Skip to content
Marapone SpecChecker New

Two contract documents.
One of them is wrong. SpecChecker — specification compliance software that checks the project manual against the drawings

The drawing says 1/2-inch drywall. Section 09 29 00 requires 5/8-inch Type X. One of those is what the inspector reads and the other is what arrives on the truck, and nobody finds out which until the wall is closed in. SpecChecker reads the project manual against the drawing set and reports every place they disagree — before the material is ordered. Opens in a browser on any machine. From $249 a month.

Open it on Pro — $249/mo → See every screen
Or send us one real manual and drawing set and we'll run it for you first →
✓ Opens in a browser ✓ Both documents quoted, both cited ✓ PDF, Word, Excel or a scan ✓ Windows, macOS & Linux · mobile Q4 2026
The real app, recorded · one project manual, two drawing sets, one addendum, 21 conflicts

27

Measurable properties

Ratings, thicknesses, strengths, grades

39

Named assemblies

Doors, AHUs, panelboards, slabs, partitions

11

Conflict types

Including the two a document has with itself

25

CSI divisions

MasterFormat spine, 47 sections mapped

No edit, no voiceover, no time-lapse

WATCH ONE REAL RUN,
START TO FINISH.

The tour above is cut from a seeded demo. This is not: it is a single unbroken recording of the shipped application doing the work, including the parts where you wait.

Want this on your own documents instead of ours? Run it on your own file — 7 days free →

Nothing redrawn for the website

EVERY SCREEN,
AT FULL RESOLUTION.

Screenshots straight out of SpecChecker, not mockups. Click any one to open it at capture size and read the actual figures in the actual cells.

Conflicts by severity, and which orders to hold today
Conflicts by severity, and which orders to hold today
A conflict in full — the spec clause and the drawing, quoted
A conflict in full — the spec clause and the drawing, quoted
The conflict matrix across assemblies and properties
The conflict matrix across assemblies and properties
What was checked, and what could not be read confidently
What was checked, and what could not be read confidently
The project manual, drawings and addenda it cross-checked
The project manual, drawings and addenda it cross-checked
The drafted RFI, one letter per spec section
The drafted RFI, one letter per spec section
The reasoning, with every citation it rests on
The reasoning, with every citation it rests on

Click to enlarge · press Esc to close

What coordination actually costs

Nobody reads 400 pages against 90 sheets.

The project manual and the drawing set are both contract documents. They are written by different people, revised on different cycles, and issued together as though they agree. Checking one against the other is a job nobody has time for, so it is done in fragments — by the PM who happens to notice, or by the supplier who calls to ask which one to price.

The failure is never dramatic. It is a door schedule that shows an unrated leaf where the manual requires 90 minutes. The rating is stamped on a label at the factory, so an unrated door cannot be upgraded on site — it is replaced, and the lead time is measured in months.

The window closes at the purchase order. Until the material is ordered, a contradiction is a question. After it, it is a change order, a delay claim, or a wall that has to come back down.

One finding, as the app reports it

Specification — 90 minutes

“C. Door D-1 at the main exit stair shall have a 90 minute fire rating”
Section 08 11 00 — METAL DOORS AND FRAMES

Drawing — 0 minutes

“D-1 915 x 2135 F HOLLOW METAL HM 0 GRADE 2 MAIN EXIT STAIR”
Sheet A-601 — Door Schedule
Critical Below specified minimum Div 08 · Doors Hold the order

Doors D-1 — fire-resistance rating: 90 minutes required, 0 minutes shown

A rating short of the spec is caught at the fire inspection, after the assembly is closed in. Doors and frames are worse again: the rating is stamped on a label at the factory, so an unrated leaf cannot be upgraded on site — it is replaced, and the lead time is measured in months.

Confidence: High · Action: hold the material order until it is answered
RFI-001 drafted, citing both documents and quoting the wording of each

One of 21 findings on the demo project. Nothing here is written by a model.

Five tools, one job

Measure it. Check it. Price it. Level it. Buy it out.

These are five products because they are five different problems, and most contractors only have one or two of them. Buy the one you're missing — or run all five and the same job carries straight through from the drawing set to the awarded subcontract.

Stage 1

Blueprint Auditor

How much is there — quantities, scope gaps, code and tender risk off the drawing set.

See it →
Stage 2

SpecChecker

Do the documents agree — the manual against the drawings, before anything is ordered.

You are here
Stage 3

AI Estimator

What should we bid — assemblies, contingency, municipal soft costs, margin.

See it →
Stage 4

Bid Leveler

Whose quote is actually cheapest — a dozen formats normalised to one scope.

See it →
Stage 5

ScopeGuard

Did anyone actually price it — every sub's letter against your own documents.

