How this is checked

Every worksheet here is a draft for preparing a HAZOP meeting. Four layers check it before a qualified team does. None of them replaces that team.

1. Automatic checks on every row

These run in your browser on every worksheet, whether it comes from the library, the building blocks or an AI draft. Red means an error, amber a warning to look at first.

Physics: deviations that cannot happen are rejected, for example reverse temperature, reverse pressure, a level deviation in a node with no vessel, or a reaction deviation with no reactor.
Completeness: every guide word and parameter pair in a node has a row or a "not credible" entry with a reason.
Safeguards are written as "Check: ..." items, because the team must confirm each one exists and works.
Risk matches the risk matrix: severity times likelihood, unmitigated (Low 1 to 4, Medium 5 to 9, High 10 to 16, Very High 20 to 25). One file holds the matrix, so a company matrix can replace it everywhere.
Consistency between severity and consequence type, for example a severity 4 row cannot be "operability only".
Specific wording: causes and actions that say "failure", "upset" or "problem" without naming equipment or a tag are flagged. Every row needs at least two credible causes and a consequence.
Tags named in a row must exist on the flowsheet.
Duplicates and repeated rows are flagged.
Citations must name a real document and page (see layer 2).

2. Citations read against the page

Rows cite free official guidance, incident investigation reports and public HAZOP studies. Each citation in the library and the building blocks was checked by reading the cited page against the row. Citations the page did not support were removed, and rows with no supporting source say "Engineering judgment". Where a page number is too vague, the citation names the section, for example OSHA 1910.119(j), mechanical integrity. If a row is later edited, its citations show as not yet checked until they are read again. The full list is on the sources page.

3. Benchmarks against real HAZOP studies

Two real HAZOP worksheets serve as answer keys: a natural gas compressor unit and a gas turbine lube oil system. Neither is used as a source for that equipment, and each is hidden from the source search during its run. The library, the building blocks and the AI draft are each scored on how many of the scenarios the real team recorded they cover. Missed scenarios are listed so they can be added. See the benchmark page for every run.

4. Engineer review and flags

A qualified engineer approves, edits or rejects each row, with a comment. Every action is logged with the reviewer's name and the date. Every row shows one of these badges:

AI draft, not reviewedWritten by AI or assembled from blocks. Not yet looked at by an engineer.

Automatic checks passedNo errors or warnings from layer 1. Still not engineer-reviewed.

Engineer-reviewedApproved, or edited and approved, by a named reviewer on a stated date.

Under re-reviewSomeone flagged a reviewed row. It stays visible but is checked again.

Anyone can use "Flag this row" to say what looks wrong. Flags go to a private queue that only reviewers can read.

What this cannot do

It does not know your plant. Tags, design pressures, set points and safeguards in a draft are placeholders until your team confirms them against your P&IDs and data sheets.
Automatic checks catch form, physics and consistency errors. They cannot prove that a cause or consequence is credible for your system.
The risk ranking is unmitigated and uses a generic matrix. Use your company matrix and LOPA for decisions.
A HAZOP is a team study. These drafts help the team prepare; they do not replace the meeting, the facilitator or the sign-off.

Draft for HAZOP preparation only. Must be reviewed by a qualified HAZOP team.