Method
Fluency is not structure.
A plan can be well-written and still rest on one untested stone. Models are unusually good at producing that kind of plan. So are people under time pressure. Fathom is a way of looking, not a way of generating more prose.
01
Claims, not paragraphs
Any design, incident writeup, or model answer is a set of claims. Some are goals. Some are constraints. Most are assumptions wearing the grammar of fact. Fathom forces the brief into a directed graph: this node must hold for that node to mean anything.
02
Blast radius is not importance
Importance is a feeling. Blast radius is a count: if this node is false, what fraction of the graph is unsupported. A colorful risk in the appendix can have blast near zero. A boring assumption about idempotency can take the product with it.
03
Confidence is epistemic
Confidence is not how loudly the brief states the claim. It is how well the brief actually supports it. An idea that “everyone does” and that has never been measured is low confidence, even if it is fashionable. Fathom will mark it that way on purpose.
04
The keystone
In an arch, the keystone is the wedge that converts a pile of stones into a structure. In a plan, it is the assumption with high blast and low support, on the path to the goal. If it yields, the rest of the document is decoration. Naming it is the whole product.
05
Load, then cascade
Stress-testing is an adversarial pass on a single node: would a competent skeptic ship on this? What-if is mechanical — flip the node to false and recompute what still stands. Neither is brainstorming. Both are closer to a proof load on a beam.
06
What this is not
Not a chatbot. Not a writing assistant. Not an eval harness, though it will make your evals less theatrical. It will not tell you what to build. It will tell you which sentence, if wrong, means you did not build it.
07
Why not just ask a chatbot
You can. You will get a fluent critique, a different one tomorrow, and no way to put a load on a single sentence. Fathom is not smarter prose. It is a graph, two numbers that people mix up (blast vs confidence), and a named keystone you can flip. If that is not worth more than a chat window, use the chat window.
08
The cases are not demos. They are tests.
Each prepared case was written against a failure a competent reviewer already knows: retrieval in RAG, idempotency in a booking agent, webhook identity in a double-charge, fail-open in a limiter, labels in a “we’ll fine-tune later” deck. If Fathom had named a veneer, we would have thrown the case out. That is the check. Live custom briefs are a model judgment — treat them as a first pass, then stress the stone yourself.