feat: add competitor comparison to README to answer 'why here?' objection

Add direct comparisons to the five alternatives a top mind would consider:
academic paper (takes 1-3 years, never reaches field), EA Forum posts (no
canonical output), Kaggle competitions (optimizes benchmark not field utility),
AI for Good initiatives (proprietary black box), running AI yourself (no review
mechanism for false confidence). Each comparison names the specific failure mode
of the alternative and why the Open Problem Lab protocol addresses it.

This is the conversion question every serious contributor asks before investing
time. Answering it directly in the README removes the objection before it forms.

Author: cschanhniem

README.md

@@ -58,6 +58,20 @@ Browse all packs in [problem-packs/](problem-packs/). Each pack contains a probl
 
 ## What Makes This Different
 
+**Compared to writing a paper:** A paper takes 1–3 years, typically never reaches field practitioners, and produces a PDF that AI systems cannot act on. Here, a verified evidence record or validated signal analysis can be in the hands of a field reviewer within days. Contribution scope is narrower — one claim, one method, one reviewable artifact — which makes it composable and incrementally improvable.
+
+**Compared to EA Forum posts or Slack discussions:** Unstructured discussion is useful for framing but produces no canonical, replicable output. A discussion here must become an Issue with a done condition, an Issue must become a Pull Request, and a Pull Request must survive domain review before anything is accepted. The output is auditable and citable.
+
+**Compared to Kaggle competitions:** Kaggle optimizes for benchmark scores, not field utility. A malaria model that achieves high AUC on a held-out dataset but has a 40% false-positive rate at sub-national scale during low-transmission seasons is not a field tool — it is a distraction. Open Problem Lab's field-reality and red-team review gates catch this before anything is published.
+
+**Compared to AI for Good initiatives:** Most "AI for Good" work is proprietary, black-box, and inaccessible to the field practitioners who would use it. Open Problem Lab is forkable, reproducible, and designed so that a district health officer can trace any accepted claim back to its source, method, and limitations.
+
+**Compared to just running AI on the problem yourself:** Useful as a first draft. Useless as a result. An unreviewed AI output has no mechanism for catching false confidence, operational overreach, or dangerous misuse. The value of this protocol is exactly what is removed when you skip it.
+
+---
+
+**The four properties that matter:**
+
 **Verification is the product.** An AI agent can produce a literature summary in seconds. What takes time — and what is actually valuable — is a submission that cites dated sources, documents its method, states its assumptions, enumerates its failure modes, and can be independently reproduced.
 
 **Negative results are first-class.** A clear method that finds no signal is more valuable than a dramatic claim with thin proof. We build for that standard.