SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.17486 · 2026

SALT: Salience-Aware Lexical Trie for Long-Context Compression

Shangqian Gao, Oteo Mamo, Hyunjin Yi, Joydhriti Choudhury, Weikuan Yu

arXiv · PDF

Code that ran

24 of 27 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, 2 violated it and 0 have no contract verdict; 19 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/2607.17486.json · badge: /api/ran/2607.17486.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
oteomamo/SALT canonical (classifier label)
classifier basis: census
24 ran · 3 no recorded run
FunctionStatusWhere it lives
append Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/ledger.py
code served (permissive licence) · get_code("3af1fd763069130b")
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 json
from pathlib import Path
import tempfile
import os

# Create a temporary directory for the session
temp_dir = tempfile.mkdtemp()

# Create a delegation record to append
rec = {
    "delegation_id": "deleg_001",
    "delegator": "user_123",
    "delegatee": "user_456",
    "permissions": ["read", "write"],
    "timestamp": "2024-01-15T10:30:00Z",
    "active": True
}

ARGS = {
    "session_dir": temp_dir,
    "rec": rec
}
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
ask_signals Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/route.py
code served (permissive licence) · get_code("01b4dda99b654372")
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

ARGS = {
    'ask': "What are the three main tasks?\n- First task\n- Second task\n* Third task\n\nAlso, can you help? Yes, I can! Check items: 1) Item one 2) Item two. Contact @john for details."
}
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 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
bind Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/personas.py
code served (permissive licence) · get_code("767e18040cb2ca99")
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 torch.nn as nn

# Define a small Persona class to represent personas
class Persona:
    def __init__(self, name, worker, path):
        self.name = name
        self.worker = worker
        self.path = path

# Define a small RosterEntry class to represent roster entries
class RosterEntry:
    def __init__(self, name, role):
        self.name = name
        self.role = role

# Define a small Roster class
class Roster:
    def __init__(self, entries):
        self.entries = entries

# Create valid test data
personas = [
    Persona(name='persona1', worker='chat', path='/path/to/persona1'),
    Persona(name='persona2', worker='worker_a', path='/path/to/persona2'),
]

roster_entries = [
    RosterEntry(name='worker_a', role='worker'),
    RosterEntry(name='worker_b', role='worker'),
]

roster = Roster(entries=roster_entries)

ARGS = {
    'personas': personas,
    'roster': roster,
}
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 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
build_context Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oteomamo/SALT/salt/agents/delegate.py
code served (permissive licence) · get_code("1567aa16a9422ba2")
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 DelegationContext. 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
from dataclasses import dataclass, field
from collections import defaultdict

torch.manual_seed(0)

# Stand-in classes for the required objects

class TrieNode:
    def __init__(self):
        self.n_sentences = 5
    
    def compress(self, query, budget_pct, tokenizer, model, device, 
                 coverage_half_life, coverage_decay_docs, shift_damping, 
                 shift_margin, shift_query_boost, per_source_themes, 
                 max_words, stable_keys, coverage_gc, coverage_max_keys, 
                 defer_commit, exclude_sent_idx):
        return {
            'selected_sent_idx': [0, 1, 3],
            'stats': {'words_used': 150, 'coverage': 0.85}
        }

class BGETokenizer:
    def __call__(self, text):
        return torch.tensor([[1, 2, 3, 4], [5, 6, 7, 8]])
    
    def encode(self, text):
        return [1, 2, 3, 4]

class BGEModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def forward(self, input_ids):
        return self.linear(input_ids.float())

class IngestDrain:
    def drain(self):
        return []

class Request:
    def __init__(self):
        self.query = "test query"
        self.budget_pct = 0.8
        self.switches = {'coverage_half_life': 1.0, 'coverage_decay_docs': 10}

class State:
    def __init__(self):
        self.trie = TrieNode()
        self.bge_tok = BGETokenizer()
        self.bge_model = BGEModel()
        self.bge_device = 'cpu'
        self.budget = 0.75
        self.ingest = IngestDrain()
        self.turn_labels = {}
        self.conversation_map = {}
        self.offload_context_cap = 5000
        self.coverage_half_life = 1.0
        self.coverage_decay_docs = 10
        self.shift_damping = 0.5
        self.shift_margin = 0.1
        self.shift_query_boost = 1.5
        self.per_source_themes = False
        self.stable_coverage_keys = False
        self.coverage_gc = True
        self.coverage_max_keys = 100

