Project summary

I'd build a tool that lets writers and podcasters see exactly which claims in an AI answer drew on their work, check the original passage, and flag missing or wrong credit. Readers should be able to go straight to the original.

The first version would cover sources whose use was observed or recorded within a participating publishing workflow. When that history is unavailable, the receipt would show it as unknown.

Dediro has two public reading samples at https://dediro.com/. A separate manual example follows a claim about local institutions in Gaza back to its source and shows where the article's wording went further than the passage: https://docs.google.com/document/d/14tCM-sYziBVIL-qRLGoFd-JtHpj8yBxFp9GjNulCaCI/preview.

The proposed tool would make that check repeatable and portable. Each receipt would keep the claim version, the original work and creator, a source passage and locator, the basis for recording source use, and the editor's assessment. It would distinguish material retrieved during research from material recorded as informing a claim. A passage match would establish that the text occurs in the supplied source. An editor would still decide what the passage supports.

What are this project's goals? How will you achieve them?

At the minimum funding level, I would spend eight weeks specifying the receipt format and building an offline evidence viewer with editing and export. Another developer could produce a compatible evidence file. A reader could inspect it without a paid account or a connection to a model provider.

At the full funding level, I would add four weeks for a creator correction workflow and an outside test. A creator could flag wrong credit, and a human reviewer would record the decision, reason and prior version. I would recruit an independent editor or operator to try an exported receipt and report whether it can be reopened and checked. Participants are recruitment targets.

How will you know it worked?

I would publish a cleared set of at least 30 claim-source examples, including at least 10 mistaken, missing or uncertain connections. The tests would include changed source versions, copied material and citations that do not support a claim. The report would separate passage matching from judgments about source use and support, and show false matches and missed cases.

Any synthetic fixtures would be labeled and counted separately from examples drawn from real source material.

For the full project, I would also report the outcome of creator correction attempts and the outside import test. If I cannot recruit an outside tester or a feature fails, the report will say so.

How will this funding be used?

I am seeking a minimum of USD 3,000 and a maximum of USD 10,000. These are prospective allocations, with no vendor or participant committed.

Minimum scope: USD 2,400 for development and editorial design time, USD 300 for usability testing, and USD 300 for documentation and accessibility review.

Full scope: USD 6,000 for implementation and editorial design, USD 2,000 for independent testing and source-credit review, USD 1,000 for accessibility and documentation, and up to USD 1,000 for necessary compute, storage and source permissions. I would use existing equipment and freely reusable sources where possible, account for actual costs and follow the agreed grant terms for unused funds.

Who is on your team? What's your track record on similar projects?

I'm Farzad Khosravi, founder of Dediro. I have been developing the project since 2024. The public reader and manual source trace are work you can inspect today. I would lead the project and recruit the independent reviewer or operator after funding. The creator-facing receipts and correction workflow are proposed work.

The code, format, documentation, cleared fixtures and evaluation would be public and openly licensed. Source works would be linked, and excerpts redistributed only with an appropriate basis. Readers and other tools could use the released materials independently of Dediro.

What are the most likely causes and outcomes if this project fails?

Source use can be unclear, even when a passage looks relevant. The receipt needs to preserve that uncertainty. It cannot establish hidden model-training or generation history from a plausible text match. The prototype would record what the workflow observed and leave unknown uses visible.

Creators may find the correction step too burdensome, outside reviewers may be difficult to recruit, and source permissions may limit examples. I would keep the initial scope small, use cleared or labeled synthetic fixtures, and publish the results and remaining gaps even if the broader workflow does not earn continued use.

How much money have you raised in the last 12 months, and from where?

Dediro has raised USD 0. I am pursuing other grant support for related work, including RJI, Mozilla and FUTO. Any overlapping award would be disclosed. I would reduce this request or agree distinct costs and deliverables before accepting funds for the same work.