Explain Claims Submit Organizers FAQ Pricing Contact Start researching
New in v2: public claims and pre-submission checks

Research you can trace

Develop research with AI, compare independent results, and trace every claim to its evidence. Use it here, or run it on your own machine.

Transformer reproduction · Paper diff
Abstract · p. 1

It achieves 28.427.928.4 BLEU on WMT 2014 English-to-German.

Switch to v2 to see a change
Understand

Any paper, at your level

Import from arXiv and pick who it's for. Every point cites the paper.

Attention Is All You Need

arXiv 1706.03762
Example

The Transformer drops recurrence and uses , so every token attends to every other token in parallel.

and positional encodings stand in for the word order an RNN gets for free. It reaches 28.4 on WMT 2014 English–German.

Tap a term or citation
Public claims

Claims from 100 ML papers, open to reproduction

Mearc lists 2,164 claims from 100 widely read arXiv papers, each with its own link. A reproduction counts toward a claim only after someone other than its author reviews it.

  1. 1

    Record your result in your Mearc project.

  2. 2

    Publish it as a file Mearc shows you in full first.

  3. 3

    Review by a second person, and then it counts.

Browse the claims
Attention Is All You NeedarXiv 1706.03762 clm_07eb8978No reproduction recorded “We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.” Abstract · automated extraction, not reviewed ReproducedPartially reproducedNot reproducedMixed results
Before you submit

Check your paper against its own evidence

Mearc reads your draft against the claims your team reviewed, finds the venues it fits, and turns each gap into a task.

Evaluation

Superseded wording

Your draft · Results“Label smoothing of 0.1 consistently improves BLEU while hurting perplexity across all model sizes.”

Reviewed claim“… improved BLEU (25.4 ± 0.2 vs 24.6 over three seeds), but a second lab's base-model run did not show the gain.”

Task: adopt the reviewed wording

Venue match

Workshop track
Quality
73
Readiness
67%
Topics
Reproduction, seeds
  • Deadline open
  • Within the page limit
  • Double-blind ready
  • AI-use disclosure missing

Snapshot

Anonymized
  • Emails and team names removed
  • Repository link removed
  • Acknowledgments removed
  • Self-citations flagged for you

Fixed by its SHA-256 hash

Verify

Try to rewrite history

Every event is hash-chained to the one before it. Edit one and see what happens.

    Hash chain intact
    Features

    One record for people and agents

    For workshop organizers

    A verification report for every submission

    Point Mearc at a folder of submitted PDFs and your call's policy. It writes one page per paper and a table for the whole event. No PDF is uploaded and no AI model is used.

    • The reproducibility items your call requires
    • AI-use disclosure, checked against your policy
    • A SHA-256 fingerprint of every file
    Ask about a pilot

    Git records which lines changed. Mearc records why.

    How it works

    Three steps

    1. 1

      Import

      An arXiv ID and a repo link.

    2. 2

      Collaborate

      Agents propose. People decide.

    3. 3

      Share

      An approved Research Passport.

    FAQ

    Questions

    Where is my research stored?

    On mearc.app, your account and research records are stored in a PostgreSQL database and your uploaded papers in private file storage. If you run Mearc on your own computer instead, its database and PDFs stay in local files there.

    Does my work go to an AI provider?

    Only if you connect your own Claude, OpenAI, or OpenRouter key. Importing from arXiv sends just the paper's ID to arxiv.org.

    Can I use Mearc without an API key?

    Yes. Every workflow runs offline, and AI output is clearly labeled as offline and high uncertainty.

    How accurate are Mearc's checks?

    Each check is scored on a blind benchmark of papers with known answers. A check is marked validated once at least 90% of its flags are correct across 20 or more flagged cases; until then the app labels it experimental. Claim extraction scores 0.87 today, so it is still experimental.

    Can I add a reproduction to a public claim?

    Yes. Record your result in Mearc, choose Publish this result, and Mearc shows you the exact file, with no project details or email addresses in it. A maintainer other than you reviews it before it counts toward the claim's status.

    Does Mearc prove who had an idea?

    No. It records who did what and which evidence they cited. That helps with credit and review, but it is not proof of original thought.

    What does early access cost?

    Nothing, and no credit card. You can export your full research record at any time.

    Early access

    Be one of the first

    We're inviting AI researchers in small groups.

    • Free, no credit card
    • Runs on your machine
    • Export anytime

    Start with a research question.

    Create a free workspace, develop an evidence plan, and write your first draft.

    Start researchingBrowse public claims

    Send feedback or report a problem

    Contact

    Talk to us

    Questions, feedback, a workshop pilot, or a problem with your account. Write to us and a person on the Mearc team will reply, or book a Mearc product call.