def format_memory_block(trie, selected, turn_labels, conversation_map):
    return "memory block text"

def memory_word_cap(state, query):
    return 2000

def report_ingest_failures(drain_result):
    pass

# Mock the salt.chat.cli module functions
import sys
from types import ModuleType
salt_chat_cli = ModuleType('salt.chat.cli')
salt_chat_cli.format_memory_block = format_memory_block
salt_chat_cli.memory_word_cap = memory_word_cap
salt_chat_cli.report_ingest_failures = report_ingest_failures
sys.modules['salt'] = ModuleType('salt')
sys.modules['salt.chat'] = ModuleType('salt.chat')
sys.modules['salt.chat.cli'] = salt_chat_cli

state = State()
req = Request()

ARGS = {
    'state': state,
    'req': req
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
salt, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DelegationContext — a DelegationContext output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DelegationContext 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
build_messages Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/delegate.py
code served (permissive licence) · get_code("d1f4954d72388c6d")
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
from pathlib import Path

class MockContext:
    def __init__(self, text):
        self.text = text

class MockRequest:
    def __init__(self, task):
        self.task = task

context = MockContext("This is some context text that provides background information for the task.")

req = MockRequest("Extract the main topic from the context above.")

ARGS = {
    'context': context,
    'req': req
}
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
content_words Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/provision.py
code served (permissive licence) · get_code("8cfe218c7dea294a")
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 set. 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 re

STOPWORDS = frozenset('a an and are as at be by for from has have in is it its no not of on or that the these this those to up what with'.split())

_WORD_RE = re.compile('[a-z0-9]+')

def content_words(note):
    return {w for w in _WORD_RE.findall((note or '').lower()) if w not in STOPWORDS and (not w.isdigit())}

# Setup code to construct valid arguments
ARGS = {
    'note': 'The quick brown fox jumps over the lazy dog with amazing skills'
}
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 set — a set output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a set 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
find_kneedle Ran
ran, honoured the contract we drafted
oteomamo/SALT/salt/engine/trie_core.py
code served (permissive licence) · get_code("0f5993b0b30ddafa")
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 numpy as np

ARGS = {
    "steps": [
        {"step": 0, "cos": 1.0},
        {"step": 10, "cos": 0.8},
        {"step": 20, "cos": 0.6},
        {"step": 30, "cos": 0.5},
        {"step": 40, "cos": 0.45},
        {"step": 50, "cos": 0.44},
    ],
    "max_keywords_ratio": 0.3,
}
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 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
is_content_word Ran
ran, violated the contract we drafted
oteomamo/SALT/salt/engine/trie_core.py
code served (permissive licence) · get_code("dcc5c7f539ccb999")
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 bool, 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 torch

ARGS = {
    'w': 'hello'
}
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 bool, 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
l2_norm Ran
ran, violated the contract we drafted
oteomamo/SALT/salt/engine/trie_core.py
code served (permissive licence) · get_code("82b8d46e62a2da66")
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 ndarray, 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 numpy as np

def l2_norm(v):
    norm = np.linalg.norm(v, axis=-1, keepdims=True)
    return v / np.maximum(norm, 1e-12)

# The function takes a single parameter 'v' which should be a numpy array
# It computes the L2 norm along the last axis and normalizes the input

# Create a small test vector with shape (2, 4) - batch size 2, dimension 4
v = np.array([
    [1.0, 2.0, 3.0, 4.0],
    [5.0, 6.0, 7.0, 8.0]
], dtype=np.float32)

ARGS = {
    "v": v
}
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 ndarray, 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
ledger_path Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/ledger.py
code served (permissive licence) · get_code("3a81a6942e639f31")
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 PosixPath. 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.

