Reconciliation

After this lesson, the reader should understand:

• 01Compare a documented rule and an observed rule without relying on text similarity.
• 02Derive the direction of a divergence from the geometry of the boundary.
• 03Classify a reconciliation outcome as faithful, divergent, or undocumented.

Comparing the wrong things

A policy rule reads "approve where the applicant meets the standard exposure test". An observed rule reads "approve where exposure ≤ 0.3". A text comparison scores them as closely related. Is the institution compliant?

The comparison cannot say, because it compared descriptions rather than boundaries. Two rules can share almost every significant word and impose entirely different thresholds; two rules can share no vocabulary at all and impose the same one. A reconciliation that means anything extracts a boundary from each side^[1] — the field, the threshold, and which side is permissive — and compares those. Normalise both to the same orientation, and the direction of any divergence then follows from arithmetic rather than interpretation: a shift that enlarges the permissive region is movement toward permissiveness, whatever the surrounding prose appears to say. The NOMOS reference implementation works boundary to boundary in exactly this way: it parses a numeric boundary per field from each side, normalises orientation, and returns one of three results — faithful (within tolerance), a threshold divergence carrying a deterministic direction and magnitude, or an undocumented boundary where practice holds a line the document never states. The fourth outcome in the framework — a documented boundary practice never exercises — is not one a behavioural reconciler can assert from behaviour alone; it shows up instead as a policy rule the scan never has cause to touch. The strongest objection targets the word 'faithful'. Agreement 'within tolerance' depends entirely on the tolerance, and a tolerance is a number someone chose. Widen it and every divergence becomes fidelity; narrow it and every rounding difference becomes drift. The objection is correct that the tolerance carries the verdict, and the answer is to treat it accordingly: as a declared parameter, set by someone with authority to set it, published alongside every result it produced^[2]. There is also a statistical discipline worth borrowing. Showing that two quantities are the same is not the same as failing to show they differ — a small sample fails to show almost anything. Equivalence testing, developed for drug trials, inverts the usual question^[3]: it asks the data to demonstrate that the difference lies inside a stated margin. A 'faithful' finding should meet that standard, not merely survive the absence of a contrary one. The method's own reach should be stated plainly. Boundary-to-boundary comparison works where both sides have a numeric threshold to extract. Categorical boundaries — 'approve for sectors A and B', 'escalate for new suppliers' — have no orientation to normalise, and comparing them still leans on matching names. The reference implementation reports those with an unknown direction rather than inventing one. And a faithful threshold can still hide disagreement in dispersion: two boundaries at the same value, one sharp and one smeared across a wide band, describe different institutions.

• What tolerance should count as agreement, and who is entitled to set it?
• Is an undocumented boundary a gap in the document, or in the practice?
• How should a reconciliation report a rule that is documented but never used?
• Why is "faithful" a result worth publishing?
• What evidence should a "faithful" finding require — the absence of a detected difference, or a demonstrated equivalence?

No responses yet. Be the first to answer one of the questions above.

Reconcile one field end to end.

1. 01Extract the documented boundary for one field.
2. 02Extract the observed boundary for the same field.
3. 03Normalise both to the same permissive orientation.
4. 04Classify the outcome: faithful, divergent, undocumented, or dormant.
5. 05If divergent, state direction and magnitude — without yet calling it good or bad.
6. 06State the tolerance you used, who would have to approve it, and whether the result changes if it halves.