See what the other side’s redline really does to your lease.
Three frontier models from three companies read the whole lease separately. Every finding is tied to the lease’s own words, and you see where they disagree.
A real lease, a real redline. Each change is read against the whole document: whom it favours, what it does to your risk, why, with a link to the lease.
A change that looked like a win, and was not
The other side deleted the clause letting the landlord demand all remaining rent on default. Alone, that helps the tenant. Read against the whole lease, two of the three models found a confession-of-judgment clause that still lets the landlord take judgment for it.
“…the Rent for the entire unexpired balance of the term.”
Checked against the lease word for word. A quote not in the lease does not count.
A look inside
Choose what to review. A new lease, or a redline that came back, read clause by clause against the whole lease. Then whose side you are on.While it runs. A progress checklist, the stage each model is on, and a place to ask questions while you wait.The findings, counted. Severe risks, risks, needs review, improvements, no action needed, and a review you work through in order.The issues, worst first. The sidebar keeps the open issues in view while you read, each with its section and status.A risk, with a redline ready. The clause from the lease, the model's suggested wording marked as not yet accepted, and your review buttons.One checklist topic, opened. Each question, the answer, and where the three models split: Claude yes, GPT no, Grok no.What each model said. Three models, side by side, on one change. Every quote was checked against the lease; the tick means it was found.Where they disagreed. Where the three models did not give the same answer, shown rather than averaged away.
Screens from reviews of the same public lease: the full review, and the review of a redline.
Two ways to use it
About an hour
A redline that came back
Every change, including ones they did not mark, read against the whole lease: the risks, what to ask for, and draft wording for your reply.
Two to three hours
A new lease
Every checklist topic, the requirements, the one-sided terms, what to ask for.
Times are from the lease we measured; longer leases take longer.
Four companies’ models read your lease and do not train on it. We keep it in your account until you delete it, and only your account can open it. Where your lease goes, in full.
This is early. Measured carefully on one lease, tried by a handful of lawyers, not legal advice. Tell us what it got wrong.
Sign in
First time with this email? Enter the access code you were given. You only need it once.
We sent an email to . Click the link in it, or type its 6-digit code here. Either works once, for 15 minutes.
Upload your standard lease, or the version you sent to the other side.
Reading the file...
The earlier version(one file)
📄
Drop the earlier version here or click to browse
PDF, DOCX, or TXT
The earlier version is:
Email me when analysis is complete (optional)
Email addresses don't match
▸ More options — how the models suggest changes, special instructions
How the models suggest changes
Used for "what to ask for" and every model-drafted change. You can change it later on the Your Changes tab.
Review Areas
Not shown (Step 1058).
Standard Review Areas
Automatic drafts are suggestions: one model drafts, two others check. Nothing goes into the Word redline until you accept it. Makes the review take longer.
Reviewed by three frontier models, checked against your lease's own words. Patent pending.
CAM™ Intelligence
AI-generated guidance — not legal advice.
Live analysis
CAM is reviewing your leases
Completed contracts unlock as soon as they finish, so you can begin reviewing while CAM continues processing the remaining leases.
Follow each major step as CAM works through the review.
📧 Get notified when done (optional)
Analysis takes a few minutes. Enter your email and we'll send you this same return link when it's ready.
Email addresses don't match
You can wait here and watch progress update, or safely close this page and come back using the link below.
Ready now
Leases ready to review
Open completed contracts immediately while CAM continues processing the rest.
Coverage check
Review Areas
CAM is checking all 32 standard issue areas.
Stage-by-stage progress
Contract status by stage
Ask while you wait
AI-generated guidance — not legal advice.
🔍
What CAM is doing right now
CAM reads the lease, splits it into numbered paragraphs, and has three models from three different companies answer the same questions about it, each on its own.
📋
What you'll get in results
For a full review: the risks to your side, worst first; what the lease requires you to do; and what to ask to have changed, with suggested wording. For “See what changed”: each change, whether it helps or hurts you, and why.
🧠
Why CAM exists
Most AI tools give one confident answer even when the evidence is thin. CAM asks three models the same questions and shows you where they disagree.
⚖️
What CAM does differently
A quote only counts if CAM finds those exact words in your lease. When the models disagree, you see it, and which model said what, instead of a hidden vote.
🎯
Applied here: commercial leases
CAM's checklist covers what a commercial real estate lease should deal with, from rent and operating costs to default, insurance and assignment.
🔎
Why the trail matters
If a finding looks wrong or surprising, you can trace it: the lease's own words it rests on, what each model answered, and where they disagreed.
🎚️
Disagreement is information
When the three models answer a question differently, that question deserves a closer human look. CAM keeps the disagreement visible instead of averaging it away.
💬
Ask questions while CAM works
Use the chat panel while CAM works to ask what it is checking, what the results will show, or what to look at first once your review is ready.
💻
View Results on Desktop
The full analysis view is optimized for desktop browsers.
