A case, already run · no key
This plan fails if retrieval is wrong.
That is the load-bearing sentence in a real “we’ll just RAG it” design. Fathom named it. A chatbot would have given you twelve risks and no structure.
Prepared case
A real RAG architecture, already decomposed.
Why not paste this into ChatGPT
Chat lists what could go wrong. Fathom names what the plan is standing on.
A chatbot
A new essay every time
- Eight bullet risks, all plausible, none ranked.
- No map of which claim the others require.
- You cannot flip one sentence and watch the rest collapse.
- Tomorrow’s answer will not be this answer.
Fathom
A structure you can load
- One named load-bearing claim — the keystone.
- A graph: this must hold for that to mean anything.
- Stress it. Flip it false. The cascade is the product.
- Prepared cases are authored against a failure we already knew.
01
Map the claims
A brief becomes a directed graph. Goals, assumptions, constraints. Not a summary. Who is standing on whom.
02
Find the stone
One claim is carrying the rest. If it is wrong, you did not build the thing you think you built.
03
Put a load on it
Attack that claim. Then flip it false. Watch which other sentences die. That is the review.
Prepared cases · no key
Start from a failure we already knew.
Each case was written against a load-bearing mistake a competent reviewer would name. If Fathom pointed somewhere else, we would have thrown the case out.
- AI product architecture4 min
We'll just RAG it
Anyone who has shipped retrieval already knows: if kNN misses, citations are theater. Fathom named that stone — not the chat UI.
- Tool-using agent5 min
The agent that books the flight
The booking agent dies if tool calls are not idempotent. Double-booking is not an edge case — it is what the architecture does under retry. Fathom names that beam.
- Production incident3 min
Payments are double-charging
Double charges after a Stripe retry almost always mean the webhook was not idempotent. Fathom names that — not 'the API was slow.'
- Systems design4 min
A million QPS rate limiter
A million-QPS limiter that fails open is a limiter in name only. The keystone is fail-open under Redis loss, not the Lua script.
- AI startup plan4 min
We'll fine-tune later
'We'll fine-tune later' assumes labeled data you do not have. The keystone is the labels, not the GPU.
Chat hides structure. Structure is the job.
Open a prepared case — no signup, no key. Paste your own brief when you want the instrument on a plan you already understand.
Open the RAG case