Dyad collapse settings

Usage

pf_dyad_config(
  tie_band = 0.05,
  max_band = 25L,
  use = "substitutable",
  policy = c("query", "operating", "arm", "top"),
  keep_soft_veto = TRUE
)

Arguments

tie_band: Candidates within this much of the top score are in contention. Matches npmatch's np_config()$tie_band.
max_band: Cap on how many candidates per grant enter the collapse, by score. This is a hard requirement, not a tuning knob: pairing is quadratic, and a token-blocked query can put thousands of equal-scoring candidates inside the band — one grant with 2,000 of them is two million comparisons, which is what hung the first run of dev/test-dyad-cascade.R. Nothing below the top handful can affect the selection anyway. Counts of what the cap drops are reported by pf_dyad_collapse().
use: Which dyad classes may be collapsed. The default is "substitutable" only; c("substitutable", "related") is available for measurement but see pf_dyads for why it should not be shipped.
keep_soft_veto: Carry a soft veto from a merged-away member onto the surviving representative, so the grant stays in review. A soft veto is npmatch's "do not auto-accept this" signal, and affiliate_suffix is the operating-arm problem in veto form — it fires on exactly the X FOUNDATION candidate when the grant says X. Without this, dropping the arm drops the veto with it. Measured on the 50k run: 71 of 700 promotions came from that path, every one of them a case where the cascade ranked the arm first and the policy overrode both the score and the veto. FALSE restores the older behaviour and recovers those 71 auto-accepts, at the cost of resting them on a rule validated at 60% on 15 cases.

Value

A list.

Why the policies differ

On the 50k run, operating promotes 629 MAYBE grants to YES, top 388 and arm 253 — and top cannot lose on score, since it keeps the highest-scoring member by construction. The gap is the soft veto. Of the 241 grants where operating and top land on different tiers, top keeps an arm in 240, operating in none, and the query carries a support marker in none. That is exactly the affiliate_suffix trigger — candidate ends in a support token, query does not — so top inherits a veto that operating escapes. The policies are not trading accuracy for score; one of them is standing on a veto.