now

current focus

Reel 1 is six days out and the experimental half is done. A 0.5B model drafts, a 3B model optionally critiques and revises, and the whole pipeline was benchmarked on 340 items. Three clean findings: cross-model critique of a broken draft rarely recovers it, self-certainty from the small model beats the big model's opinion as a gating signal, and naive LoRA distillation on 205 gated revisions catastrophically forgets.

The full writeup is an 8-page paper (paper/main.pdf in the experiment repo) plus a README that walks through all four phases. Repo stays private until the reel ships on Tuesday; the teaser post goes out Friday morning with one number from the experiment.

shipping next

• Fri Apr 24 — tease post on X + IG story. One number, no spoilers. Locked.
• Sun Apr 26 — script + storyboard the reel. Test hook with two or three people.
• Mon Apr 27 — record primary footage.
• Tue Apr 29 — Reel 1 ships on IG (11am ET) and X (9:30am ET). Experiment repo flips public. /experiments/001 gets the full writeup.
• Tue May 6 — Reel 2, gridworld agent with noisy intent conditioning. Environment work starts Wed Apr 30.
• Tue May 13 — Reel 3 + first long-form post. Four gate mechanisms head-to-head.

recently finished

• 2026-04-23 · Experiment 001 (sys-1/sys-2 LLM cascade) completed end to end. Four phases (uniform critique gate, task-aware gate, self-gate, LoRA distillation). Paper drafted, compiled, and proofed.
• 2026-04-23 · relectrik.com live on Vercel with custom domain, TLS, OG preview, hi@ forwarding preserved.
• 2026-04-23 · Site v1 built end to end: hero, /experiments, /writing, /now, /about, RSS, sitemap, dynamic OG image. Night-owl syntax theme.
• Brand system v1 locked. Electric Lab aesthetic, Nathan Jee neuron mark, red scalpel accent.
• Handle claimed: @relectrik_lab on IG and X.
• Content plan drafted through Jun 10 (Reels 1–7).