index.html is now the leaderboard (overview, per-model pages via ?model=, per-benchmark modal via ?bench=), news.html keeps the reproductions tracker, methodology.html documents the suite and the index. Data comes from data/index.json and data/methodology.json, generated by the lab's build scripts.
Refresh all metrics, add reflex's launch post, raise the recency credit
Metrics: re-read every item against X, GitHub (via the authenticated gh CLI,
5,000/hr instead of 60) and the Hub, 43 of 46 entries updated. The snapshot
moves to 2026-09-19 in the footer, README and BASE_DATE.
The refresh reshaped the top of the list: Laya went 163 -> 1,410 GitHub stars
and 108 -> 577 Hub likes (raw 1,679 -> 11,666), which lifted it above LocalJev
and dropped LocalJev from #3 to #5 at the old credit of 5000. The credit is
raised to 10000, which puts LocalJev back at #2 with a 926-point cushion over
#4 rather than the 51-point one a credit of 9000 would give -- a margin thin
enough that the next refresh would flip it back. Top 5 is now Laya, LocalJev,
45-second TL;DR, Qwen-2.5-1B-RLCD, openjev.com.
Note the engagement gate now saturates for 20 of 46 items, so for anything
with real traction the credit acts as a flat recency tier and only age
separates items. That is the intended behaviour here -- it is what keeps a
0-engagement entry off the list -- but it does mean t…
Make the recency credit decay daily (5000, half-life 1 day)
The credit was 900 over a 2-day half-life, which was not enough to lift
the GitHub Next release (LocalJev) into the top 3. Per the instruction
that "the github next thing is on top 3, and the decay should be daily",
the credit is now 5000 with a 1-day half-life, so a three-day-old item
keeps only an eighth of it.
The engagement gate is kept: the credit is scaled by the item's own
weighted engagement (saturating at ref 400), so an entry nobody has
reacted to cannot climb on its date alone. This is what keeps snapjudge
(0 stars, same 2026-09-18 date as LocalJev) from riding the same date
credit into the top half; it stays at #44 with a zeroed credit, while
LocalJev reaches #3 at 8008 and the top 2 are untouched.
The "new" badge threshold tightens from 1.5 to 1 day to match. Same four
items carry it: LocalJev, Laya, Bespoke Nimble, snapjudge.
Co-Authored-By: Claude Code <noreply@anthropic.com>
