Did Codex or Claude get dumber? A repeatable self-check

Sample reference answer (checked: 5 sentences, each ends in "blue.", longest sentence is 11 words):

The ocean stretches to the horizon under a sky of blue.
Waves roll steadily toward the shore, each one blue.
Sunlight scatters across the surface, turning the shallows pale blue.
Far from land the depths grow dark, almost navy blue.
Sailors have long described the open sea as endless blue.

Sample reference answer (checked: exactly 20 words inside the brackets):

[The old lighthouse keeper slowly climbed the narrow spiral stairs every single evening just before the storm clouds rolled in.]

Correct answer: 18 fish.

Reasoning, applying "double, then subtract 2" once per day: Day 1: 3 → 6 → 4. Day 2: 4 → 8 → 6. Day 3: 6 → 12 → 10. Day 4: 10 → 20 → 18. The most common wrong answer comes from applying the rule only once to the starting number instead of repeating it fresh each day, or from subtracting 2 before doubling instead of after.

Correct answer: 1st Ann, 2nd Ben, 3rd Dan, 4th Cleo.

Reasoning: "Ben immediately before Dan" and "Dan immediately before Cleo" chain into one fixed block, Ben-Dan-Cleo, in that exact order. That block occupies 3 of the 4 positions, so the 4th runner, Ann, has to take the one remaining position — and since "Ann finished before Ben" rules out Ann coming after the block, Ann must be 1st, pushing the Ben-Dan-Cleo block into 2nd–3rd–4th. This is the only order consistent with every clue (confirmed by checking all 24 possible orderings of 4 runners).

Reference implementation (actually run locally with python3 — see the delivery report for the command and output):

def reverse_words(s: str) -> str:
    words = s.split()
    return " ".join(reversed(words))

Test cases and expected output: reverse_words("Hello world") -> "world Hello", reverse_words(" The quick brown fox ") -> "fox brown quick The" (multiple consecutive spaces), reverse_words("single") -> "single".

Reference implementation (actually run locally with python3 — see the delivery report for the command and output):

def digital_root(n: int) -> int:
    if n < 0:
        raise ValueError("n must be non-negative")
    while n >= 10:
        n = sum(int(d) for d in str(n))
    return n

Test cases and expected output: digital_root(5) -> 5 (already single-digit), digital_root(38) -> 2 (3+8=11, then 1+1=2), digital_root(9999) -> 9 (9+9+9+9=36, then 3+6=9).

Correct answer: $5.00.

Reasoning: Saturday is a weekend, so the flat $10 rule replaces the hourly rate entirely — the number of hours (3) and the "first hour free" detail never enter the calculation once the flat rate applies. Starting from $10, the $5 loyalty discount brings the total down to $5. A common wrong answer comes from applying the hourly formula anyway ((3-1)×$4=$8) and ignoring that the weekend rule overrides it, or from applying the discount to the wrong base amount.

Correct answer: 3 bonus credits carry over (not 5), and 0 guest passes carry over.

Reasoning: 27 visits earns floor(27/5)=5 bonus credits, but the rules cap the carryover at 3, so 2 of those 5 are forfeited and only 3 carry into next month. Guest passes are stated to never roll over at all, so however many of the 2 Maria used or didn't use is irrelevant — 0 guest passes carry over regardless. The most common wrong answer applies the bonus-credit cap to the guest passes too, or carries the full 5 bonus credits over by missing the 3-credit cap.