Launched this week
BayesCore
Bayesian evidence scoring for any structured claim
3 followers
Bayesian evidence scoring for any structured claim
3 followers
Drop any document with a claim. Bayescore extracts its IS(subject, criterion) hypothesis, derives falsifiable predicates, and scores each using a two-pass adversarial LLM evaluation. Absence of evidence reduces confidence toward zero. Formula locked: score = Σ(weight × confidence) × 100. We ran it on ourselves. 46/100, Grade D. Five failing predicates. Published because it's right, not flattering. bayescore.com/self-eval
Hey Product Hunt! I'm Bugra, the solo founder of Bayescore. Here's why I built it.
I was reviewing a startup pitch — a founder who'd clearly worked hard — and I had no principled way to give feedback. Everything I said could be dismissed as opinion. There was no rubric. And the rubrics I found online were someone else's checklist, not derived from what the document was actually trying to prove.
So I built a tool that extracts the rubric from the document itself.
How it works:
1. Drop any document with a claim in it
2. Bayescore extracts IS(subject, criterion) — the evaluation hypothesis implied by the document
3. You get a score (0–100), a grade (A–F), per-predicate rationales, and the highest-leverage gaps to fix
Who's been using it:
- Founders reviewing their own decks:** "Tell me exactly what's missing, not what sounds good"
- Grant writers:** Score the proposal against IS(proposal, fundable) before submission
- Accelerators and investors:** Drop 10 applications, get a consistent predicate breakdown across all of them
- Marketers:** Score a brief against IS(campaign, worth_running) — does the evidence actually support the spend?
- Builders evaluating ideas:** IS(product, worth_building) — before you write a line of code
The thing I built that surprised me most:**
The extraction step. Getting an LLM to identify an evaluation domain from an arbitrary document — and express it as an *external evaluator's* falsifiable hypothesis rather than an internal description of what the document claims — turned out to be the hard problem. The prompt has to distinguish between "IS(startup, launch_ready)" and "IS(the founder's claims, true)." One is an evaluation. The other is a tautology.
I ran it on Bayescore itself.**
Score: 46/100, Grade D.
Three predicates pass: value proposition, problem worth solving, risk identified. Five fail: customer validation (0), demand signal (0), acquisition channel (partial), domain expertise (partial), go-to-market (partial). The full breakdown is at bayescore.com/self-eval.
I published it because if Bayescore hid its own score, you couldn't trust any score it gave.
What I'd love feedback on:
1. Does the IS(subject, criterion) framing make sense when you try it on your own documents?
2. Custom domains — you can drop any document, extract its IS hypothesis, and share the evaluation link. Is that useful for you?
3. What document type would you throw at this that I haven't thought of?
Try it at bayescore.com — paste anything with a claim, takes ~30 seconds.
Thanks for hunting us today. I'll be here all day answering everything, especially the hard questions.
— Bugra