothing. A pattern like this would naturally start with one data type — perhaps noon reports — and extend from there if it proves useful.
Where this could fit
If Class NK ever wanted to explore the pattern, the most natural starting place is probably noon reports — high volume, an existing rule set (CII / EU MRV / FuelEU), and an existing Class NK pipeline (ZETA) that an agent team could sit alongside. The change wouldn’t require a new platform or a new business mechanic — it would be an additional layer on top of work Class NK already does well.
Whether and when that would be worth pursuing is for Class NK to decide.
See the demo walkthrough The agent team chain end-to-end, with commentary.