Retry the unresolved residual with fuzzy candidates

Generates fuzzy pairs for every identity the cascade did not clear, scores them through npmatch::np_compare() and npmatch::np_score(), unions them with the pairs the cascade already retained, and re-runs selection and tiering over that union.

Usage

pf_fuzzy_recover(
  m,
  ref,
  cutoff = 0.9,
  top_k = 5L,
  block = c("zip5", "city_state"),
  skip_personal = TRUE,
  max_pool = 20000L,
  verbose = TRUE
)

Arguments

m: A pf_match object from pf_match_grants(), with its pairs attribute intact.
ref: The same reference bundle the match was run against.
cutoff, top_k, block, skip_personal, max_pool: Passed to pf_fuzzy_candidates().
verbose: Report the residual size, pairs added and tier movement.

Value

A list with result (the re-tiered frame), before and after tier tables, moved (the identities whose tier changed, with their fuzzy similarity), and n (counts).

Why the union, and not the fuzzy pairs alone

Selecting over the fuzzy pairs by themselves would let a fuzzy row win an identity that an exact pass had already beaten on the same identity – 26 of the 33 recoverable NO cases did have candidates, scored and rejected, so this is the common case rather than an edge one. npmatch::np_cascade() makes its own final decision over the deduped union of every pass, and this reproduces that: the fuzzy pairs are one more pass's worth of rows, competing on the same footing.

A consequence worth stating, because it is the regression question: an identity that already cleared YES is not in the residual, no pair is added to it, and its selection is therefore bit-identical. The function asserts that rather than trusting it.

What it is NOT limited to

The residual is every identity that has not cleared YES, which on the production run is 845 NO and 1,044 MAYBE. "NO" is a tier assigned after scoring, so it cannot be blocked on; and npmatch::np_cascade() drops an identity from pending only when a pass clears it to YES, which is the same rule. So this changes the MAYBE tier too – it adds rows to the adjudication queue and can change which row a MAYBE identity is holding. That is where a regression has to look; the YES set is protected by construction.