SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2608.08239 · 2026

The Replay Gap: Static Evaluation of Model Switching in LLM Agents Scores the Wrong World

Ashritha Gonuguntla

arXiv · PDF · Open in the Atlas

Code that ran

6 of 6 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 3 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 3 failures of our instrument; 0 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/2608.08239.json · badge: /api/ran/2608.08239.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
AshrithaG/replay-gap canonical (classifier label)
classifier basis: census
6 ran · 0 no recorded run
FunctionStatusWhere it lives
action_divergence Ran
ran; our contract draft was wrong, not the code
AshrithaG/replay-gap/src/replay_gap/metrics.py
code served (permissive licence) · get_code("3dafafdf0809bf82")
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.

# Setup code for action_divergence function

def levenshtein(a: list[str], b: list[str]) -> int:
    if not a:
        return len(b)
    if not b:
        return len(a)
    prev = list(range(len(b) + 1))
    for i, x in enumerate(a, 1):
        cur = [i]
        for j, y in enumerate(b, 1):
            cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (x != y)))
        prev = cur
    return prev[-1]

def action_divergence(base_cmds: list[str], branch_cmds: list[str]) -> dict:
    """Edit distance (normalized) + index of first differing action."""
    dist = levenshtein(base_cmds, branch_cmds)
    denom = max(len(base_cmds), len(branch_cmds), 1)
    first_diff = None
    for i, (x, y) in enumerate(zip(base_cmds, branch_cmds)):
        if x != y:
            first_diff = i
            break
    if first_diff is None and len(base_cmds) != len(branch_cmds):
        first_diff = min(len(base_cmds), len(branch_cmds))
    return {'edit_distance': dist, 'normalized_edit_distance': dist / denom, 'first_divergent_action': first_diff, 'n_actions_base': len(base_cmds), 'n_actions_branch': len(branch_cmds)}

# Construct arguments
base_cmds = ["move_forward", "turn_left", "pickup", "drop"]
branch_cmds = ["move_forward", "turn_right", "pickup", "move_backward"]

ARGS = {
    "base_cmds": base_cmds,
    "branch_cmds": branch_cmds
}
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 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
assistant_indices Ran
ran, honoured the contract we drafted
AshrithaG/replay-gap/src/replay_gap/branching.py
code served (permissive licence) · get_code("4eee563f1acfb14a")
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 int; 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 torch

torch.manual_seed(0)

# Construct a list of message dictionaries for the assistant_indices function
messages = [
    {'role': 'user', 'content': 'Hello'},
    {'role': 'assistant', 'content': 'Hi there'},
    {'role': 'user', 'content': 'How are you?'},
    {'role': 'assistant', 'content': 'I am doing well'},
    {'role': 'assistant', 'content': 'Thank you for asking'},
]

ARGS = {
    'messages': messages
}
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 list, ndim 1, dtype family int — 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
commands Ran
ran; our contract draft was wrong, not the code
AshrithaG/replay-gap/src/replay_gap/metrics.py
code served (permissive licence) · get_code("2759df64bedf7884")
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 torch

torch.manual_seed(0)

# Construct a valid messages list with assistant turns containing actions
messages = [
    {
        'role': 'user',
        'content': 'What can you do?'
    },
    {
        'role': 'assistant',
        'content': 'I can execute commands',
        'extra': {
            'actions': [
                {'command': 'ls -la'},
                {'command': 'pwd'}
            ]
        }
    },
    {
        'role': 'user',
        'content': 'Execute another command'
    },
    {
        'role': 'assistant',
        'content': 'Sure',
        'extra': {
            'actions': [
                {'command': 'echo hello'},
                {'command': 'date'}
            ]
        }
    },
    {
        'role': 'assistant',
        'content': 'And one more',
        'extra': {
            'actions': [
                {'command': 'whoami'}
            ]
        }
    }
]

from_step = 2

ARGS = {
    'messages': messages,
    'from_step': from_step
}
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 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
levenshtein Ran
ran, honoured the contract we drafted
AshrithaG/replay-gap/src/replay_gap/metrics.py
code served (permissive licence) · get_code("0ec3f8fbc6028884")
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)

# The function levenshtein takes two lists of strings and computes the Levenshtein distance
# It needs:
# - a: list[str]
# - b: list[str]

a = ["cat", "dog", "bird"]
b = ["bat", "fog"]

ARGS = {
    "a": a,
    "b": b,
}
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
n_assistant_steps Ran
ran, honoured the contract we drafted
AshrithaG/replay-gap/src/replay_gap/branching.py
code served (permissive licence) · get_code("cf3c08d85122d751")
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
import torch.nn as nn

# Construct a list of message dicts with mixed roles
messages = [
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi there"},
    {"role": "user", "content": "How are you?"},
    {"role": "assistant", "content": "I'm doing well"},
    {"role": "assistant", "content": "Thanks for asking"},
]

ARGS = {
    "messages": messages
}
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
prefix_for_fork Ran
ran; our contract draft was wrong, not the code
AshrithaG/replay-gap/src/replay_gap/branching.py
code served (permissive licence) · get_code("3a5eee7eec29353f")
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.

# Setup code for prefix_for_fork function

ARGS = {
    'messages': [
        {'role': 'user', 'content': 'Hello'},
        {'role': 'assistant', 'content': 'Hi there'},
        {'role': 'user', 'content': 'How are you?'},
        {'role': 'assistant', 'content': 'I am doing well'},
        {'role': 'user', 'content': 'Great!'},
        {'role': 'assistant', 'content': 'Thanks for asking'},
    ],
    'fork_step': 2,
}
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 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

Also present in other papers' code

Of 6 samples lifted from this paper's linked repositories, 6 carry a function identity and 0 do not. Of the 6 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 6 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

LLM routers promise efficiency by matching each request to the cheapest adequate model, and are increasingly applied per step inside multi-step agents. Yet agentic routers are evaluated like single-turn routers: by replaying logged trajectories and substituting another model's recorded outputs, assuming the rest of the trajectory is unaffected. We test this assumption with branching rollouts: we fork live SWE-bench agent trajectories at controlled points, rebuild the environment, continue each fork with a different model, and compare against same-model control forks that isolate sampling and replay noise. Across six paired runs (∼900 rollouts), swaps exceed their matched control floors by +0.25 to +0.66 normalized edit distance (multiplicity-corrected CIs exclude zero), rewriting 61-94% of post-fork actions; 74-77% of early swaps diverge at the first post-fork action, versus 6-35% of controls -leaving only 3% of replayed states valid. Divergence decreases with fork depth in both directions. All five outcome flips we observe occur in swap arms -upgrades rescuing unsolved instances, a downgrade losing the sole solve -and zero occur across 359 control forks. Scoring these same swaps with a log-stitching replay evaluator, replay mispredicts every success-relevant outcome call and predicts patches with 0.00-0.11 similarity to reality. Auditing the noise floor, temperature-0 "determinism" is configuration-dependent: FP8-served controls diverge on over 90% of forks while AWQ-served ones remain near-identical; and under tight budgets the stronger model more often exhausts its steps without submitting. Replay-based benchmarks score the wrong world for agentic routing; we release our harness and all trajectories.

For agents

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

get_harvested_code_for_paper("2608.08239")
get_code_for_paper("2608.08239")
have("2608.08239")

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

curl https://syntology.ai/api/ran/2608.08239.json

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

[![Syntology run record](https://syntology.ai/api/ran/2608.08239.svg)](https://syntology.ai/paper/2608.08239)

Connect an agent — have() is free.