What
The per-arm telemetry line reports an objective that omits every continuous column's contribution.
OutputProtocol.onImprovement (CliMode.kt:480) prints %%%klause-arm: label=<arm> objective=<n> time=<ms>, fed from Solvable.objectiveValue — the Long channel, which carries only the Boolean and integer terms. On khb05250 an arm reports objective=27500000 against a run whose reported objective is 2.3e8.
#1739 fixed the same omission on the o line by carrying continuousObjectiveValue beside the exact Long, and deliberately left this one: widening it changes a contract the bench parses as an integer.
Why
Arm attribution drives the credit campaigns (#414, #554). On a model whose cost sits mostly on continuous columns the numbers are not the objective, and the shortfall is not a constant offset — it varies with each incumbent's continuous part — so it can reorder arms rather than just shifting them.
The Long channel is exact and must stay so for the wide-coefficient formats (PB and WCNF objectives can outrun 2⁵³), so this wants the same two-channel treatment #1739 used, plus a decision about what the bench's parser should accept.
What
The per-arm telemetry line reports an objective that omits every continuous column's contribution.
OutputProtocol.onImprovement(CliMode.kt:480) prints%%%klause-arm: label=<arm> objective=<n> time=<ms>, fed fromSolvable.objectiveValue— theLongchannel, which carries only the Boolean and integer terms. Onkhb05250an arm reportsobjective=27500000against a run whose reported objective is 2.3e8.#1739 fixed the same omission on the
oline by carryingcontinuousObjectiveValuebeside the exactLong, and deliberately left this one: widening it changes a contract the bench parses as an integer.Why
Arm attribution drives the credit campaigns (#414, #554). On a model whose cost sits mostly on continuous columns the numbers are not the objective, and the shortfall is not a constant offset — it varies with each incumbent's continuous part — so it can reorder arms rather than just shifting them.
The
Longchannel is exact and must stay so for the wide-coefficient formats (PB and WCNF objectives can outrun 2⁵³), so this wants the same two-channel treatment #1739 used, plus a decision about what the bench's parser should accept.