Skip to content
Private AI for consumer protectionBamberg, Germany · Research in practice

Clarity before
you agree.

PANDA / LegalSAN

Understand your rights. Spot manipulation. Take action. PANDA is building an AI assistant for consumer protection, from the terms you accept to the problems you need to resolve.

It starts with LegalSAN: our compact, private AI model that flags potentially unfair contract clauses on your own device. Our browser extension prototype now also marks dark patterns while you shop, warns about price changes before you pay and shows where to cancel a subscription.

01 / Reading between the linesIllustrative clause · not a live result
“We may terminate your account at any time, for any reason, without notice.”
Potential concern / 01

Unilateral termination

The provider reserves the right to end your access without explaining why or warning you first.

See what PANDA finds
24 MBOn-device model
08Clause categories
0 uploadsLegalSAN scans stay on your device
Supported byEXISTUniversity of Bamberg

01Live demonstration

Read it. Then decide.

Choose a sample contract or paste your own text. The model downloads on your first scan and processes the clauses inside this browser.

A file-sync service with the clause stack you meet most often.21 clauses detected

Nimbus Drive Terms of Service 1. Acceptance of these Terms By accessing, browsing, registering for or otherwise using Nimbus Drive in any manner, you acknowledge that you have read these Terms and agree to be bound by them. If you do not agree, you must stop using the Service immediately. 2. Changes to the Service and to these Terms We reserve the right, at our sole discretion, to modify, suspend or replace these Terms at any time. Revised Terms take effect as soon as they are posted on our website, and your continued use of the Service after that date constitutes your acceptance of the revised Terms. We are not obliged to notify you individually of any change. 3. Your content You retain ownership of the files you upload. We may, however, review, refuse to display, remove or delete any content at our sole discretion and without prior notice, including where we believe in good faith that it violates these Terms. We have no obligation to store your content or to return any copy of it to you after removal. 4. Suspension and termination We may suspend or terminate your account and your access to the Service at any time, with or without notice, for any reason or for no reason, at our sole discretion. Fees already paid are non-refundable in the event of termination. You may close your account at any time from your account settings. 5. Service availability The Service is provided on an "as is" and "as available" basis without warranties of any kind, whether express or implied. We do not warrant that the Service will be uninterrupted, timely, secure or error-free. 6. Limitation of liability To the maximum extent permitted by law, Nimbus Drive, its affiliates, officers and suppliers shall not be liable for any indirect, incidental, special, punitive or consequential damages, including loss of data, loss of profits or business interruption, arising out of or in connection with your use of the Service. Our total aggregate liability to you for all claims shall not exceed the greater of the amount you paid us in the twelve months preceding the claim or twenty euros. 7. Governing law and venue These Terms and any dispute arising out of them shall be governed by and construed in accordance with the laws of the State of Delaware, without regard to its conflict of law provisions. You agree that any claim shall be brought exclusively in the state or federal courts located in Wilmington, Delaware, and you consent to the personal jurisdiction of those courts. 8. Dispute resolution Any dispute, claim or controversy arising out of these Terms shall be resolved by final and binding arbitration administered under the rules of a recognised arbitral institution. You waive any right to a trial by jury and any right to participate in a class, collective or representative action. 9. Support We aim to answer support requests within two business days. Security incidents affecting your data will be reported to you and to the competent supervisory authority without undue delay.

Research prototype. Flags are signals for closer reading, not legal conclusions.

02The technology

Small enough to stay yours.

LegalSAN brings a focused language model to the browser. Local processing makes privacy a property of the product.

The local advantage

Your text.
Your device.
Your decision.

Explore the research
01

Separate the clauses

The document is divided into individual clauses for analysis.

02

Recognise the patterns

A transformer encoder scores each clause against eight categories of potentially unfair terms.

03

Make the concern visible

Flagged passages are highlighted in context, with a category and an explanation.

Architecture
6-layer transformer
Weights
int8 · 24 MB
Context
128 tokens per clause
Execution
WebAssembly · Web Worker

03Research & performance

Focused intelligence. Measurable results.

