hindsight-skills

Memory teaches AI your facts. Hindsight teaches it your patterns of being wrong, the calibration layer every AI advisor will need.

Comment

4th at YC Hackathon and GBrain opensource

Hindsight

Hindsight is the calibration layer every AI advisor is going to need.

Memory teaches AI your facts. Hindsight teaches it your patterns of being wrong.

Five new GBrain skills + a GStack Browser extension. Submitted to the YC GStack × GBrain hackathon, May 16 2026 by team Hindsight.

What's in this org

Built by

  • Rayan Arya
  • Rushil Jaiswal
  • Keshav Kotamraju

Demo

https://drive.google.com/drive/folders/19mLqKOWKwfYlEnFAGvXsOz7omVFLwEbZ

The skills

  • extract-takes — finds gradeable claims in brain pages
  • resolve-outcomes — grades takes against public evidence (curated outcomes + Claude fallback)
  • hindsight-profile — aggregates resolved takes into a calibration profile (quantitative stats + qualitative patterns)
  • find-contradictions — surfaces pairs where the brain owner contradicted themselves across time
  • calibrated-advise — gives advice that applies the calibration profile to a current question, using ZeroEntropy retrieval + The Hog fresh-signal

Demo data

  • 228 Paul Graham essays scraped (2001–2026) — data/corpus/
  • 30 hand-curated outcomes; 15 attached to the live extracted-take corpus for the calibration sample — data/outcomes.json
  • 67% overall hit rate, 4 sharp pattern statements including the meta-pattern that high conviction backfires on macro/structural calls
  • 1 hand-verified self-contradiction PG flagged himself — data/contradiction-fallback.json

Install

git clone https://github.com/rayan-arya/hindsight-skills
cd hindsight-skills
npm install
ANTHROPIC_API_KEY=... npm run dev

Server runs on port 3001. POST to /skills/[name] to invoke each skill.

Track C scripts (corpus scrape + The Hog fresh-signal service) live in data/scripts/ and have their own requirements.txt.

The first commercial wedge

Partner memos at venture firms — concentrated, dated, falsifiable claims with public outcome data, and partners pay for anything that makes them less wrong. We think the calibration layer is what every AI memory product is going to need within 18 months.

Built With

Share this project:

Updates