See it →

SpecChecker stands alone. It needs the project manual and the drawing set — nothing else, and no other product in the range. It is the cheapest of the five to justify, because the question it answers arrives on every job whether or not you bid it.

What it looks for

Eleven ways two documents disagree.

Every finding carries both statements, both references, the words each document used, a confidence score, and one sentence on what it costs to get wrong. The counts below are what the demo project actually produced from four PDFs.

Below specified minimum

The drawing is under the floor the manual sets. A 45-minute door where the spec says 90; a 22 kA panelboard where the spec says 42 kA.

9 found
Shown, not specified

The drawings call for something the manual never covers. Nobody wrote the performance requirement, so nobody can reject the submittal.

5 found
Lesser product on drawing

A downgrade — annealed glass where the spec requires safety glass, standard-grade hardware where the spec requires heavy duty.

2 found
Values disagree

Both documents state a number for the same property of the same assembly, and the numbers are different. The plainest kind, and still routinely missed.

2 found
Above specified maximum

Over the ceiling the manual sets — a 2.2 W/m²·K window where the spec caps U-value at 1.60.

1 found
Addendum not carried

A two-page PDF that arrived by email during tender and changed one number, while the manual and the drawings both keep saying the old thing.

1 found
Specified, not shown

The manual requires it and no callout anywhere in the drawing set deals with it. Silence on a binding requirement is not agreement.

1 found
Unit ambiguity

The same number in two different units — R-value against RSI, imperial against metric — which reads as agreement and is not.

none on this job
Stale cross-reference

A clause pointing at a section that was cut, or a note pointing at a sheet that was never issued.

none on this job
Manual vs. itself

A manual written by three people over two years. There is no precedence rule to fall back on inside a single document, so this one has to be asked.

none on this job
Drawings vs. themselves

An old general note on one sheet and the revised one on another. Same problem, and the same answer: put it in writing to the consultant.

none on this job

21 open findings — 7 critical, 7 high, 6 medium, 1 low — of which 13 should stop a material order today. Read from 25 pages across four PDFs: 39 requirements out of the manual, 63 callouts off the sheets, and 7 RFIs drafted and waiting to go out.

Inside the app

Every screen, no mock-ups.

Everything below is the shipping application running one real audit: Northgate Community Centre, a City of Toronto facility at 4120 Northgate Boulevard — one project manual, an architectural set, a structural/mechanical/electrical set and Addendum 02, all issued for construction. Nothing is canned. The reader, the parsers and the cross-check engine are the ones your own documents go through.

01 — Dashboard

Orders to hold, not findings to admire.

The headline number is not the finding count. It is how many material orders should not be released today — because that is the only version of this question a project manager can act on before lunch.

  • ▸A written review that names the worst conflict, the divisions it concentrates in, and how many RFIs would clear it
  • ▸"Deal with these first" pulls the blocking findings to the top with both documents quoted inline
  • ▸Every finding carries what it costs to get wrong and one sentence on what to do about it
02 — The conflict matrix

The table that goes to the consultant.

One row per conflict: division, section, item, property, what the specification says, the clause it came from, what the drawing says, the sheet it came from, the kind of conflict, severity, confidence and the action. It exports to Excel unchanged, which is the form this argument is usually had in.

  • ▸Filter to blocking only, or include what you have already resolved
  • ▸Every row cites the section number and the sheet number — no row without both
  • ▸A version snapshot is written on every cross-check, so what the documents said at tender stays recoverable
03 — Coverage

On a findings list, silence looks exactly like agreement.

This is the screen most products would not ship. An audit can only report a conflict where both documents say something about the same property. Where the drawings are silent there is nothing to disagree with — so the app tells you its own coverage before you treat a low finding count as good news. On the demo: 26 properties specified, 62% had a drawing to compare against.

  • ▸Clearance is shown by package, not by division — nobody releases "Division 08", they release the door package
  • ▸Each package is HOLD or CLEAR, with the properties behind the call
  • ▸A setup checklist blocks the confident-looking empty result: an audit run against one document, or against drawings nobody told the app were drawings
04 — RFIs

One letter per section, quoting both documents.

The app does not decide which document governs — that is a question for the consultant of record, which is exactly why the recommended action is to ask them in writing. One click drafts the letter per section, addressed to the consultant, citing both documents and quoting the wording of each, with a response date.

