Evidence and the Rubric

"Every reading points at the exact words it stands on. Change the rubric, and history is re-read — never rewritten."

Evidence Is the Anchor

The single most important field in a reading is evidence: the verbatim span, in its original language, that the reading rests on. "Mild" is a derived label; "a little stiff" is the truth it was derived from, and the reading carries the truth right beside the label. This is what keeps the typed layer honest. Any later analysis that cites a symptom must be able to trace back through the reading's evidence to Dad's actual words on an actual day — a claim you cannot trace to a span is a claim the engine is not allowed to make.

Normalization Without Erasure

Severity is normalized onto a small ladder — mild, moderate, severe, unspecified — and trend onto improving, worsening, stable, unspecified. But normalization happens on top of the evidence, never in place of it. The canonical field values are English so the aggregates in Track 6 can group them, while the evidence stays in whatever language Dad wrote. So "꽤 아팠다" becomes severity: moderate with the Korean span preserved. The comparable label and the citable truth coexist; neither is sacrificed to the other.

Evidence is verbatim and language-preserving; the rubric is versioned. The reading records which interpreter and which rubric generation produced it, so a smarter rubric later re-reads the crumbs and stamps a new generation — the old readings stay, the evidence never moves, and a comparison always knows which lens it is looking through.

The Boundary Lives Inside the Rubric

One more thing the rubric carries: the no-diagnosis boundary, stated inside the rubric itself. The instruction that reads a crumb also tells the model, in the same breath, that it is producing observations and never a diagnosis. The safety line from Track 2 is not bolted on after the reading — it is part of the contract the reading is produced under. That is the pattern to copy: when a rule must always hold, put it where the work happens, not in a checker that runs afterward and hopes to catch a miss.

Code

A rubric bump re-reads; evidence never moves·python

# Reading under rubric v2:
{
  "types": ["symptom"], "site": "knee", "severity": "mild",
  "evidence": "\uc880 \uc2dc\uadfc\ud558\ub2e4",   # verbatim, original language
  "rubric_version": 2
}

# Rubric v3 reads the same wording as moderate.
# Forge RE-READS the crumb and stamps a new generation:
{
  "types": ["symptom"], "site": "knee", "severity": "moderate",
  "evidence": "\uc880 \uc2dc\uadfc\ud558\ub2e4",   # unchanged - same truth
  "rubric_version": 3
}
# The v2 reading still exists. The words never changed.
# A comparison across time knows which rubric each label used.

Exercise

A reading has evidence 'a little stiff' and severity mild under rubric v2. Rubric v3 starts reading 'a little stiff' as a slightly higher severity. Name exactly which fields of the reading change, which stay identical, and what happens to the v2 reading.

Hint

Severity and rubric_version change; the evidence span, the crumb, and the crumb's ULID stay identical; the v2 reading is not deleted — it remains in the record so comparisons know which generation each label came from.

Evidence and the Rubric

Evidence Is the Anchor

Normalization Without Erasure

The Boundary Lives Inside the Rubric

Code

External links

Exercise

Progress

Comments 0