A compact model trained for the fine print. LegalSAN brings specialised contract analysis to the browser, with the research behind it documented in our working paper.

95.88%

Overall accuracy

Reported prior LegalBench evaluation.

88.46%

Binary macro-F1

Fair versus potentially unfair · prior LegalBench run.

These reported LegalBench figures come from a different evaluation than the paper’s 1,607-clause UNFAIR-ToS study. Binary macro-F1 averages both classes; the paper’s binary F1 measures the unfair class. Section 5.3 explains the distinction and the missing records needed to verify the prior run.

Model comparison

Small by design.

Model size · millions of parameters

  1. DistilBERT2.8× the parameters
    66 M
  2. Legal-BERT4.7× the parameters
    110 M
  3. RoBERTa-large15.1× the parameters
    355 M

24 MBLegalSAN download · INT8

Focused on the fine print. Compact enough to run on your device.

Bars use a shared linear scale and compare parameter counts, not accuracy or download size. Detailed performance comparisons belong in the preprint.

Performance & scale

A small model. A strong signal.

Macro-F1 · logarithmic parameter scale

Swipe horizontally to explore the full chart.

Macro-F1 versus model parameter countLegalSAN: 23.5 M parameters, 0.787 macro-F1. DistilBERT: 66 M parameters, 0.793 macro-F1. Legal-BERT: 110 M parameters, 0.830 macro-F1. RoBERTa-large: 355 M parameters, 0.812 macro-F1. Gemma 2 27B: 27 B parameters, 0.565 macro-F1. Llama 3.1 70B: 70 B parameters, 0.581 macro-F1. Llama 3.1 405B: 405 B parameters, 0.577 macro-F1. Shown separately because parameter counts are undisclosed: GPT-4o mini 0.454, Claude 3.5 Haiku 0.446, GPT-4.1 nano 0.374. LLM scores are from the LegalEvalHub unfair_tos leaderboard; encoder scores are historical figures. Evaluation protocols differ, so this is not a controlled head-to-head comparison.MACRO-F1SIZE UNDISCLOSED0.000.200.400.600.801.0010 M100 M1 B10 B100 B1 TPARAMETERS · LOG SCALELegalSAN0.787 · on-deviceDistilBERT0.793Legal-BERT0.830RoBERTa-large0.812Gemma 2 27B0.565Llama 3.1 70B0.581Llama 3.1 405B0.577GPT-4o mini0.454Claude 3.5 Haiku0.446GPT-4.1 nano0.374Not positioned on the size axis

LLM scores are f1_macro from the LegalEvalHub unfair_tos leaderboard; encoder scores are historical figures from our earlier comparison. Evaluation protocols differ, so this is not a controlled head-to-head benchmark, and these macro-F1 scores are distinct from the LegalBench headline metrics above. Models with undisclosed parameter counts are shown separately. Full methodology will accompany the preprint.

Inside the model / Distillation

436 MB to 94 MB.

78.4% smaller FP32 weights

A smaller model learns from Contracts-BERT, then fine-tunes for eight clause categories. The best reported student reaches 0.7691 macro-F1 versus the teacher’s 0.8115: a reduction of 0.0424.

Weight sizeFP32 · 0–436 MB
  1. Contracts-BERTTeacher · fine-tuned435.6 MB
  2. Distilled studentB_attn_6L · best learning rate94.1 MB
Classification performanceMacro-F1 · 0–1
  1. Contracts-BERTTeacher · fine-tuned0.8115
  2. Distilled studentB_attn_6L · best learning rate0.7691

Source: supplied training model card. Macro-F1 averages eight category F1 scores; it is not accuracy. Each chart starts at zero. FP32 weight sizes are separate from the INT8 browser download and runtime memory.

The experiment behind the numbers

A model card records the model’s origin, training setup and evaluation results. Download the experiment record to inspect the details.

Download model card (JSON)

04What it recognises

The details that deserve a second look.

