SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.01170 · 2026

Online Reasoning Calibration: Test-Time Training Enables Generalizable Conformal LLM Reasoning

Tommi Jaakkola, Stephen Bates, Ashia Wilson, Zekai Wang, Menghua Wu, Cai Zhou, Chenyu Wang, Qianyu Zhu, Flora Shi

arXiv · PDF · Open in the Atlas

Code that ran

13 of 16 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (1 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 1 has no contract verdict; 11 failures of our instrument; 3 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2604.01170.json · badge: /api/ran/2604.01170.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
wzekai/ORCA community
Papers with Code archive
13 ran · 3 no recorded run
FunctionStatusWhere it lives
TTTProbe Ran
ran on a synthesized input · constructed an object; no result was computed
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("a2dc42bef2c35e5f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, TTTProbe: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import math
import torch.nn as nn
import torch.nn.functional as F

torch.manual_seed(0)

ARGS = {
    "d_phi": 4,
    "d_hidden": 8,
    "use_ln": False,
    "use_residual": False,
    "learnable_eta": False,
    "base_lr": 0.01,
    "share_kq": False,
    "no_kq": False,
    "use_mlp": False,
    "use_pca": False,
    "pca_dim": 256,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TTTProbe — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a TTTProbe output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
binom_p Ran
ran, violated the contract we drafted
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("da0ccfc6f8e9bdb3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from scipy.stats import binom

# Construct valid arguments for binom_p function
# The function takes:
# - loss: array-like of binary outcomes (0s and 1s)
# - delta: float, the null hypothesis probability

# Create a small loss array representing binary outcomes
loss = np.array([0, 1, 1, 0, 1, 1, 0, 0])

# Create a delta value between 0 and 1
delta = 0.5

ARGS = {
    'loss': loss,
    'delta': delta,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
gelu_bwd Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("636d4e375feacf84")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# The function gelu_bwd takes a single parameter x
# which is used in arithmetic operations (tanh, multiplication, etc.)
# So x should be a torch tensor

x = torch.randn(2, 4)

ARGS = {
    "x": x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_data_splits Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("5afda302a2e44381")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Construct metadata dict with splits information
metadata = {
    'splits': {
        'train': [0, 1, 2],
        'val': [3, 4],
        'test': [5, 6, 7]
    },
    'splits_wu': {
        'train': [0, 1],
        'val': [2, 3],
        'test': [4, 5, 6, 7]
    }
}

ARGS = {
    'metadata': metadata,
    'scheme': 'default'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_loss Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("9218d0d6347d01af")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Analyzing the function:
# - pred: a list/sequence that supports indexing and .index() method, elements compared with >= lam
# - true: a list/sequence that supports indexing and len()
# - lam: a scalar value used in comparison

# The function treats pred as a list of numbers and true as a list of binary values (0 or 1)
# It finds the first index where pred[i] >= lam, then returns 1 - true[idx]

pred = [0.2, 0.5, 0.8, 0.3]
true = [0, 1, 1, 0]
lam = 0.6

ARGS = {
    "pred": pred,
    "true": true,
    "lam": lam
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_run_dir Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("8a4bb8a49682c245")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import tempfile

class Args:
    def __init__(self):
        self.method = 'ttt'
        self.d_hidden = 64
        self.base_lr = 0.001
        self.use_ln = True
        self.use_residual = True
        self.learnable_eta = True
        self.share_kq = True
        self.use_mlp = False
        self.use_pca = False
        self.pca_dim = 256
        self.no_meta_train = False
        self.no_online_update = False
        self.epochs = 100
        self.outer_lr = 0.001
        self.smooth_window = 10
        self.no_kq = False
        self.output_dir = tempfile.mkdtemp()
        self.run_name = None
        self.label_mode = 'supervised'

ARGS = {
    'args': Args()
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
ln_bwd Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("70a242e297e815b3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create small tensors
batch_size = 2
feature_dim = 4

