Carry the last score to each missed scheduled visit

1library(yamaa)23adqs <- yamaa_domain("spec.yaml")$output4adqs

1import yamaa23adqs = yamaa.yamaa_domain("spec.yaml").output4adqs

1# Reference solution for the yamaa benchmark adam-adqs-missed-visit-locf (R track).2#3# Written from the benchmark's full prompt and its inputs alone, as an agent4# would write it. Harbor's oracle agent runs it to check that the prompt can5# be solved and that the grader scores a correct answer 1.67library(dplyr, warn.conflicts = FALSE)8library(readr)9library(tidyr)1011adsl <- read_csv(12  "/app/input/adsl.csv",13  col_types = cols(.default = col_character())14)15qs <- read_csv(16  "/app/input/qs.csv",17  col_types = cols(18    AVISITN = col_integer(),19    AVAL = col_double(),20    .default = col_character()21  )22)2324# Collected scores; an empty score is left out, so it counts as missed.25collected <- qs |>26  filter(PARAMCD %in% "ACTOT", !is.na(AVAL)) |>27  left_join(select(adsl, USUBJID, EFFFL), by = "USUBJID") |>28  mutate(DTYPE = NA_character_) |>29  select(STUDYID, USUBJID, PARAMCD, AVISIT, AVISITN, AVAL, DTYPE, EFFFL)3031# One LOCF record for each missed scheduled visit in the efficacy32# population; its value is the closest earlier visit with a score, where33# a collected zero carries forward and no earlier score leaves no value.34scheduled <- tibble(35  AVISIT = c("Week 8", "Week 16", "Week 24"),36  AVISITN = c(8L, 16L, 24L)37)38eff_subjects <- adsl |>39  filter(EFFFL %in% "Y") |>40  select(STUDYID, USUBJID, EFFFL)4142missed_keys <- eff_subjects |>43  crossing(scheduled) |>44  anti_join(45    select(collected, USUBJID, AVISITN),46    by = c("USUBJID", "AVISITN")47  )4849carried <- missed_keys |>50  left_join(51    collected |> select(USUBJID, EARLIER_VISITN = AVISITN, EARLIER_AVAL = AVAL),52    by = "USUBJID",53    relationship = "many-to-many"54  ) |>55  filter(!is.na(EARLIER_AVAL), EARLIER_VISITN < AVISITN) |>56  group_by(STUDYID, USUBJID, AVISIT, AVISITN, EFFFL) |>57  slice_max(EARLIER_VISITN, n = 1, with_ties = FALSE) |>58  ungroup() |>59  transmute(USUBJID, AVISITN, CARRIED = EARLIER_AVAL)6061missed <- missed_keys |>62  left_join(carried, by = c("USUBJID", "AVISITN")) |>63  mutate(PARAMCD = "ACTOT", AVAL = CARRIED, DTYPE = "LOCF") |>64  select(STUDYID, USUBJID, PARAMCD, AVISIT, AVISITN, AVAL, DTYPE, EFFFL)6566adqs <- bind_rows(collected, missed) |>67  arrange(USUBJID, AVISITN)6869dir.create("/app/output", showWarnings = FALSE)70write_csv(adqs, "/app/output/adqs.csv", na = "")