01

Limitation of liability

The provider caps or excludes its own responsibility for damage it causes you.

Watch for Blanket exclusions of "any and all" damages, or caps set at a token amount.

02

Unilateral termination

Your account can be suspended or closed at the provider’s discretion, without cause or notice.

Watch for "At any time", "without notice", "for any reason or no reason".

03

Unilateral change

The provider may rewrite the contract later and treat your continued use as consent.

Watch for Changes effective on posting, with no right to reject them.

04

Content removal

Your content can be deleted or hidden at the provider’s discretion, sometimes irreversibly.

Watch for Removal "without notice" combined with no obligation to keep a copy.

05

Contract by using

Merely browsing or using the service is deemed acceptance of the whole agreement.

Watch for Consent inferred from access rather than from an explicit agreement.

06

Choice of law

A foreign legal system is imposed, which may be far less protective than your own.

Watch for Law of the provider’s home state, chosen without regard to where you live.

07

Jurisdiction

Disputes must be brought in distant courts, making enforcement impractical for you.

Watch for Exclusive venue in a single county or country far from the consumer.

08

Arbitration

Court and class actions are replaced by private arbitration chosen by the provider.

Watch for Jury-trial and class-action waivers bundled into the arbitration clause.

05 In your browser

Protection where you click.

The LegalSAN browser extension works like an ad blocker, but for manipulation. It checks pages while you browse and needs no account. Rules and models run on your device, so no page text leaves it.

  1. 01On by default

    Unfair terms

    On English terms-of-service pages, the clause model from the demo above flags passages that deserve a second look.

  2. 02On by default

    Dark patterns while you shop

    Pre-ticked add-ons, guilt-trip decline buttons, countdowns that restart, stock and activity claims, and automatic renewals are marked on the page.

  3. 03On by default

    Before you pay

    At checkout, the order summary is compared with what you saw earlier: new fees, a subtotal that grew, or an item that now costs more.

  4. 04Opt-in

    Checkout guard

    When a checkout has warnings, the first click on Place order or Pay now is held until you choose Go back or Pay anyway.

  5. 05On by default

    Cancel a subscription

    On account pages, it points to the controls that lead to cancelling, wording that makes cancelling hard, and direct links for big services. It never clicks for you.

Chrome prototype, currently being hand-tested on real shops and subscriptions. Shopping checks cover English pages. An experimental dark pattern text model can be switched on in the side panel.

06 The PANDA roadmap

From understanding to action.

Consumer protection reaches beyond the fine print. LegalSAN is our foundation for an assistant that helps you recognise unfair terms, detect manipulative design and act on your rights.

  1. 01 / UnderstandLive prototype

    Clarity in the contract.

    LegalSAN highlights potentially unfair clauses across eight categories. A compact model, running locally in your browser, puts the fine print in context.

    Today Scan terms before signing up for a service.

    Try the working model
  2. 02 / DetectPrototype · In testing

    See through dark patterns.

    Protection in the browsing flow. Our browser extension marks interfaces that pressure, mislead or obstruct you on shopping pages, and explains what each one does.

    In the prototype Pre-ticked add-ons, fake countdowns, guilt-trip buttons and automatic renewals, found by rules from published research. A compact text model is being evaluated on real pages.

    See the extension
  3. 03 / ActStarted · First steps

    Help that follows through.

    An agentic service to help handle contract cancellations and everyday legal issues: prepare requests, submit them with your approval and track the follow-up.

    First steps, in the prototype A summary of what changed before you pay, an opt-in guard that holds the pay button while there are warnings, and a helper that shows where to cancel a subscription.

    What we’re working toward Cancel an unwanted subscription, dispute a charge or organise evidence for a complaint, with a path to qualified legal support when needed.

Available today: the LegalSAN clause scanner on this page. Dark pattern checks, checkout warnings and the cancellation helper run in our browser extension prototype, which is in testing. Services that act on your behalf are planned; launch dates are not yet announced.

Make an informed choice

Start with
the fine print.

Scan a contract