Evidence-grounded discovery for accountability work
LLMs paraphrase, approximate, and hallucinate citations. I built a tool that only surfaces what it can validate character-by-character against source documents — verbatim excerpts, exact page numbers, zero inference.
Built for journalism, oversight, investigations, and research. Currently a working prototype — looking for pilot collaborators.
What I built it to do, what I deliberately kept out, and where your judgment still matters.
The validation is bulletproof. The storytelling is still yours.
Why this exists and how you can help steer it.
I'm Andrew. I built this because I needed it.
I volunteer on boards and do governance work in my spare time. Last year I had to review ten years of board papers, audit reports, and committee transcripts to understand how a major project went wrong.
I tried ChatGPT. It found things that were "close enough" — paraphrased quotes, approximate summaries, contextually similar passages. But never verbatim text. When I asked for the exact page number, it confidently pointed to the wrong place.
I tried Claude with extended context. It could handle more documents but had the same problem: it would synthesize, it would interpret, it would "helpfully" rephrase. I still had to manually verify every single excerpt.
The problem was not that LLMs are bad at reading — they are incredible. The problem was that for accountability work, "close enough" is not good enough. If I am going to cite something in a report or put it in a story, I need the exact words, the exact page, and I need to know it is real.
So I built Accounter. It only surfaces what it can validate character-by-character. No inference. No paraphrasing. No helpful summaries. Just: "This exact text appears on this exact page of this specific document."
I built it in a month. It works for my use case. Now I need to know if it works for anyone else's.
I need real document sets and candid feedback.
Accounter is a working prototype. The validation works but I don't know yet if the salience ratings prioritize the right signals, if the auto-tags help or clutter, or whether going from 355 signals to a story is still too manual.
I'm looking for 3-5 pilot collaborators who:
What Accounter produces
What I'm not promising:
If you're wrestling with hundreds of documents, let's talk.
A permanent, structured record of everything worth attention in your documents. Each signal pairs an editable label with a verbatim evidence excerpt and its precise source location. These are real signals from my HS2 pilot run — the same data explored in the case studies.
I'm looking for 3-5 pilot collaborators. Tell me about your document corpus and what you're trying to understand.