1# Reference solution for the yamaa benchmark adam-adqs-missed-visit-locf (Python track).2#3# Written from the benchmark's full prompt and its inputs alone, as an agent4# would write it. Harbor's oracle agent runs it to check that the prompt can5# be solved and that the grader scores a correct answer 1.67from pathlib import Path89import polars as pl1011adsl = pl.read_csv("/app/input/adsl.csv", infer_schema=False)12qs = pl.read_csv("/app/input/qs.csv", infer_schema=False).with_columns(13    pl.col("AVISITN").cast(pl.Int64),14    pl.col("AVAL").cast(pl.Float64),15)1617# Collected scores; an empty score is left out, so it counts as missed.18collected = (19    qs.filter(pl.col("PARAMCD") == "ACTOT", pl.col("AVAL").is_not_null())20    .join(adsl.select("USUBJID", "EFFFL"), on="USUBJID", how="left")21    .with_columns(DTYPE=pl.lit(None, dtype=pl.String))22    .select(23        "STUDYID", "USUBJID", "PARAMCD", "AVISIT", "AVISITN", "AVAL", "DTYPE", "EFFFL"24    )25)2627# The protocol schedules Weeks 8, 16, and 24 for the efficacy population.28scheduled = pl.DataFrame(29    {"AVISIT": ["Week 8", "Week 16", "Week 24"], "AVISITN": [8, 16, 24]}30)31eff_subjects = adsl.filter(pl.col("EFFFL") == "Y").select(32    "STUDYID", "USUBJID", "EFFFL"33)34planned = eff_subjects.join(scheduled, how="cross")35missed_keys = planned.join(36    collected.select("USUBJID", "AVISITN"), on=["USUBJID", "AVISITN"], how="anti"37)3839# Each missed visit carries the closest earlier visit with a score, where40# a collected zero carries forward and no earlier score leaves no value.41earlier_by_subject: dict[str, list[tuple[int, float]]] = {}42for row in collected.sort(["USUBJID", "AVISITN"]).iter_rows(named=True):43    earlier_by_subject.setdefault(row["USUBJID"], []).append(44        (row["AVISITN"], row["AVAL"])45    )4647missed_rows = []48for row in missed_keys.sort(["USUBJID", "AVISITN"]).iter_rows(named=True):49    carried = None50    for visitn, aval in earlier_by_subject.get(row["USUBJID"], []):51        if visitn < row["AVISITN"]:52            carried = aval53        else:54            break55    missed_rows.append({**row, "PARAMCD": "ACTOT", "AVAL": carried, "DTYPE": "LOCF"})5657missed = pl.DataFrame(58    missed_rows,59    schema={60        "STUDYID": pl.String,61        "USUBJID": pl.String,62        "EFFFL": pl.String,63        "AVISIT": pl.String,64        "AVISITN": pl.Int64,65        "PARAMCD": pl.String,66        "AVAL": pl.Float64,67        "DTYPE": pl.String,68    },69    strict=False,70).select(71    "STUDYID", "USUBJID", "PARAMCD", "AVISIT", "AVISITN", "AVAL", "DTYPE", "EFFFL"72)7374adqs = pl.concat([collected, missed], how="diagonal").sort(["USUBJID", "AVISITN"])7576Path("/app/output").mkdir(exist_ok=True)77adqs.write_csv("/app/output/adqs.csv")

1/* opensas reference solution for the yamaa benchmark adam-adqs-missed-visit-locf.2   Reads the benchmark inputs and derives the requested output.3   Execute with the existing opensas runtime: opensas /app/output/result.sas. */45data adsl;6  length STUDYID USUBJID EFFFL $1024;7  infile "/app/input/adsl.csv" dsd dlm="," firstobs=2 truncover lrecl=32767;8  input STUDYID : $char1024. USUBJID : $char1024. EFFFL : $char1024.;9run;1011data qs;12  length STUDYID USUBJID PARAMCD AVISIT $1024;13  infile "/app/input/qs.csv" dsd dlm="," firstobs=2 truncover lrecl=32767;14  input STUDYID : $char1024. USUBJID : $char1024. QSSEQ PARAMCD : $char1024. AVISIT : $char1024. AVISITN AVAL;15run;1617proc sql;18  create table collected as select a.STUDYID,a.USUBJID,a.PARAMCD,a.AVISIT,a.AVISITN,a.AVAL,'' as DTYPE,b.EFFFL from qs as a left join adsl as b on a.STUDYID=b.STUDYID and a.USUBJID=b.USUBJID where not missing(a.AVAL);19quit;20data schedule; length AVISIT PARAMCD $1024; set adsl; if EFFFL='Y'; PARAMCD='ACTOT'; do AVISITN=8,16,24; AVISIT=catx(' ','Week',strip(put(AVISITN,best32.))); output; end; run;21proc sql;22  create table missed as select a.* from schedule as a left join collected as b on a.STUDYID=b.STUDYID and a.USUBJID=b.USUBJID and a.PARAMCD=b.PARAMCD and a.AVISITN=b.AVISITN where missing(b.AVISITN);23  create table candidates as select a.*,b.AVAL,b.AVISITN as sourcevisit from missed as a left join collected as b on a.STUDYID=b.STUDYID and a.USUBJID=b.USUBJID and a.PARAMCD=b.PARAMCD and b.AVISITN<a.AVISITN;24quit;25proc sort data=candidates; by STUDYID USUBJID PARAMCD AVISITN descending sourcevisit; run;26data carried; length DTYPE $11; set candidates; by STUDYID USUBJID PARAMCD AVISITN; if first.AVISITN; DTYPE='LOCF'; run;27data result; length DTYPE $11; set collected carried; run;2829proc sql;30  create table final as select STUDYID, USUBJID, PARAMCD, AVISIT, AVISITN, AVAL, DTYPE, EFFFL from result;31quit;32proc export data=final outfile="/app/output/adqs.csv" dbms=csv replace; run;