How we use AI

A model reads. Code checks. You decide.

Hazzard uses language models to read plans at a scale no team can, and nothing a model writes reaches your record unchecked.

Book a walkthroughSee how it works →

  1. A model reads

    • Reads scanned pages, tables and figures
    • Pulls out every hazard mention, word for word
    • Drafts justifications from verified evidence

  2. Code checks

    • Every quote against the page it cites
    • Every drafted sentence for a verified source
    • Hazard names against the taxonomy

  3. Your team decides

    • Which hazards are on your list
    • Which draft to accept, edit or rewrite
    • Every 1–5 score, and who approves it

1 · Reading the plan

First, the PDF becomes something we can check.

Before any AI writes a word, every plan goes through seven steps. Five are ordinary software that gives the same result every time. AI helps with two: reading the pages plain text can’t, and making the plan searchable.

  1. PDF
    a1c3…9f
    412 pages
    Opens fine
    1Software

    Check the file

    We make sure the PDF opens, count its pages and give it a fingerprint, so the same plan is never stored twice.

  2. TextScanTableMap
    2Software

    Sort every page

    Each page is sorted: ordinary text, a scan, a table, or a map or picture.

  3. 1
    2
    3
    4
    5
    3Software

    Copy out the words

    The words and tables are copied out in reading order. Scanned pages are read with text recognition.

  4. Scans, tables, maps
    Read by AI
    4AI model

    Read the hard pages

    An AI vision model looks only at scans, tables and maps, the pages plain text can't capture.

  5. 1Introduction
    2Hazard profiles
    2.1Wildfire
    2.2Flooding
    3Mitigation
    5Software

    Rebuild the outline

    The table of contents, chapters and sections are put back together.

  6. WordsTables & mapsOutline

    One trusted copy

    Every citation points here

    6Software

    Make one trusted copy

    Everything becomes one digital copy of the plan. Every quote and footnote later points back to it.

  7. evacuation notices
    4.2 Wildfire history
    6.1 EvacuationNotices reached 61% in the first hour
    7AI model

    Make it searchable

    Each section is indexed with a short summary, so the right passage turns up in seconds.

2 · Extraction

What the model may say is a form, not an essay.

Code decides which sections a model reads. For each hazard mention, the model must fill in a closed form: the hazard in the document’s own words, the kind of claim, how strong the evidence is, which way it points, a verbatim quote and where it sits.

  • Mentioned Named in a document with no assertion of local exposure
  • Asserted A document asserts exposure, without supporting data
  • Substantive A document carries real risk data — frequency, extent, history
  • Counter-evidence “No tsunami has ever struck” is shown as such, and never counts as exposure.

One mention, as the model must return it

Hazard, as written
wildland fires
Matched to
Wildland Fire
Kind of claim
History
Evidence tier
Substantive
Stance
Supports
Where
Pine Hollow Fire After-Action Report, p. 4, block p004-b002
Quote
“the District responded to 64 wildland fires”

Fields outside this form are refused. The quote is checked next.

3 · The quote check

If the quote isn't on the page, it's gone.

Every quote the model returns is checked in code against the block it cites, word for word, ignoring only capital letters and spacing. A quote that isn’t there is discarded and the reason logged next to what was kept. This is not an instruction to the model; it is a check the model can’t talk its way past.

Pine Hollow Fire After-Action Reportp. 4 · p004-b002

Between 2010 and 2024 the District responded to 64 wildland fires, the majority of which burned less than one acre. The Pine Hollow Fire of August 2018 burned 1,240 acres and destroyed 14 homes.

The model says the page says

“the District responded to 64 wildland fires”

Found in block p004-b002 · kept, p. 4

The model says the page says

“the District responded to more than 80 wildland fires”

Not on the page · discarded, reason logged

Then the hazard name finds its place in the taxonomy

  1. 1Code

    Aliases

    “Wildfire”, “WUI fire”, “brush fire” — a fixed list per hazard.

  2. 2Code

    Learned mappings

    Names a person has already confirmed.

