PANOPTES
panoptes-e86ef9e3

calib·HumanEval/33

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
1.000
inter-judge spread
0.000
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 sort_third to fn_237b0ff87e so the judges can't pattern-match a memorized HumanEval solution.



def fn_237b0ff87e(l: list):
    """This function takes a list l and returns a list l' such that
    l' is identical to l in the indicies that are not divisible by three, while its values at the indicies that are divisible by three are equal
    to the values of the corresponding indicies of l, but sorted.
    >>> fn_237b0ff87e([1, 2, 3])
    [1, 2, 3]
    >>> fn_237b0ff87e([5, 6, 3, 4, 8, 9, 2])
    [2, 6, 3, 4, 8, 9, 5]
    """
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.



def fn_237b0ff87e(l: list):
    """This function takes a list l and returns a list l' such that
    l' is identical to l in the indicies that are not divisible by three, while its values at the indicies that are divisible by three are equal
    to the values of the corresponding indicies of l, but sorted.
    >>> fn_237b0ff87e([1, 2, 3])
    [1, 2, 3]
    >>> fn_237b0ff87e([5, 6, 3, 4, 8, 9, 2])
    [2, 6, 3, 4, 8, 9, 5]
    """
    l = list(l)
    l[::3] = sorted(l[::3])
    return l
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
1.000
$0.0082
gpt-4o
1.000
$0.0032
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 = 0.325/ log 102.303
2 semantic clusters via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=9 · c2=1
semantic-entropy
gpt-4o
H = 0.325/ log 102.303
2 semantic clusters via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=9 · c2=1

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 1.000 · 1967 tokens · 3700ms

The solution correctly extracts elements at indices divisible by 3 (0, 3, 6, ...) using slice `l[::3]`, sorts them, and assigns them back. This matches both provided examples: [1,2,3]→[1,2,3] and [5,6,3,4,8,9,2]→[2,6,3,4,8,9,5]. It also handles edge cases: empty list returns [], single element returns it unchanged, and the original list is not mutated (copy is made). The implementation is idiomatic, concise, and correct.

gpt-4o
score 1.000 · 1071 tokens · 1445ms

The candidate solution correctly implements the task as described. It sorts the elements at indices divisible by three while leaving other elements unchanged. The solution passes the provided examples and handles edge cases such as empty lists and lists with fewer than three elements. The use of slicing and sorting is idiomatic and clear.