PANOPTES
panoptes-e86ef9e3

calib·HumanEval/20

humaneval-calibprobe · code

This is one (task, candidate response) pair flowing through the full PANOPTES pipeline. Each section below is a stage of the analysis: the task itself, the candidate solution being evaluated, every judge's score + rationale, the sampling-pass dispersion that captures within-judge noise, and the uncertainty-quantification metrics computed on top.

mean score
0.725
inter-judge spread
0.450
judges polled
2
sampling draws
20
1. the task

The function signature + docstring presented to both the model under test and to every judge. The entry-point name was obfuscated from find_closest_elements to fn_59e2617f12 so the judges can't pattern-match a memorized HumanEval solution.

from typing import List, Tuple


def fn_59e2617f12(numbers: List[float]) -> Tuple[float, float]:
    """ From a supplied list of numbers (of length at least two) select and return two that are the closest to each
    other and return them in order (smaller number, larger number).
    >>> fn_59e2617f12([1.0, 2.0, 3.0, 4.0, 5.0, 2.2])
    (2.0, 2.2)
    >>> fn_59e2617f12([1.0, 2.0, 3.0, 4.0, 5.0, 2.0])
    (2.0, 2.0)
    """
2. the candidate response judged

The full solution every judge is grading. For these runs the candidate is the reference solution prepended with the task signature so it parses as a complete program.

from typing import List, Tuple


def fn_59e2617f12(numbers: List[float]) -> Tuple[float, float]:
    """ From a supplied list of numbers (of length at least two) select and return two that are the closest to each
    other and return them in order (smaller number, larger number).
    >>> fn_59e2617f12([1.0, 2.0, 3.0, 4.0, 5.0, 2.2])
    (2.0, 2.2)
    >>> fn_59e2617f12([1.0, 2.0, 3.0, 4.0, 5.0, 2.0])
    (2.0, 2.0)
    """
    closest_pair = None
    distance = None

    for idx, elem in enumerate(numbers):
        for idx2, elem2 in enumerate(numbers):
            if idx != idx2:
                if distance is None:
                    distance = abs(elem - elem2)
                    closest_pair = tuple(sorted([elem, elem2]))
                else:
                    new_distance = abs(elem - elem2)
                    if new_distance < distance:
                        distance = new_distance
                        closest_pair = tuple(sorted([elem, elem2]))

    return closest_pair
3a. point-pass scores (temperature 0)

One call per judge at temperature=0. This is the "best single guess" each judge has. Disagreement here is structural. The judges literally see this candidate differently.

claude-sonnet
0.950
$0.0083
gpt-4o
0.500
$0.0048
3b. sampling-pass dispersion (temperature 1)

n draws per judge at temperature=1. Dots are individual draws; vertical bar is the mean; shaded band is ±1σ. Wide band = the judge is uncertain even with itself; narrow band = it's consistent.

4. uncertainty quantification

Three statistical methods computed on top of the raw scores above. Each one tells you something different about how much to trust the headline number.

semantic-entropy
claude-sonnet
H = 0.000/ log 102.303
1 semantic cluster via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=10
semantic-entropy
gpt-4o
H = 1.221/ log 102.303
4 semantic clusters via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=5 · c2=2 · c3=1 · c4=2

How to read these: self-consistency = does the judge agree with itself when resampled? Narrower CI means more consistent. Semantic entropy = do the judge's rationales cluster into one meaning, or several? Higher entropy means the judge is internally conflicted about why. Decomposition = how much of the total variance comes from within-judge sampling noise (aleatoric, hard to fix) vs between-judge disagreement (epistemic, fixable by calling more judges).

5. judge rationales

The natural-language explanation each judge gave alongside its score. These are produced via tool-use structured output, so the score and the rationale are guaranteed to come from the same forward pass.

claude-sonnet
score 0.950 · 2082 tokens · 4264ms

The solution correctly finds the two closest numbers by comparing all pairs (O(n²)), handles equal values (distance=0), and returns them sorted as (smaller, larger). Both provided doctests would pass. The only minor issue is that it compares each pair twice (idx, idx2) and (idx2, idx), which is redundant but not incorrect. This is a minor efficiency concern, not a correctness bug. The solution handles edge cases like equal numbers and negative numbers correctly.

gpt-4o
score 0.500 · 1312 tokens · 2623ms

The candidate solution attempts to find the closest pair of numbers in a list by comparing each pair of numbers. However, it has a significant inefficiency and a logic flaw. The solution uses a nested loop to compare each pair of numbers, which results in O(n^2) complexity. While this is not incorrect, it is inefficient for large lists. More critically, the solution does not ensure that the returned pair is ordered as (smaller number, larger number) as required by the task. The use of `tuple(sorted([elem, elem2]))` does ensure ordering, but the logic for updating the closest pair is flawed because it does not correctly handle cases where multiple pairs have the same minimum distance. Additionally, the solution does not handle edge cases like lists with duplicate numbers correctly, as it may return the same number twice if they are the closest pair. Overall, the solution is partially correct but has significant issues that would surface on realistic input.