  • ▸Draft for everything open, or only for what is blocking a material order
  • ▸Grouped by section so the consultant who owns it gets one letter, not eleven
  • ▸Editable before it goes out, and tracked from draft to issued to answered to closed
  • ▸Exports as an RFI package, a PDF audit report, or the whole conflict matrix as Excel

05 — It works out what each file is

Drop the files on the window — PDF, Word, Excel, or a scan. CSI section numbers and PART blocks mean a manual; sheet numbers, title blocks and schedules mean drawings. The guess is shown on screen and you can override it. Every document is copied into the project's own folder, because a finding that cites page 4 is worth nothing once that PDF has moved off somebody's desktop.

06 — It learns what you call noise

Mark any finding real or noise and the verdict persists. Mute a property outright and it stops being compared. The pattern of what conflicts on your projects — which divisions, which consultants, which properties — accumulates inside your own workspace and makes the second audit measurably quieter than the first.

Why the findings hold up

A misquoted clause in an RFI is not recoverable.

Text recognition misreads exactly the characters that matter here: 5/8 as 5/B, 90 as 9O. So every finding carries a confidence, and it is not decoration — it decides where the finding sits on the list and whether it is marked "verify".

Certain

Clean text layer, a numbered spec section, a PART 2 product clause. The strongest thing a document can say.

High

Clean text, or a tagged schedule row under a header the app understood.

Moderate

A general note — read correctly; its scope is what is uncertain — or an inferred subject.

Low

Recognised by text recognition from a scanned page. Shown, sorted below the certain ones, and marked verify — open the sheet before quoting it.

It does not decide who is right

Which document governs is a question for the consultant of record, not for software. SpecChecker's job is to find the disagreement, quote both sides accurately, and put the question in writing to the person who can answer it.

It reads text, not geometry

It does not scale dimensions off a plan or read anything stated graphically. It reads what the two documents say — clauses, schedules, general notes — and compares statements to statements.

A general clause applies broadly, on purpose

It cannot tell that door D-8 is an office door. A clause that names no tag applies to the assembly, so a general 45-minute requirement is compared against every door in the schedule. Some of those are worth confirming and some are not — the RFI asks the right question either way.

You overrule the engine

Every finding can be resolved, dismissed or escalated; every property can be muted; every extracted value edited. Those calls persist and compound across projects — inside your own workspace, never pooled with another firm's and never used to train anything.

The knowledge base

Twenty-seven properties, and you can add the twenty-eighth.

A spec clause and a drawing callout are only ever in conflict about a measurable property of a named assembly — a thickness, a rating, a strength, a grade. Everything the engine does reduces to pulling those out of both documents, putting them in the same units, and seeing whether they agree.

Fire & life safety

Fire-resistance ratings, sprinkler design density, hardware grade, and the exit-stair and corridor assemblies they attach to.

Envelope

Thermal resistance and RSI, U-values, SHGC, glazing make-up, air and vapour barrier requirements.

Structure

Concrete compressive strength, reinforcing steel grade, structural steel grade, compaction requirements.

Interiors

Gypsum board thickness and core type, partition assemblies, door material and hardware grade.

Mechanical

Airflow, external static pressure, filter efficiency (MERV), piping material, and the units they belong to.

Electrical

Short-circuit interrupting ratings, conductor and raceway material, panelboard and distribution equipment.

Commercial

Warranty terms — the clause that costs nothing at tender and is worth arguing about five years later.

Adding a property is data, not code

Every property is an entry in attribute_rules.json — its unit table, the patterns that recognise it, whether more is better, how much difference is a real difference, and one sentence on what it costs to get wrong. Add an entry and the engine checks it on the next run. The CSI spine is MasterFormat, so the taxonomy is the one your estimators already read; the jurisdiction layer is named fields you can swap.

How regions & rate packs work →
From document issue to answered RFI

Five steps, and the last one is the point.

01

Create the project

Name, number, client, consultant of record, and the issue status of each document set. The consultant is who the RFIs get addressed to.

02

Drop in the documents

Manual, drawing set, addenda, exported schedules. Each file is classified by what it looks like; the guess is shown and can be overridden.

03

Run the cross-check

Both sides are read into the same shape — one property, of one assembly, in one canonical unit, with the page or sheet it came from — and compared.

04

Work the findings

Sorted by what stops a material order, not by how many there are. Resolve, dismiss or escalate — and read the Coverage tab before calling it clean.

05

Send the RFIs

This is the step that saves the money. One letter per section, citing both documents, asking which governs — while the answer is still free.

06

Keep the record

A version snapshot from every cross-check. What the documents said at tender, before three addenda landed, stays recoverable — on a claim, that is the whole argument.

MaraponeAI

MaraponeAI never finds a conflict.

Every finding comes from the parsers and the cross-check engine, out of text on a numbered page of a real document, and every one cites where it came from. The model explains, prioritises and drafts the letter.

