PANOPTES
panoptes-e86ef9e3

calib·HumanEval/7

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 filter_by_substring to fn_2733b90a7c so the judges can't pattern-match a memorized HumanEval solution.

from typing import List


def fn_2733b90a7c(strings: List[str], substring: str) -> List[str]:
    """ Filter an input list of strings only for ones that contain given substring
    >>> fn_2733b90a7c([], 'a')
    []
    >>> fn_2733b90a7c(['abc', 'bacd', 'cde', 'array'], 'a')
    ['abc', 'bacd', 'array']
    """
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_2733b90a7c(strings: List[str], substring: str) -> List[str]:
    """ Filter an input list of strings only for ones that contain given substring
    >>> fn_2733b90a7c([], 'a')
    []
    >>> fn_2733b90a7c(['abc', 'bacd', 'cde', 'array'], 'a')
    ['abc', 'bacd', 'array']
    """
    return [x for x in strings if substring in x]
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
$0.0071
gpt-4o
1.000
$0.0031
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 = 0.000/ 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 1.000 · 1800 tokens · 3739ms

The solution correctly implements the filter using a list comprehension with `substring in x`, which handles all cases: empty list returns [], strings containing the substring are included, and those without are excluded. Matches both doctest examples and handles edge cases like empty strings, multi-character substrings, and case sensitivity correctly. Clean and idiomatic Python.

gpt-4o
score 1.000 · 958 tokens · 1963ms

The candidate solution correctly implements the function to filter a list of strings based on the presence of a given substring. It uses a list comprehension to iterate over the input list and checks if the substring is present in each string. The solution passes the provided examples, including an edge case with an empty list. The implementation is idiomatic and clear, with no apparent bugs or inefficiencies.