PANOPTES
panoptes-696da4d5

HE/15

humaneval · 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
3
sampling draws
15
1. the task

The function signature + docstring presented to both the model under test and to every judge.



def string_sequence(n: int) -> str:
    """ Return a string containing space-delimited numbers starting from 0 upto n inclusive.
    >>> string_sequence(0)
    '0'
    >>> string_sequence(5)
    '0 1 2 3 4 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 string_sequence(n: int) -> str:
    """ Return a string containing space-delimited numbers starting from 0 upto n inclusive.
    >>> string_sequence(0)
    '0'
    >>> string_sequence(5)
    '0 1 2 3 4 5'
    """
    return ' '.join([str(x) for x in range(n + 1)])
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-haiku
1.000
$0.0027
claude-sonnet
1.000
$0.0071
gpt-4o-mini
1.000
$0.0002
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.

decomposition
aggregate (all judges)
Var = 0.00e+0
total predictive variance, 3 judges
aleatoric (irreducible)
0.00e+0
epistemic (reducible)
0.00e+0
self-consistency
claude-haiku
1.000[1.000, 1.000]
posterior mean · 90% Bayesian-bootstrap CI
variance: 0.00e+0
n samples: 5
self-consistency
gpt-4o-mini
1.000[1.000, 1.000]
posterior mean · 90% Bayesian-bootstrap CI
variance: 0.00e+0
n samples: 5
self-consistency
claude-sonnet
1.000[1.000, 1.000]
posterior mean · 90% Bayesian-bootstrap CI
variance: 0.00e+0
n samples: 5
semantic-entropy
claude-haiku
H = 0.000/ log 51.609
1 semantic cluster via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=5
semantic-entropy
gpt-4o-mini
H = 0.000/ log 51.609
1 semantic cluster via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=5
semantic-entropy
claude-sonnet
H = 0.000/ log 51.609
1 semantic cluster via bidirectional NLI
entropy as fraction of max
cluster sizes
c1=5

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-haiku
score 1.000 · 1797 tokens · 2744ms

The solution correctly implements the required functionality. It uses `range(n + 1)` to generate numbers from 0 to n inclusive, converts each to a string, and joins them with spaces. The logic is sound: `range(n + 1)` produces [0, 1, ..., n], the list comprehension converts each to string, and `' '.join()` produces the space-delimited output. Both provided examples pass: `string_sequence(0)` returns '0' and `string_sequence(5)` returns '0 1 2 3 4 5'. Edge cases are handled correctly: negative n would produce an empty range (empty string), and large n would work as expected. The code is idiomatic Python and clear.

claude-sonnet
score 1.000 · 1741 tokens · 3172ms

The solution correctly generates a space-delimited string of numbers from 0 to n inclusive using `range(n + 1)` and `' '.join(...)`. It handles the edge case of n=0 (returns '0') and the general case (e.g., n=5 returns '0 1 2 3 4 5'). The implementation is idiomatic, clear, and efficient.

gpt-4o-mini
score 1.000 · 867 tokens · 1523ms

The solution correctly implements the task by generating a space-delimited string of numbers from 0 to n inclusive. It passes the provided examples and handles edge cases, such as n = 0, correctly. The code is clear and idiomatic, using a list comprehension and the join method effectively.