Solution #1: Using ROW_NUMBER with random.value functions

WITH x AS ( 
SELECT CASE WHEN sal<=999  THEN '0-999' 
            WHEN sal<=1999 THEN '1000-1999' 
            WHEN sal<=2999 THEN '2000-2999' 
            ELSE                '3000+' 
       END range, 
       ename, sal 
FROM scott.emp 
), y AS ( 
SELECT ename, sal, range,  
       ROW_NUMBER()OVER(PARTITION BY range  
                        ORDER BY dbms_random.value) rn 
FROM x 
) 
SELECT range, ename, sal 
FROM y 
WHERE rn<=2 
ORDER BY range

Solution #2: Using DECODE, MAX() KEEP and UNION ALL

WITH x AS ( 
SELECT DECODE(1, SIGN(999-sal), '0-999', SIGN(1999-sal), '1000-1999', 
                 SIGN(2999-sal), '2000-2999', '3000+') range, 
       ename, sal, ROWNUM || dbms_random.value rnd 
FROM scott.emp 
) 
SELECT range, MAX(ename)KEEP(DENSE_RANK FIRST ORDER BY rnd) ename, 
              MAX(sal)  KEEP(DENSE_RANK FIRST ORDER BY rnd) sal 
FROM x 
GROUP BY range 
UNION ALL 
SELECT range, MAX(ename)KEEP(DENSE_RANK LAST ORDER BY rnd) ename, 
              MAX(sal)  KEEP(DENSE_RANK LAST ORDER BY rnd) sal 
FROM x 
GROUP BY range 
ORDER BY range

Solution #3: Using SIN for random value simulation and multi-column UNPIVOT with MAX() KEEP function

WITH x AS ( 
SELECT DECODE(1, SIGN(999-sal), '0-999', SIGN(1999-sal), '1000-1999', 
                 SIGN(2999-sal), '2000-2999', '3000+') range, 
       ename, sal,  
       SIN(ROWNUM*TO_NUMBER(SUBSTR( 
                             extract(second  
                                     from current_timestamp),-3)) 
           ) rnd 
FROM scott.emp 
), y AS ( 
SELECT range, MAX(ename)KEEP(DENSE_RANK FIRST ORDER BY rnd) ename1, 
              MAX(sal) KEEP(DENSE_RANK FIRST ORDER BY rnd) sal1, 
              MAX(ename)KEEP(DENSE_RANK LAST ORDER BY rnd) ename2, 
              MAX(sal) KEEP(DENSE_RANK LAST ORDER BY rnd) sal2 
FROM x 
GROUP BY range 
) 
SELECT range, ename, sal 
FROM y 
UNPIVOT ( 
  (ename, sal) for (t1, t2) in ((ename1,sal1), (ename2,sal2)) 
) 
ORDER BY range