PANOPTES
panoptes-e86ef9e3

calib·HumanEval/13

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
0.975
inter-judge spread
0.050
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 greatest_common_divisor to fn_f22f28074d so the judges can't pattern-match a memorized HumanEval solution.



def fn_f22f28074d(a: int, b: int) -> int:
    """ Return a greatest common divisor of two integers a and b
    >>> fn_f22f28074d(3, 5)
    1
    >>> fn_f22f28074d(25, 15)
    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_f22f28074d(a: int, b: int) -> int:
    """ Return a greatest common divisor of two integers a and b
    >>> fn_f22f28074d(3, 5)
    1
    >>> fn_f22f28074d(25, 15)
    5
    """
    while b:
        a, b = b, a % b
    return a
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
$0.0082
gpt-4o
1.000
$0.0029
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 = 0.639/ log 102.303
3 semantic clusters via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=8 · c2=1 · c3=1
semantic-entropy
gpt-4o
H = 0.000/ log 102.303
1 semantic cluster via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=10

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 · 1820 tokens · 5457ms

The solution implements the Euclidean algorithm correctly. It passes both provided examples: gcd(3,5)=1 and gcd(25,15)=5. It handles the general case correctly. One minor edge case: if both a=0 and b=0, it returns 0, which is a mathematically conventional result (gcd(0,0) is typically defined as 0). For negative integers, the algorithm returns a result with the sign of `a` (e.g., gcd(-6, 4) returns -2 instead of 2), which could be considered a bug depending on the specification, but the task doesn't mention negative inputs. Overall, the solution is idiomatic, efficient, and correct for the stated task.

gpt-4o
score 1.000 · 898 tokens · 1300ms

The candidate solution correctly implements the Euclidean algorithm to find the greatest common divisor (GCD) of two integers. It passes the provided examples and handles edge cases such as when one of the integers is zero, returning the absolute value of the other integer, which is the correct behavior for GCD. The implementation is idiomatic and clear.