from pathlib import Path

ARGS = {
    'session_dir': '/tmp/test_session'
}
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 PosixPath — a PosixPath output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PosixPath 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
load_persona Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/personas.py
code served (permissive licence) · get_code("a8589040349e3800")
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 Persona. 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 tempfile
from pathlib import Path

# Create a temporary valid persona file
temp_dir = tempfile.mkdtemp()
persona_file = Path(temp_dir) / "test_persona.txt"

# Write a valid persona file with all required fields
persona_content = """---
name: alice_persona
worker: gpt4
role: target
notes: A test target persona for validation
---
You are a helpful AI assistant named Alice. 
Respond to all queries with enthusiasm and clarity.
Always try to be helpful and informative."""

persona_file.write_text(persona_content, encoding='utf-8')

ARGS = {
    'path': str(persona_file)
}
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 Persona — a Persona output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Persona 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
load_personas Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/personas.py
code served (permissive licence) · get_code("d339d5d3738be461")
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, 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 tempfile
from pathlib import Path

# Create temporary directory with persona files
temp_dir = tempfile.TemporaryDirectory()
temp_path = Path(temp_dir.name)

# Create a valid persona file
persona_file = temp_path / "alice.md"
persona_file.write_text("""---
name: alice
worker: gpt4
role: target
notes: A helpful assistant persona
---
You are Alice, a helpful and friendly AI assistant. Answer questions clearly and concisely.""")

# Create another valid persona file with verify role
persona_file2 = temp_path / "bob.md"
persona_file2.write_text("""---
name: bob
worker: chat
role: verify
---
You are Bob, a verification assistant. Check the accuracy of responses.""")

# The paths parameter should be a list/sequence of file or directory paths
ARGS = {
    'paths': [temp_path]
}
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, ndim 1, dtype family float — a tuple 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
note_for Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/provision.py
code served (permissive licence) · get_code("13781df55d2f0b12")
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

class Job:
    def __init__(self):
        self.notes = "Job notes: "
        self.window = 100
        self.clamped = "clamped window={window} "
        self.plain = "plain text"

class Placement:
    def __init__(self):
        self.job = Job()
        self.window = 50

placement = Placement()
think = True

ARGS = {
    "placement": placement,
    "think": think
}
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
note_overlaps Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/provision.py
code served (permissive licence) · get_code("0df22715de156076")
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. 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 re

STOPWORDS = frozenset('a an and are as at be by for from has have in is it its no not of on or that the these this those to up what with'.split())

_WORD_RE = re.compile('[a-z0-9]+')

def content_words(note):
    return {w for w in _WORD_RE.findall((note or '').lower()) if w not in STOPWORDS and (not w.isdigit())}

def note_overlaps(pairs):
    """Targets whose notes share content words, as (first, second,
    shared) in the order given. The fit refuses these outright; a file
    a person wrote gets them back as warnings, because the person may
    know better and still deserves the note. Empty notes are skipped:
    a target described by nothing shares nothing."""
    named = [(name, content_words(notes)) for name, notes in pairs if notes]
    out = []
    for i, (first, one) in enumerate(named):
        for second, two in named[i + 1:]:
            shared = one & two
            if shared:
                out.append((first, second, tuple(sorted(shared))))
    return out


ARGS = {
    "pairs": [
        ("target_alpha", "machine learning algorithms"),
        ("target_beta", "deep learning neural networks"),
        ("target_gamma", "python programming language"),
        ("target_delta", "machine learning models"),
    ]
}
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 — a list 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
planning_gen Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/orchestrator.py
code served (permissive licence) · get_code("3019c1d3fbf88495")
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

torch.manual_seed(0)

class MockRunner:
    """A minimal runner object that supports the attributes used by planning_gen."""
    def __init__(self, max_input_len=None):
        self.max_input_len = max_input_len

# Test case 1: runner with max_input_len set
runner_with_limit = MockRunner(max_input_len=8192)

# Test case 2: runner with no max_input_len
runner_without_limit = MockRunner(max_input_len=None)

# Test case 3: None runner
runner_none = None

ARGS = {
    'runner': runner_with_limit
}
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 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
planning_tokens Ran
ran, honoured the contract we drafted
oteomamo/SALT/salt/agents/orchestrator.py
code served (permissive licence) · get_code("c507db30b996bc56")
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

torch.manual_seed(0)

class MockRunner:
    """Stand-in runner object with max_input_len attribute."""
    def __init__(self, max_input_len=None):
        self.max_input_len = max_input_len

