PANOPTES
panoptes-e86ef9e3

calib·HumanEval/11

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
0.900
inter-judge spread
0.200
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 string_xor to fn_ba2f4acae5 so the judges can't pattern-match a memorized HumanEval solution.

from typing import List


def fn_ba2f4acae5(a: str, b: str) -> str:
    """ Input are two strings a and b consisting only of 1s and 0s.
    Perform binary XOR on these inputs and return result also as a string.
    >>> fn_ba2f4acae5('010', '110')
    '100'
    """
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


def fn_ba2f4acae5(a: str, b: str) -> str:
    """ Input are two strings a and b consisting only of 1s and 0s.
    Perform binary XOR on these inputs and return result also as a string.
    >>> fn_ba2f4acae5('010', '110')
    '100'
    """
    def xor(i, j):
        if i == j:
            return '0'
        else:
            return '1'

    return ''.join(xor(x, y) for x, y in zip(a, b))
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
0.800
$0.0079
gpt-4o
1.000
$0.0037
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 = 0.611/ log 102.303
2 semantic clusters via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=7 · c2=3

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 0.800 · 1846 tokens · 4436ms

The solution correctly implements binary XOR for equal-length strings and passes the canonical example ('010' XOR '110' = '100'). The `xor` helper correctly returns '0' when bits are equal and '1' otherwise. However, using `zip` silently truncates when the strings have different lengths — if `a` and `b` differ in length, the extra characters are ignored rather than raising an error or handling them. The task description doesn't explicitly address unequal lengths, but this is a plausible edge case. The logic is otherwise sound and idiomatic.

gpt-4o
score 1.000 · 1016 tokens · 2077ms

The candidate solution correctly implements the binary XOR operation for two input strings consisting of '1's and '0's. The solution defines a helper function `xor` that returns '0' if the bits are the same and '1' if they are different, which is the correct behavior for XOR. The use of `zip` ensures that the XOR operation is applied pairwise to corresponding bits from both strings, and `join` is used to concatenate the results into a single string. The example provided in the docstring is correctly handled by the implementation. The solution is idiomatic, clear, and handles edge cases such as empty strings correctly due to the use of `zip`.