The citation verifier
Checked 33 references. Two were fabricated: a real author spliced onto an invented journal, title and statistics. Both replaced with the real papers.
Asked to finish the paper, it refused to invent Results.
CSUN
Things built here with AI. Most have no settled purpose yet. If one is useful to you, take it.
Checked 33 references. Two were fabricated: a real author spliced onto an invented journal, title and statistics. Both replaced with the real papers.
Asked to finish the paper, it refused to invent Results.
A paper arguing AI needs verification went through three AI reviews. They invented four citations.
None reached the manuscript. Every one was checked first.
A course reading that reads itself aloud in the instructor's cloned voice. Twenty-one tracks.
Students get a lecture without anyone recording one.
A pipeline drafted a paper from real data. A second reader, given only the manuscript and the data, rejected it the same session.
An AI system that rejects its own work instead of defending it.
The pipeline behind the three above. It can say no.
Dead ends count as results. Whoever wrote something cannot certify it.
Two models debated the future of business schools. One later noticed it had also been the referee, and found it had overstated in its own favour.
An Oregon-Trail-style game that teaches the research pipeline. The AI played it through the browser to test it.
Management cases as detective work. You get the scene and work out which failure caused it.
One instruction about the socio-technical systems literature, taken to a finished conceptual paper in a single session.
Four models, five questions, six schools: 600 responses. Human coders worked blind. Asked whether an AI could code instead, it said no.
Forty-five management theories as tools students judge for era and fit. The case generator was never wired to a key.
A narrated deck with a drawing on nearly every slide and no image files. All of it drawn in code.
One paper halved repeatedly down to a single word, with what breaks at each step.
Every journal on the impact list: voice, limits, and the unwritten rule that kills submissions.
Eighty-five faculty as a co-authorship network. Surname matching proposed 77 pairs; 30 survived scrutiny.
Six sliders for voice, with a sentence that rewrites itself as you drag. Two of the dials overlap.
Coded award criteria checked against the signed gift agreements, and whether processing order changed who got money.
What the college thinks about AI. Anonymous, refreshed daily, published unedited.
Eighty-five faculty researched in parallel. Verification caught fabricated claims before publish.
Role-specific AI briefings, written for what each role has to decide.
Thirty-four figures of speech in three classical families, each marked for whether it appears in one essay and whether it is a measured habit of the writer across a 25-year corpus.
Built to answer whether a model can be told to sound like someone.
OpenAI, Anthropic and Google scored on six dimensions. The page stores its own history and computes the deltas.
Want one? Several outlived what they were built for. Say so and it is yours.
Missing links. Most have no live URL: the hosting was retired. Local copies survive. Screenshots to come.