Or email yourself the link:
Menu ▾
Overview
Open Issues
Generating portfolio summary…
Critical — Immediate action required High — Significant deviation, review recommended Medium — Moderate deviation, monitor Low — Minor deviation, informational Conforms — Aligns with reference lease
Primary Order:
Lease Summary
Deviations (Review Recommended)
Additional Provisions
Provisions identified outside the evaluated scope with no standard provision match.
Export audit trail:
Ask About This Analysis
Scope
Models:
Response:
Synthesized by:
AI-generated analysis — not legal advice.
How CAM™ Intelligence Analyzes Your Lease
How CAM analyzes leases — two modes
CAM stands for Constrained Assertion Method, Vered.ai’s framework for evidence-aware AI analysis. In lease text, Common Area Maintenance (CAM) is a separate lease issue area.
CAM operates in two analysis modes. The behaviors described on this page apply differently depending on which mode produced your report:
Coverage Analysis (single document) — When you upload a single lease without a reference template, CAM runs Coverage Analysis. This mode uses CAM’s schema-governed assertion architecture to evaluate the lease against a curated library of expected provisions. Each issue area is classified by materiality (nothing adverse found, worth reviewing, needs attention) and surfaced with exposure prose framed for the perspective you selected (tenant, landlord, or neutral). Coverage Analysis surfaces what’s missing, what’s incomplete, and what’s structurally asymmetric — without requiring a reference document to compare against.
Comparison Analysis (lease vs template) — When you upload a tenant lease alongside a reference template, CAM runs Comparison Analysis. This mode uses CAM’s full multi-evaluator pipeline: independent evaluator roles, challenger review, severity assessment, and the disagreement governance described below. The confidence indicators, Finding Reliability scores, and Assertion Sensitivity Gap all apply to Comparison Analysis findings.
Contract Interaction Review — After Coverage Analysis completes, CAM runs a third pass that reads the entire lease as a whole rather than provision by provision. This pass looks for three things: whether a missing provision’s substance is supplied by another clause elsewhere in the document; whether cure, remedy, or rights language is running in favor of the correct party; and whether two or more provisions combine to create exposure that neither reveals on its own. This layer operates strictly within the four corners of the lease — it does not credit common law, background doctrine, or implied terms. Findings appear in the Contract Interaction Review tab and are flagged on individual provision cards where they apply.
The remainder of this page describes Comparison Analysis behavior in detail. Coverage Analysis surfaces issues based on schema-driven classification with materiality grading rather than multi-evaluator agreement; the confidence-score machinery described below does not apply to Coverage findings.
👀
Most AI tools hide disagreement. CAM surfaces it.
When AI evaluators disagree about a clause, conventional tools quietly resolve that disagreement and deliver one confident-sounding answer. You never see the uncertainty. CAM works differently: each provision is evaluated independently by multiple AI roles, without any of them seeing the others’ conclusions. Where they agree, the finding is reliable. Where they don’t, that disagreement is shown to you — because a split evaluation is itself important information.
📊
Most AI tools ask the model how confident it is. CAM doesn’t — because that answer isn’t reliable.
AI models routinely express high confidence in wrong answers. A model’s self-reported certainty is not a meaningful signal — it reflects how the model “feels” about its answer, not whether the answer is actually well-supported. CAM doesn’t ask evaluators to rate their own certainty. Instead it measures things that are independently observable: did all evaluators reach the same conclusion without seeing each other’s work? Is the reasoning grounded in explicit clause text, or inferred? Did the finding survive a dedicated challenge step? Does the clause depend on how other parts of the lease are read? The scores you see in CAM reflect those factors — agreement, evidence quality, reasoning completeness, and structural complexity — not a self-assessment.
⚖
CAM knows when not to assert.
Most AI systems are built to always give you an answer. CAM is built around a different question: “Is there enough evidence to assert this conclusion reliably?” When the answer is no — because evaluators disagreed, or the clause is too context-dependent to characterize definitively — CAM withholds the finding rather than guessing. A withheld finding is not a failure. It is the system being honest about the limits of what it can confidently conclude.
How the pipeline works:
Unlike single-AI tools, CAM uses a multi-stage evaluation pipeline designed for accuracy and transparency:
Multiple AI evaluation roles independently analyze each provision without seeing each other’s conclusions — producing genuinely independent assessments, not a consensus.
An automated rules engine checks for specific deviation patterns: exception clauses, definition changes, numerical differences, obligation shifts, and more.
A challenger role probes flagged findings to test whether the deviation is substantive or cosmetic, and to surface hidden dependencies.
Severity is assessed based on legal and financial impact, so you know what to address first.
Each finding shows you exactly what each evaluator concluded, which rules triggered, and how the final verdict was reached. Full transparency, not a black box.
Confidence indicators
After evaluating each clause, CAM assigns a confidence label. These appear in the sidebar and on every finding card:
●●●● Verified — All reviewers agreed. The finding is reliable and well-grounded in the clause text.
●●●○ Impact Unclear — Finding is confirmed, but the practical impact depends on how specific terms or cross-references are interpreted. Read the interpretation note.