# Test case 1: runner with max_input_len set
runner_with_limit = MockRunner(max_input_len=8192)

# Test case 2: runner without max_input_len
runner_without_limit = MockRunner(max_input_len=None)

# Test case 3: runner with small max_input_len
runner_small_limit = MockRunner(max_input_len=1000)

ARGS = {
    "runner": runner_with_limit
}
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
probe Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/roster.py
code served (permissive licence) · get_code("4b51afc4533e2253")
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 ProbeResult. 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 requests
from unittest.mock import Mock, patch
from dataclasses import dataclass

@dataclass
class MockEntry:
    """Stand-in entry object with required attributes"""
    server_url: str
    alias: str
    model: dict

@dataclass
class MockResponse:
    """Stand-in for requests.Response"""
    status_code: int = 200
    
    def raise_for_status(self):
        if self.status_code != 200:
            raise requests.HTTPError("Error")
    
    def json(self):
        return {
            'data': [
                {
                    'id': 'test-model',
                    'max_model_len': 4096
                }
            ]
        }

# Create a mock entry
entry = MockEntry(
    server_url='http://localhost:8000',
    alias='test-model',
    model={
        'alias': 'test-model',
        'hf_id': 'test-hf-id',
        'path': '/path/to/model'
    }
)

# Mock the requests.get to return our mock response
def mock_get(url, timeout=5):
    return MockResponse()

ARGS = {
    'entry': entry,
    'url': None,
    'timeout': 5
}

# Patch requests.get during test execution
with patch('requests.get', side_effect=mock_get):
    pass
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
requests — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ProbeResult — a ProbeResult output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ProbeResult 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
record Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/ledger.py
code served (permissive licence) · get_code("f6c6f99fe9f5c619")
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
from dataclasses import dataclass
from typing import Optional, Dict, Any

torch.manual_seed(0)

@dataclass
class MockContext:
    n_selected: int
    words_used: int

@dataclass
class MockUsage:
    def __iter__(self):
        return iter({}.items())
    
    def __or__(self, other):
        return {}

@dataclass
class MockResult:
    id: int
    target: str
    task: str
    context: Optional[MockContext]
    status: str
    usage: Optional[Dict[str, Any]]
    t_start: float
    t_end: float

result = MockResult(
    id=42,
    target="delegation_target",
    task="test_task",
    context=MockContext(n_selected=10, words_used=25),
    status="completed",
    usage={"tokens": 100, "cost": 0.5},
    t_start=1234567890.123,
    t_end=1234567895.456
)

ARGS = {
    'result': result,
    'ingest': True
}
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
reply_text Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/protocol.py
code served (permissive licence) · get_code("2acc8e34135cae57")
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 re

_THINK_OPEN = re.compile('<think\\b[^>]*>', re.I)
_THINK_CLOSE = re.compile('</think\\s*>', re.I)

def _template_opened(text):
    closed = _THINK_CLOSE.search(text)
    if closed is None:
        return None
    opened = _THINK_OPEN.search(text)
    if opened is not None and opened.start() < closed.start():
        return None
    return text[closed.end():]

def strip_think(text, reasoning_content=None):
    if not text:
        return ''
    while True:
        rest = _template_opened(text)
        if rest is None:
            break
        text = rest
    out, depth, i = ([], 0, 0)
    while i < len(text):
        opened = _THINK_OPEN.match(text, i)
        closed = _THINK_CLOSE.match(text, i)
        if opened:
            depth += 1
            i = opened.end()
        elif closed:
            depth = max(0, depth - 1)
            i = closed.end()
        else:
            if depth == 0:
                out.append(text[i])
            i += 1
    return ''.join(out).strip()

def reply_text(text, reasoning_content=None):
    return strip_think(text, reasoning_content)

