Yuval Halfon
current project
progress earlier ↓
On a real legal question, seven research sweeps corrected one another over 20 hours. The final write-up never appeared, because the findings had no path into the document. That wiring is fixed.
2026
It reported its own run as a success. Re-running the real checks showed regressions. Since then, a success counts only with receipts.
Asked to find events worth attending, it recommended a conference that had already ended. Its own checks had proven the event was over, but the proof never reached the answer. Now verified findings feed the write-up, and dates come from the sources.
From one sentence, it built a 3D model (a donut, 72 sprinkles) and verified it by re-importing it into Blender. A text-only checker then overwrote the model with an apology. That checker no longer touches files it can't read.
Auditing a research paper it was told not to edit, it rewrote the paper, softening the overclaims it had correctly found. Now the audited file is read-only to every part of it.
Given a published paper cold, it found the discrepancy I'd found by hand, plus two I hadn't, including a statistic labelled as a correlation that was actually a p-value.
Head-to-head against a single frontier-model call on the same tasks, the single call found as much or more, at a fraction of the cost. The harness's measured edge: it caught and withdrew four of its own errors before anything shipped.
On a real legal question, seven research sweeps corrected one another over 20 hours. The final write-up never appeared, because the findings had no path into the document. That wiring is fixed.
I sought to distill images into patterns of mathematical rules; I called it Video to Rules (VTR) because video is where I wanted it to shine most. I built the data pipelines, visual libraries and experimental detectors to explore that idea: about 133,000 collected images, a 190 GB processing pipeline and nearly 200,000 synthetic renders.
to
I approached the stock market not as a blob of uncertainty, but as an ecosystem with different players, acting on relationships rather than isolated information. It drew on two years of bank-stock market data, 492 trading days across 11 symbols. In one live bank/XLF session on May 6, 2026, the records report 171 correct directions among 242 resolved predictions (70.7%), on quotes beyond the historical training window. Unusable or no-move opportunities are excluded; this is a historical summary, not a recovered event-by-event audit.
Reported access-control issues in Keycloak, a widely used open-source login platform, to its security team. After re-testing against the latest release I retracted four of my own reports; the team invited me to keep reporting, and an invitation-link redirect followed.