Technology & education · worldwide

Examined the way
research is examined.

Most writing about technology and education is opinion wearing the clothes of evidence. Here, claims are sourced, statistics carry the institution and the year that produced them, and the reference list is part of the article — then the whole thing opens up to a community that spans every country.

Sections

A capsule bar shows a model's advertised context window with a dashed marker where reliable accuracy actually ends, alongside stats on Llama 4 Scout's 10M-token claim, its 22% NoLiMa accuracy at 32K, and Mistral 7B's real 16K limit.
Artificial Intelligence 2h ago

Long-Context Windows: What Independent Benchmarks Show Versus What Vendors Claim

Context windows now advertise millions of tokens, but independent benchmarks — RULER, NoLiMa, Lost in the Middle and a newer prep…

A DNA double helix with a highlighted variant callout illustrates DeepMind's AlphaGenome Atlas, which precomputes effects for all 9 billion possible human DNA changes into a searchable, ~1-petabyte database.
Artificial Intelligence 3h ago

DeepMind Publishes Precomputed Predictions for All 9 Billion Possible Human DNA Changes

Google DeepMind has published precomputed predictions for all 9 billion possible single-letter human DNA changes in a new, search…

Three vertical bars compare SCORE's headline pass rates — 73.5% reproducible, 57% robust, 55.1% replicated — under the line "Reproducible. Robust. Replicable. Rarely all three."
Research Practice 14h ago

Reproducible, Robust, Replicable: What a 3,900-Claim Audit of Social Science Found

A seven-year DARPA-funded project checked 3,900 social-science claims three ways — reproducing, stress-testing, and replicating t…

A "words vs. action" gauge shows 68% of CoARA action plans use "open science" language, but most specific practices are documented in under 20% of the 248 plans.
Education 15h ago

A Count of 248 Research-Assessment Reform Plans Finds Values Stated, Operational Change Lagging

A content analysis of 248 action plans filed with Europe's Coalition for Advancing Research Assessment finds open-science languag…

A bispecific antibody schematic bridges a T-cell and a myeloma cell, alongside headline stats showing 77.8% complete response, 92% two-year progression-free survival, and a hazard ratio of 0.15 for teclistamab.
Health 21h ago

Randomized Trial Finds Teclistamab Outperforms Standard Early Treatment in High-Risk Smoldering Myeloma

A phase 2 trial in Nature Medicine found that early teclistamab treatment produced deep, durable remissions in high-risk smolderi…

A shield-checked chat bubble beside SB 1119's headline stats: a $15,000 max penalty per child, a one-hour session cap for minors, and the July 2027 compliance deadline.
Artificial Intelligence 1d ago

California Signs Adam's Law, the First Statute to Require Safety Audits of AI Companion Chatbots

California's SB 1119, signed September 10, 2026, is the first law to mandate independent child-safety audits of AI companion chat…

Why this exists

If something cannot be sourced, it does not get published.

That single rule shapes everything else. An article here is written to be checked: follow a claim to its origin, see which institution produced a number and in which year, and take the reference list with you.

The same standard applies to the community. Polls are voted on by verified members so one person means one vote, and the results are published openly — raw counts, not a headline summarising them.

Nothing is quietly deleted or rewritten. Corrections are marked on the article itself, and published work keeps its address in the archive.