PANOPTES
panoptes-e86ef9e3

calib·HumanEval/9

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 rolling_max to fn_c58bbfc44a so the judges can't pattern-match a memorized HumanEval solution.

from typing import List, Tuple


def fn_c58bbfc44a(numbers: List[int]) -> List[int]:
    """ From a given list of integers, generate a list of rolling maximum element found until given moment
    in the sequence.
    >>> fn_c58bbfc44a([1, 2, 3, 2, 3, 4, 2])
    [1, 2, 3, 3, 3, 4, 4]
    """
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_c58bbfc44a(numbers: List[int]) -> List[int]:
    """ From a given list of integers, generate a list of rolling maximum element found until given moment
    in the sequence.
    >>> fn_c58bbfc44a([1, 2, 3, 2, 3, 4, 2])
    [1, 2, 3, 3, 3, 4, 4]
    """
    running_max = None
    result = []

    for n in numbers:
        if running_max is None:
            running_max = n
        else:
            running_max = max(running_max, n)

        result.append(running_max)

    return result
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
$0.0076
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 = 0.000/ log 102.303
1 semantic cluster via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=10
semantic-entropy
gpt-4o
H = 0.500/ log 102.303
2 semantic clusters via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=8 · c2=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 1.000 · 1868 tokens · 3902ms

The solution correctly implements rolling maximum. It handles the canonical example [1,2,3,2,3,4,2] → [1,2,3,3,3,4,4]. Edge cases: empty list returns [] (loop doesn't execute), single element returns that element, negative numbers work correctly since max() handles them, and a list of all equal numbers returns the same values. The logic is clean and idiomatic — using None as sentinel to initialize running_max is a standard Python pattern. No bugs detected.

gpt-4o
score 1.000 · 1007 tokens · 1583ms

The candidate solution correctly implements the rolling maximum functionality. It initializes a running maximum and iterates through the list, updating the running maximum as needed and appending it to the result list. The solution handles edge cases such as an empty list (returning an empty list) and lists with negative numbers or duplicates correctly. The implementation is clear and idiomatic, matching the expected output in the provided example.