That line is drawn where it is because an RFI goes on the project record with your name on it. A hallucinated requirement would make you look careless, in writing, on a project you have to keep working on for two more years.

Ask it can I order the doors or show division 08 and the instruction is parsed by rule and previewed before it changes anything on screen.

Privileged working product, treated as such

A contractor's annotations on the contract documents — which clauses they think are wrong, which ones they are preparing to claim on — are privileged working product. They sit in a workspace scoped to your account, are read only by our own model, and are never used to train anything.

The model never finds anything

Findings come from a deterministic reasoning engine, traced to a clause and a page. The model writes the explanation and the RFI draft on top of that — it is never the thing that decided a conflict exists.

Your data outlives the subscription

Projects, every audited document, RFI history, muted properties and a snapshot from every cross-check export on demand in PDF, Excel and CSV. Cancelling closes the programs; it does not take what you made with them.

Our own text recognition

Scanned sheets are read by a recogniser we run ourselves. No third-party OCR service is in the path, and low-confidence reads are labelled rather than smoothed over.

What a plan actually opens

A finished program, open in a browser tab.

One flat price per firm. No per-file charges, no metering, no implementation fee. The working program the minute you sign in — and your first payment back if it doesn't earn its keep.

The arithmetic

One finding on the demo project — a 90-minute exit-stair door shown unrated on the schedule — is a door and frame package reordered after the opening is framed. The rating is stamped at the factory, so the leaf is replaced rather than upgraded, and the lead time is measured in months.

The Pro plan is $249 a month and runs on every set of documents after that one. That is the whole value argument: you are not buying software hours, you are buying the question asked while the answer is still free.

Included in every plan

  • ✓The complete program the moment you sign in — nothing to install, no licence key, no machine to provision
  • ✓Unlimited projects and documents on your plan's seats. No per-project or per-file fee, ever
  • ✓All 27 properties across 39 assemblies, 11 conflict types, and the CSI MasterFormat spine
  • ✓The whole knowledge base as editable JSON — add a property without touching code
  • ✓Every improvement the day it ships — no version to buy, no upgrade to schedule
  • ✓Windows, macOS and Linux. Unlocks this tool on the mobile app, Q4 2026, included
  • ✓Cancel any month from your dashboard, no long-term contract — see the plans

Not included — that’s a customised setup

From $1,500

  • ✗Tuning on your spec templates, your consultants' drafting conventions and your house standards
  • ✗Properties and assemblies specific to what you build — a rule set outside the shipped 27
  • ✗Wiring into Procore, Sage, Bluebeam or your own document management system
  • ✗Workflows the product doesn’t cover — RFI backlogs, submittal logs, deficiency lists
  • ✗Bespoke development beyond the five programs — that is quoted as project work

Buying the product first doesn’t lock you out of a customised setup later — most teams start here, then commission the parts that are specific to them. See all three levels →

Or all five on the Pro plan: $249/mo — the Blueprint Auditor, SpecChecker, the AI Estimator, the Bid Leveler and ScopeGuard, $1,000 less than buying them separately, plus one setup call that runs a real job of yours from the drawing set to the awarded subcontract.

Open it on Pro — $249/mo → Book a walkthrough on your own documents

Run it on the set you're about to order from.

Send us the project manual and the drawing set. We will run them through SpecChecker and walk you through what it found — conflict by conflict, with the clause and the sheet behind each one — before you spend anything.

Open it on Pro — $249/mo → Book a walkthrough first

Billed monthly or yearly · tax at checkout · Secured by Stripe · cancel in one click, any time

SpecChecker reports where two contract documents disagree. It does not decide which one governs, and it does not replace a design review, a code review or professional judgement.

For the person who has to sign off on it

The technical detail, before you ask for it.

Every claim on this page has a page behind it. No form, no call — read it now.

Architecture

Five layers, every component named, and the one design rule the whole thing is built around.

Trust & security

Where the data sits, who can reach it, retention, deletion and the security pack for your IT review.

Integrations

What it reads and writes, and how it sits beside Procore, Autodesk, Sage or a shared drive without replacing them.

Regions & rate packs

Why jurisdiction is data in editable files rather than buried in code, what a finished pack contains, and how your province, state or country gets built.

Plans & pricing

What each plan opens, how many seats it covers, and how cancelling from your own dashboard works.

FAQ

Support, cancelling, exports, data deletion, and what happens if we disappear.

Real jobs, real numbers

Three documented builds

6–12h/week recovered, RFI turnaround 9.2→5.4 days, +24% win rate in 60 days — each with the build cost, the break-even and the quote it came from.

Your numbers, not ours

ROI calculator

Put in your seat count, hourly rate and document volume. It will also tell you when a subscription is the better buy.