Every Claim Cites

"If a sentence in the analysis cannot point at the crumbs it stands on, the engine is not allowed to say it."

Rule One: No Claim Without Evidence

The first of the three analysis rules is the strictest: every claim carries its crumb IDs, and the interpreter generation it read. An analysis statement that cannot be traced to specific dated crumbs is an unfalsifiable summary, and an unfalsifiable summary is a bug — not a stylistic lapse. The reader must always be able to ask "how do you know that?" and get an answer that resolves to Dad's own words on specific days. Analysis that cannot answer that question is exactly the guessing-machine failure the whole surface is designed to avoid.

Enforced at the Storage Boundary

This rule is not a request to the model; it is a wall in the code. When narration is produced, the storage boundary checks it: a narration with no crumb citations, or one that cites an ID outside its own window, is refused — nothing lands. The artifact is only stored if every claim it makes is anchored to crumbs inside the window it analyzed. Putting the check at the storage layer, not in a hopeful prompt, is the same lesson as Track 4's rubric boundary: when a rule must always hold, enforce it where the data becomes durable.

A claim without evidence is a bug, refused at the storage boundary. Narration cites inline crumb IDs; the boundary rejects any narration missing them or citing outside the window. The engine cannot store a sentence it cannot ground. Provenance is not decoration here — it is the admission ticket.

The Reader Sees the Receipts

For Dad, the inline [crumb:ID] markers render as small [n] references keyed to a Sources legend, where each entry shows the cited crumb's date and words. The stored artifact keeps the real IDs; the UI shows the human-readable receipt. So an analysis is not a wall of assertions to take on faith — it is a set of observations, each with a visible "here is the note this came from" attached. That is what turns a summary into something a doctor, or Dad, can actually check.

Code

Cited narration lands; uncited narration is refused·text

# ACCEPTED - every claim anchored to crumbs in the window:
"Two flares this window, each within a week of a heavy leg
 session [crumb:01J8A] [crumb:01J8C]. Three lighter weeks had
 no flare [crumb:01J8F]. Worth asking a doctor about load."
   -> stored. renders [1][2][3] with a dated Sources legend.

# REFUSED at the storage boundary:
"Your inflammation is clearly trending worse lately."
   -> no crumb IDs. nothing to trace. REJECTED, nothing stored.

"Back in March you had the same issue."   # window is July
   -> cites outside the analyzed window. REJECTED.

Exercise

Write one health claim about yourself in a single confident sentence. Now rewrite it so every part of it points at a specific dated note you could actually produce. Which version would survive someone asking 'how do you know?' — and what did the first version quietly assume?

Hint

The confident version usually smuggles in a trend or a cause you never actually recorded. The cited version can only say what the notes support — which is exactly the constraint that makes it trustworthy.

Rule One: No Claim Without Evidence

Enforced at the Storage Boundary

The Reader Sees the Receipts

Code

External links

Exercise

Progress

Comments 0