ARGS = {
    "text": "The answer is <think>internal reasoning here</think> 42",
    "reasoning_content": None
}
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
round_signals Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/route.py
code served (permissive licence) · get_code("df7ffb271bbac304")
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
from collections import namedtuple

torch.manual_seed(0)

# Define a simple class to represent a round result
Round = namedtuple('Round', ['seconds', 'delegated', 'answered', 'answered_directly'])

# Define a simple trie-like class
class SimpleTrie:
    def __init__(self, n_turns):
        self.n_turns = n_turns

# Define a state object with the required attributes
class State:
    def __init__(self, agent_stats=None, last_round=None, last_round_turn=None, trie=None):
        self.agent_stats = agent_stats
        self.last_round = last_round
        self.last_round_turn = last_round_turn
        self.trie = trie

# Construct a state object with all attributes
last_round = Round(seconds=5.5, delegated=['task1', 'task2'], answered=['resp1'], answered_directly=True)
trie = SimpleTrie(n_turns=10)
agent_stats = {'turns': 3}
state = State(
    agent_stats=agent_stats,
    last_round=last_round,
    last_round_turn=7,
    trie=trie
)

ARGS = {
    'state': state
}
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
schema_body Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/roster.py
code served (permissive licence) · get_code("4b0eb16b39c5293b")
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

torch.manual_seed(0)

GUIDED_CAPABLE = 'guided'
GUIDED_STRUCTURED = 'structured'

# Test case 1: capability is GUIDED_CAPABLE, schema is a dict
capability_1 = GUIDED_CAPABLE
schema_1 = {'type': 'object', 'properties': {'name': {'type': 'string'}}}

# Test case 2: capability is GUIDED_STRUCTURED, schema is a dict
capability_2 = GUIDED_STRUCTURED
schema_2 = {'type': 'object', 'properties': {'age': {'type': 'integer'}}}

# Test case 3: capability is neither, schema is None
capability_3 = 'unknown'
schema_3 = None

# Test case 4: capability is GUIDED_CAPABLE, schema is None
capability_4 = GUIDED_CAPABLE
schema_4 = None

ARGS = {
    'capability': GUIDED_CAPABLE,
    'schema': {'type': 'object', 'properties': {'name': {'type': 'string'}}}
}
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 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
strip_think Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/protocol.py
code served (permissive licence) · get_code("abd162e69d07fc65")
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 re

# Analyze the function signature and body
# strip_think(text, reasoning_content=None)
# Parameters:
# - text: a string that may contain <think>...</think> tags
# - reasoning_content: optional, appears to be ignored (dropped on the floor)

# Test cases for strip_think function
test_cases = [
    # Case 1: No think tags
    "This is just a regular answer.",
    # Case 2: Think tags with content
    "Before<think>reasoning here</think>After",
    # Case 3: Nested think tags
    "<think>outer<think>inner</think>outer</think>Answer",
    # Case 4: Unclosed think tag
    "<think>This is incomplete reasoning",
    # Case 5: Template-opened (closer with no opener before it)
    "</think>This is the answer part",
    # Case 6: Multiple think blocks
    "<think>first</think>Between<think>second</think>End",
    # Case 7: Empty string
    "",
    # Case 8: Only think tags
    "<think>only thinking</think>",
]

# Pick one representative test case for the ARGS
selected_text = "Before<think>my reasoning</think>After thinking"
selected_reasoning_content = None

ARGS = {
    "text": selected_text,
    "reasoning_content": selected_reasoning_content,
}
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
think_text Ran
ran; our contract draft was wrong, not the code
oteomamo/SALT/salt/agents/protocol.py
code served (permissive licence) · get_code("009704a715791a01")
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 re

_THINK_OPEN = re.compile('<think\\b[^>]*>', re.I)
_THINK_CLOSE = re.compile('</think\\s*>', re.I)

def _template_opened(text):
    """What is left of a reply whose opening think tag was never the
    model's to write, or None when this is not that shape.