# x: input tensor, shape (batch_size, feature_dim)
x = torch.randn(batch_size, feature_dim)

# grad_output: gradient w.r.t. output, same shape as x
grad_output = torch.randn(batch_size, feature_dim)

# gamma: scale parameter for LayerNorm, shape (feature_dim,)
gamma = torch.randn(feature_dim)

# eps: small constant for numerical stability (default provided)
eps = 1e-06

ARGS = {
    "x": x,
    "grad_output": grad_output,
    "gamma": gamma,
    "eps": eps
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
ln_fwd Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("e6563e79a4cb495a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create small tensors for layer normalization
batch_size = 2
feature_dim = 4

# Input tensor: (batch_size, feature_dim)
x = torch.randn(batch_size, feature_dim)

# Learnable parameters: gamma (scale) and beta (shift)
# These should have shape (feature_dim,) to match the last dimension of x
gamma = torch.ones(feature_dim)
beta = torch.zeros(feature_dim)

# Epsilon for numerical stability
eps = 1e-6

ARGS = {
    'x': x,
    'gamma': gamma,
    'beta': beta,
    'eps': eps
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_dataset Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("5ae740cdec591624")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import pickle
import tempfile
import os

# Create a temporary pickle file with mock dataset structure
temp_dir = tempfile.mkdtemp()
dataset_path = os.path.join(temp_dir, 'dataset.pkl')

mock_data = {
    'problems': [
        {
            'step_embeddings': [[0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8]],
            'step_labels': [0, 1],
            'step_labels_consistent': [0, 1]
        },
        {
            'step_embeddings': [[0.2, 0.3, 0.4, 0.5], [0.6, 0.7, 0.8, 0.9], [0.1, 0.1, 0.1, 0.1]],
            'step_labels': [1, 1, 0],
            'step_labels_consistent': [1, 1, 0]
        }
    ]
}

with open(dataset_path, 'wb') as f:
    pickle.dump(mock_data, f)

ARGS = {
    'dataset_path': dataset_path,
    'label_mode': 'supervised'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_datasets Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("2f2303b2ebebeabe")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import pickle
import tempfile
import os
import torch
import numpy as np

torch.manual_seed(0)

# Create temporary directory for pickle files
tmpdir = tempfile.mkdtemp()

# Create first dataset pickle file
dataset1_path = os.path.join(tmpdir, 'dataset1.pkl')
dataset1 = {
    'problems': [
        {
            'step_embeddings': np.random.randn(3, 4).tolist(),
            'step_labels': [0, 1, 1],
            'step_labels_consistent': [0, 1, 1]
        },
        {
            'step_embeddings': np.random.randn(2, 4).tolist(),
            'step_labels': [1, 1],
            'step_labels_consistent': [1, 1]
        }
    ],
    'embed_dim': 4,
    'batch_size': 10,
    'splits': {
        'train': [0],
        'calibration': [],
        'test': [1]
    }
}
with open(dataset1_path, 'wb') as f:
    pickle.dump(dataset1, f)

# Create second dataset pickle file
dataset2_path = os.path.join(tmpdir, 'dataset2.pkl')
dataset2 = {
    'problems': [
        {
            'step_embeddings': np.random.randn(4, 4).tolist(),
            'step_labels': [0, 0, 1, 1],
            'step_labels_consistent': [0, 0, 1, 1]
        }
    ],
    'embed_dim': 4,
    'batch_size': 5,
    'splits': {
        'train': [0],
        'calibration': [],
        'test': []
    }
}
with open(dataset2_path, 'wb') as f:
    pickle.dump(dataset2, f)

ARGS = {
    'dataset_paths': [dataset1_path, dataset2_path],
    'label_mode': 'supervised'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_run_name Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("fcf82641bf583592")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

class Args:
    def __init__(self):
        self.method = 'ttt'
        self.no_kq = False
        self.d_hidden = 64
        self.base_lr = 0.001
        self.use_ln = True
        self.use_residual = True
        self.learnable_eta = True
        self.share_kq = True
        self.use_mlp = True
        self.use_pca = True
        self.pca_dim = 256
        self.no_meta_train = False
        self.no_online_update = False
        self.epochs = 150
        self.outer_lr = 0.002
        self.smooth_window = 15

