PANOPTES
panoptes-e86ef9e3

calib·HumanEval/19

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
17
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_numbers to fn_7b5e464ce3 so the judges can't pattern-match a memorized HumanEval solution.

from typing import List


def fn_7b5e464ce3(numbers: str) -> str:
    """ Input is a space-delimited string of numberals from 'zero' to 'nine'.
    Valid choices are 'zero', 'one', 'two', 'three', 'four', 'five', 'six', 'seven', 'eight' and 'nine'.
    Return the string with numbers sorted from smallest to largest
    >>> fn_7b5e464ce3('three one five')
    'one three five'
    """
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_7b5e464ce3(numbers: str) -> str:
    """ Input is a space-delimited string of numberals from 'zero' to 'nine'.
    Valid choices are 'zero', 'one', 'two', 'three', 'four', 'five', 'six', 'seven', 'eight' and 'nine'.
    Return the string with numbers sorted from smallest to largest
    >>> fn_7b5e464ce3('three one five')
    'one three five'
    """
    value_map = {
        'zero': 0,
        'one': 1,
        'two': 2,
        'three': 3,
        'four': 4,
        'five': 5,
        'six': 6,
        'seven': 7,
        'eight': 8,
        'nine': 9
    }
    return ' '.join(sorted([x for x in numbers.split(' ') if x], key=lambda x: value_map[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
0.950
$0.0086
gpt-4o
1.000
$0.0035
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 71.946
1 semantic cluster via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=7

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 · 2008 tokens · 5604ms

The solution correctly implements the task: it splits the input string by spaces, filters out empty strings (handling edge cases like extra spaces), sorts the words using a value map from word to integer, and joins them back. The canonical example 'three one five' → 'one three five' works correctly. Edge cases like empty string input would return '' (correct). The only minor concern is that splitting by ' ' (single space) rather than using split() without arguments means multiple consecutive spaces would produce empty strings that are filtered out, which is handled by the `if x` filter. This is correct behavior. The solution is clean, idiomatic, and handles all plausible edge cases well.

gpt-4o
score 1.000 · 1089 tokens · 1811ms

The candidate solution correctly implements the task requirements. It defines a mapping from string numerals to their integer values, splits the input string into individual numerals, filters out any empty strings, and sorts them based on their integer values using the mapping. The sorted numerals are then joined back into a space-delimited string. The solution handles the example case correctly and is robust to edge cases such as empty input or repeated numerals. The code is clear and idiomatic.