  3. 3Model

    One model call

    Only for names nothing else can place, applied at 0.8 confidence or higher.

  4. 4Person

    A person's queue

    Anything left stays honestly unmatched until someone decides.

4 · The writing

Evidence first. Writing second.

A model writes justifications only after the evidence has passed. It is handed the quotes that survived the check, the facts your team kept and the federal record, and nothing else. A fact you exclude is never given to it. A sentence that cites nothing, or cites something it wasn’t given, is dropped and counted.

Book a walkthrough

Only what already passed

  • Verified quote

    “64 wildland fires from 2010 through 2024…” · C, p. 4

  • Verified quote

    “18,400 acres of wildland–urban interface…” · B, p. 12

  • Federal record

    FEMA NRI wildfire risk: relatively high · NRI 1.20

The model drafts

Every sentence must name what it cites

Citation check · code

The county recorded 64 wildland fires from 2010 through 2024, most under one acre.1 About 18,400 acres of interface lie along the western foothills.2

Wildfire activity in the region is expected to double by 2040.

Cites nothing it was given, so it is dropped and counted.

A person accepts, edits or rewrites it. Until then it is a draft.

What the checks can’t do. A check can prove a quote is on the page; it cannot prove a summary reads it fairly. So every sentence carries a footnote to a page you can open, counter-evidence is labelled rather than counted, and nothing a model writes appears on your record until a person accepts it.

5 · The decisions

The machine suggests. Your team decides.

Every hazard is a suggestion until a person includes or excludes it. Drafts are drafts until a person accepts, edits or rewrites them, and only accepted text appears. A draft may suggest a level; only a person sets the score. Re-running the pipeline never touches a decision already made.

Every decision, upload, approval and re-run writes an audit event to a history with no update path and no delete path.

See the approval flow →

  • The page

    The source page with the passage highlighted from the block's own coordinates.

  • The run

    The model (gpt-5.6-terra today), the prompt version (4.6.0), the taxonomy version, the counts and the timings.

  • The prompt

    The exact text the model received for your document, one click from the run.

  • The rejections

    What the run threw away and why, in the same ledger as what it kept.

  • The people

    Who decided what, when, and what it replaced.

6 · The result

Every sentence resolves to its page.

What reaches your brief is the accepted reasoning, each sentence footnoted to the quote and page it rests on. Open the footnote and you are looking at the document itself.

Read the full sample brief →

Likelihood

3 of 5, Possible · verified

The county fire district recorded 64 wildland fires from 2010 through 2024, most under one acre; the largest, the 2018 Pine Hollow Fire, burned 1,240 acres.1 The 2022 wildfire protection plan maps 18,400 acres of wildland–urban interface along the western foothills, with about 3,100 homes within half a mile of continuous fuels.2

Sources for this assessment

  1. 1“Between 2010 and 2024 the District responded to 64 wildland fires, the majority of which burned less than one …”C, p. 4
  2. 2“The WUI boundary encompasses approximately 18,400 acres, concentrated in the western foothills, with roughly 3…”B, p. 12

Worked example: Cedar Ridge County is fictional. The scores illustrate the format and are nobody's decision.

What we don’t do

No training on your documents.

Model calls go to OpenAI's API, which does not train on API data. Every call asks OpenAI not to keep the response, except deep research, which runs in the background and is kept by OpenAI for up to 30 days so an interrupted run can resume.

No model output stored without a verified source.

A mention is a quote, a page and a block, or it is not kept. A drafted sentence cites what it was given, or it is dropped.

No silent numbers.

A skipped feed is recorded as skipped and never reads as zero. A hazard name nothing can place stays unmatched in a queue a person works.

No hidden prompt.

Every run stamps the model and prompt version that produced it, and you can read the exact prompt your document was run with.

Book a walkthrough

Bring a plan. Watch it get read.

Tell us about your jurisdiction. We set up the workspace, you upload a plan you already have, and every claim that comes out of it is cited to a page you can open.

Already have an account? Sign in

Request a walkthrough

Takes about two minutes. A person reads every request.

Define your footprint on the map first