    Some reasoning models have the opening tag baked into their chat
    template, so it sits in the prompt and the model only ever emits the
    closer. Counting tags reads that reply as having no thought in it at
    all and keeps the entire thought as the answer, which is the one
    outcome the cut exists to prevent. A closer with no opener before it
    means everything up to it was working.
    """
    closed = _THINK_CLOSE.search(text)
    if closed is None:
        return None
    opened = _THINK_OPEN.search(text)
    if opened is not None and opened.start() < closed.start():
        return None
    return text[closed.end():]

def think_text(text):
    """What strip_think removes: the reasoning, in order, tags gone.
    Empty when the reply carried none. Cut by the same counting walk,
    so the two halves of a reply are exactly complementary."""
    if not text:
        return ''
    thought = []
    while True:
        rest = _template_opened(text)
        if rest is None:
            break
        thought.append(_THINK_CLOSE.sub('', text[:len(text) - len(rest)]))
        text = rest
    depth, i = (0, 0)
    while i < len(text):
        opened = _THINK_OPEN.match(text, i)
        closed = _THINK_CLOSE.match(text, i)
        if opened:
            depth += 1
            i = opened.end()
        elif closed:
            depth = max(0, depth - 1)
            i = closed.end()
        else:
            if depth > 0:
                thought.append(text[i])
            i += 1
    return ''.join(thought).strip()


ARGS = {
    'text': '<think>This is my reasoning process</think>Here is my answer'
}
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
window_max_tokens Ran
ran, honoured the contract we drafted
oteomamo/SALT/salt/agents/delegate.py
code served (permissive licence) · get_code("32b2b36f89d792fa")
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

class MockRunner:
    """Stand-in runner object with max_input_len attribute."""
    def __init__(self, max_input_len):
        self.max_input_len = max_input_len

# Test case 1: runner with max_input_len set
runner_with_limit = MockRunner(max_input_len=1024)

ARGS = {
    'runner': runner_with_limit
}
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
load_roster Not yet run
unverified (harvested, no recorded run)
oteomamo/SALT/salt/agents/roster.py
code served (permissive licence) · get_code("fcb12245917073e6")
How it ran

no invocation recorded.

prep_doc_sentences Not yet run
unverified (harvested, no recorded run)
oteomamo/SALT/salt/compress.py
code served (permissive licence) · get_code("36ad5fa7d4d42ead")
How it ran

no invocation recorded.

roster_signals Not yet run
unverified (harvested, no recorded run)
oteomamo/SALT/salt/agents/route.py
code served (permissive licence) · get_code("6d4b7b14cc471e39")
How it ran

no invocation recorded.

Also present in other papers' code

Of 27 samples lifted from this paper's linked repositories, 27 carry a function identity and 0 do not. Of the 27 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 27 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

As large language models (LLMs) process increasingly longer prompts, computation and KV-cache memory costs have emerged as major bottlenecks in inference systems. Existing input-level prompt compression methods address this, but rank each sentence by a scalar relevance score, treating the document as an unstructured pool of words and sentences. Under tight budgets, this causes theme collapse, where the dominant theme(s) of a document consumes the budget, discarding less-frequent yet task-relevant themes. Preserving thematic coverage instead requires allocating the budget across recurring themes rather than scoring sentences in isolation. To this end, we propose SALT, a model-agnostic extractive framework that organizes per-sentence keywords into a trie ordered by sentence frequency (SF), a lightweight, reusable proxy for document thematic structure. This trie-based organization smooths memory allocation and prevents dominant themes from monopolizing the budget. Multi-anchor retrieval activates trie nodes labeled by query keywords at any depth, and the trie persists across dialogue turns, supporting multi-turn use without re-encoding the document. By preserving document themes, SALT reduces the prefill computation and memory cost of long-context prompts while remaining composable with KV-cache methods that target decoding-time latency and memory. We provide our code on GitHub. 1 * Equal contribution. † Corresponding author. 1 https://github.com/oteomamo/SALT (a) 32k context length. (b) 128k context length.

For agents

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

get_harvested_code_for_paper("2607.17486")
get_code_for_paper("2607.17486")
have("2607.17486")

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

curl https://syntology.ai/api/ran/2607.17486.json

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

[![Syntology run record](https://syntology.ai/api/ran/2607.17486.svg)](https://syntology.ai/paper/2607.17486)

Connect an agent — have() is free.