Insights
AI and evidence
Keeping a research pipeline honest: grounding, parity, absence discipline and the failures they were built after.
The llms.txt question, answered with the evidence
We publish an llms.txt file and we will tell you plainly that the evidence says it does almost nothing today. Here is the data, and here is why we ship it anyway.
ReadWhy we let a human overrule the adversarial checker
We measured what would happen if our skeptical second pass ran automatically. It would have demoted the strongest competitor most and handed our client back first place. A checker tuned to argue, applied mechanically, recreates the bias it was built to remove.
ReadWe audited our own gate and it caught 2 of 7 fabrication classes
Quote verification is necessary and nowhere near sufficient. Here are the five fabrication classes it cannot see, why every one of them passes a byte-for-byte check, and what we built as a result.
ReadNegative controls for AI research
A verification step nobody tests is a verification step you are trusting on faith. Plant a deliberately false claim on every run, and fail loudly if it ever gets accepted.
ReadAn AI analyst that cannot contradict the dashboard
Give a language model its own database access and it will eventually quote a number that disagrees with the chart above it. The fix is architectural: make it read through the same code that renders the tile.
ReadThe grounding gate: how to stop an AI research pipeline inventing quotes
The dangerous failure is not invention from nothing. It is a quote that is ninety percent right and ten percent improved. Byte-for-byte verification against a cached source, with rejection rather than flagging.
Read Find out what your marketing
is actually achieving.
One conversation, no obligation: what you sell, who you lose deals to, and what you cannot currently see. If a build is a fit, it is scoped and quoted from there. There is no signup and no self-serve tier.
No prices on this site. Every engagement is scoped and quoted after the first meeting.