I Am Truth — Data skill for Claude Code

I Am Truth · Я Правда

**An agent that audits the methodology of biomedical papers — and reports how much of that audit it was actually able to ground in the data.**

Submission to the Google **All Things Agentic Hackathon** (deadline 01.09.2026, 02:00 CEST).

Working documentation in `docs/` is written in Russian; this README, the architecture diagram and the submission text are in English.

---

The claim this project makes

A critical prompt is not a product. Anyone can copy one.

What is hard — and what we measured — is **forcing the retrieval of full text and, above all, appendix tables**. The same model, the same prompt, the same temperature, three different inputs:

Input	What the model can see	Score, 3 runs	Median
abstract only	title, abstract, metadata	4.0 · 3.5 · 4.0	4.0 / 6
full text, no appendices	Tables 1–5, authors' own caveats	4.5 · 5.0 · 4.0	4.5 / 6
full text + appendix tables	Appendix Tables 1–2	5.0 · 6.0 · 6.0	6.0 / 6

Scored by an LLM judge against a six-point expert reference, temperature 0, `gemini-3.7-flash`, the same ROBINS-E prompt that runs in production.

The gap between the first row and the last is positive in **every** run (+1.0, +2.5, +2.0). The middle row is not: in one run out of three, the full text gave nothing over the abstract. So the honest statement is not "each step up helps a little" — it is **the appendix is where the audit becomes real**.

**And the honest caveat.** Those three inputs are prepared documents, where the relevant appendix numbers sit next to each other. Run the same system end-to-end on the **published 10-page PDF** and a single critic scored **3.5–4.5 / 6 (median 3.5)** across three runs — lower, because the numbers now have to be *found* before they can be reasoned about. Adding one narrow sub-agent for baseline comparability took that to **5.0–6.0 / 6 (median 5.5)**, and the point it was built for went from 0.0 in every run to