ARGS = {
    'args': Args()
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
run_test Ran
ran; our contract draft was wrong, not the code
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("84570830f93e1e84")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

def create_args():
    # Analyzing run_test signature:
    # - preds: list of prediction sequences (used in zip with trues)
    # - trues: list of true label sequences (used in zip with preds)
    # - delta: float, target risk upper bound
    # - epsilon: float, FWER confidence level (default 0.05)
    # - bins: int, number of lambda values to sweep (default 10000)
    
    # From get_loss function:
    # - pred: sequence of floats (compared with lam via >=)
    # - true: sequence with indexable elements, last element true[-1]
    # - Returns 1 - true[idx] where idx is an index
    # So true should contain values between 0 and 1 (binary-like)
    
    # Create small synthetic data
    np.random.seed(0)
    
    # preds: list of prediction sequences (floats in [0,1])
    preds = [
        [0.9, 0.3, 0.7],  # prediction sequence 1
        [0.5, 0.2, 0.8],  # prediction sequence 2
    ]
    
    # trues: list of true label sequences (binary values 0 or 1)
    trues = [
        [1, 0, 1],  # true labels sequence 1
        [0, 1, 1],  # true labels sequence 2
    ]
    
    # delta: target risk upper bound
    delta = 0.1
    
    # epsilon: FWER confidence level
    epsilon = 0.05
    
    # bins: number of lambda values to sweep (keep small for fast execution)
    bins = 100
    
    return {
        'preds': preds,
        'trues': trues,
        'delta': delta,
        'epsilon': epsilon,
        'bins': bins
    }

ARGS = create_args()
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
smooth Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("7ca8a94ba75bdc59")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from collections import deque

# Create small prediction arrays for testing
pred = np.array([1.0, 2.0, 3.0, 4.0, 5.0])
window = 2

ARGS = {
    "pred": pred,
    "window": window
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
build_ttt_probe Not yet run
unverified (harvested, no recorded run)
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("4d4c1a03a41eb8bc")
How it ran

no invocation recorded.

calibrate_ttt Not yet run
unverified (harvested, no recorded run)
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("8a1b9fa2efea9911")
How it ran

no invocation recorded.

to_cumulative Not yet run
unverified (harvested, no recorded run)
wzekai/ORCA/code/calibrate.py
code served (permissive licence) · get_code("2175b04ff110081a")
How it ran

no invocation recorded.

Also present in other papers' code

Of 16 samples lifted from this paper's linked repositories, 16 carry a function identity and 0 do not. Of the 16 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 16 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

While test-time scaling has advanced the performance of large language models, state-of-the-art results impose a heavy computational burden. Better dynamic allocation of test-time compute would help mitigate this cost, but one key challenge is the lack of calibration in post-trained language models. Here, we present Online Reasoning Calibration (ORCA), a technique for calibrating the sampling process that draws upon conformal prediction and test-time training. Specifically, we introduce a meta-learning procedure that updates the calibration module for each input. This allows us to provide valid confidence estimates under distributional shift, such as that in thought patterns across different stages of reasoning, or in prompt distributions between model development and deployment. ORCA not only provides theoretical guarantees on conformal risks, but also empirically shows higher efficiency and generalization across different reasoning tasks. At risk level δ=0.1, ORCA improves Qwen2.5-32B efficiency on in-distribution tasks with savings up to 47.5% with supervised labels and 40.7% with self-consistency labels. Under zero-shot out-of-domain settings, it improves MATH-500 savings from 24.8% of the static calibration baseline to 67.0% while maintaining a low empirical error rate, and the same trend holds across model families and downstream benchmarks. Our code is publicly available at https://github.com/wzekai/ORCA.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2604.01170")
get_code_for_paper("2604.01170")
have("2604.01170")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2604.01170.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2604.01170.svg)](https://syntology.ai/paper/2604.01170)

Connect an agent — have() is free.