●○○○ Inconclusive — Insufficient basis to assert a finding. Treat as a flag, not a confirmed deviation.
How the confidence scores work
Each provision receives three diagnostic scores, visible in the expanded audit detail:
Finding reliability
How much can you rely on this finding? Calculated as A × E × R × 100 where:
A (Agreement) — Did all three reviewers reach the same conclusion? Unanimous (3–0) = 1.0, majority (2–1) = 0.75, no consensus = 0.5.
E (Evidence) — Is the conclusion grounded in explicit clause text, or inferred? E ranges from ~0.65 (weak, inferred) to ~0.95 (strong, grounded in direct text). If the challenger confirms the deviation as substantive, E rises; if it finds the difference cosmetic, E falls. This is why two findings with identical reviewer agreement can have very different reliability scores — one may be anchored in exact clause language while the other depends on interpretation.
R (Reasoning) — Did the finding go through every stage of the pipeline? R ranges from ~0.80 (challenge skipped) to 1.00 (full chain: extraction → evaluation → challenge → severity). Hidden dependencies — cross-referenced sections, cascading definitions — also reduce R because they make the reasoning path harder to verify independently. Most DEVIATES findings score 0.90 or above; lower values flag provisions where the pipeline had to make assumptions.
A score of 80+ is strong. Below 50, the finding is withheld rather than surfaced.
Stricter review (drop score)
How does this hold up under the most conservative reading? CAM recalculates the score with stricter evidence standards and applies a structural fragility penalty. The “drop” is the difference — called the Assertion Sensitivity Gap (ASG).
A small drop (ASG < 15) means the finding is stable regardless of how strictly you read it.
A large drop (ASG > 28) means the finding looks persuasive but weakens under scrutiny — this is what triggers “Impact Unclear.”
This is the key innovation of CAM. Traditional AI tools report only one confidence score. CAM reports two — exposing findings that appear reliable but aren’t.
Clause complexity
How tied is this clause to other parts of the lease? Factors include:
Internal cross-references (e.g. “subject to Section 9.1”)
Definition overrides (a defined term was changed elsewhere)
Negation or conditional scope changes
High complexity doesn’t mean the finding is wrong — it means the clause’s meaning depends on context that may require expert interpretation.
My reviews
Kept until you delete them. Deleting removes the lease, the findings and the files.
Where your lease goes, and for how long
Checked on 14 October 2026 against our code and each company’s published terms. This is a review service for a small number of invited users.
Who reads your lease
Your lease is sent, in full, to models at four companies. Each reads it separately; none of them is shown another’s answers. The findings you see are built by our own code from their answers.
Anthropic (Claude) — keeps what we send for up to 30 days, for abuse monitoring, then deletes it. Not used to train models.
OpenAI (GPT) — keeps an abuse-monitoring log for up to 30 days. Not used to train models. We have turned off every sharing and logging option the account offers.
xAI (Grok) — keeps requests and responses, encrypted, for 30 days, for abuse monitoring, then deletes them. Not used to train models. xAI does not promise a processing region.
Google (Gemini) — keeps prompts and outputs for 55 days, for abuse monitoring; authorised Google staff may review flagged content. We use the paid service, under which Google does not use your content to improve its products.
These are the companies’ own terms, and they can change them. Links: Anthropic · OpenAI · xAI · Google.
What we keep, where, and for how long
Your lease, the findings, and every file made from them (the marked-up PDF, the Word redline) are kept on an encrypted disk in a data centre in Virginia, USA, in your account until you delete them. Reviews started before accounts existed are removed 7 days after the review finishes.
Delete sooner with the Delete button on the review page. It removes the lease, the findings and the files, and tells you if anything could not be removed.
What we keep after that: counts and timings only — how many questions, how long each stage took, what it cost — with no text from your lease.
The emails contain nothing from your lease and no link that opens a review: the review-ready email links to the review, which opens only when you are signed in, and the sign-in link works once and expires after 15 minutes. A copy kept by any mailbox, ours included, opens nothing.
Our servers’ logs record which review ran and when, not what it said.
Who can open a review
Only your account. A review opens only for the account that started it, signed in. Its address opens nothing for anyone else, so a forwarded link shows them nothing. There is no way to share a review yet.
Every review belongs to an account. One access code, shared by invited users, is needed once, to create an account. It is never stored on our servers in readable form.
Accounts. If you sign in, your reviews are listed under “My reviews” and open from any browser you sign in on. Signing in is a one-time link sent to your email; we store no password. Your account holds your email address and the dates you signed in. Deleting your account deletes every review in it.
What we do not do
We do not use your lease for anything except your review.
We do not train anything on it, and we do not let the model companies do so.
We do not look at it unless you ask us to.
This is not legal advice about your own confidentiality duties. If you need a written agreement before uploading a client document, ask us before you upload.
Create an Evaluation Rule
This rule will be applied to future runs. Edit the suggestion below to be as specific as helpful.
My Evaluation Rules
These rules are injected into the AI evaluation prompt on future runs. Rules apply to all future analyses using this access code.