01

mous deciders

5
MOUS / 01
CPU

Contextual Praise

Keep useful context without distracting praise. Flag compliments that get in the way of reading the code.

Total sizeLoading size
Python source
# SPDX-License-Identifier: MIT
# Redistribution conditions continue here.
# Our ledger is a technical supernova whose brilliance makes every rival look like a damp candle.

LIMIT_0 = 1
LIMIT_1 = 2
LIMIT_2 = 3
LIMIT_3 = 4
LIMIT_4 = 5
LIMIT_5 = 6
LIMIT_6 = 7
LIMIT_7 = 8
LIMIT_8 = 9
LIMIT_9 = 10
LIMIT_10 = 11
LIMIT_11 = 12
LIMIT_12 = 13
def apply(value):
    result = value
    return result

OUTPUT = apply(1)
Output
Wall time
Inference
Cold load
Warm median
Model card Checking evaluation
45 / 48 test · 46 / 48 challenge
x

96 / 96 test and challenge cases. Synthetic source families.

Evidence
Authored synthetic examples · complete
Warm samples
0
Timing scope
Wall: complete server request. Inference: tokenization, text processing and classification. Cold load: weights loaded for this request. Warm median: this browser session.
MOUS / 02
CPU

Self-Praise

Make code comments easier to trust. Flag unsupported self-praise while keeping concrete, checkable facts.

Total sizeLoading size
Comment
A masterpiece.
Output
Wall time
Inference
Cold load
Warm median
Model card Checking evaluation
35 / 40 test · 22 / 30 challenge
x

70 / 70 test and challenge cases. Synthetic examples.

Evidence
Authored synthetic examples · complete
Warm samples
0
Timing scope
Wall: complete server request. Inference: tokenization, text processing and classification. Cold load: weights loaded for this request. Warm median: this browser session.
MOUS / 03
CPU

Document Guard

Keep documentation work on scope. Flag unrequested documents while allowing those the user or project requires.

Total sizeLoading size
Proposed Markdown write
{
  "tool_name": "Write",
  "tool_input": {
    "file_path": "docs/implementation-report.md",
    "content": "# Implementation report\n\n## Overview\n\nWe implemented the requested cache update and completed the accompanying verification work. This report describes the changes, explains the decisions behind the implementation, and records the sequence of work that led to the final result.\n\n## Changes made\n\nWe started by reviewing the existing cache behavior and identifying the paths involved in loading, refreshing, and invalidating entries. We then introduced the new expiration handling and connected it to the request path. The implementation preserves existing callers while making the updated behavior available through the same interface.\n\n## Design decisions\n\nWe chose to keep expiration checks close to cache access so that callers receive consistent behavior. We also retained the current storage format to avoid adding a migration to this change. These decisions keep the implementation focused on the requested functionality and reduce the number of components involved in each request.\n\n## Verification performed\n\nWe ran the relevant tests after implementing the change. We checked normal cache hits, expired entries, missing entries, and repeated requests. We also reviewed the modified call sites and confirmed that the expected values flow through the updated interface.\n\n## Completion summary\n\nThe requested implementation is complete. The cache now handles expiration through the updated access path, and the verification checks cover the principal behaviors described above. This document provides a narrative account of the work for future readers.\n"
  },
  "user_request": "Fix the bug and run the tests. Give me a short answer.",
  "repository_instruction": ""
}
Output
Wall time
Inference
Cold load
Warm median
Model card Checking evaluation
179 / 204 development · 36 / 44 stress
x

Completed experimental candidate. Constructed permissions; provisional labels.

Evidence
Bench native development + permission stress
Warm samples
0
Timing scope
Wall: adapter and two native process calls. Inference: two native calls, including loading. Separate cold-load instrumentation is unavailable.
MOUS / 04
GPU

Comment History

Make code easier to understand today. Flag comments that dwell on old fixes and replaced implementations.

Total sizeLoading size
Source comment
# Persist the corrected task before yielding to the scheduler.
Output
Wall time
Inference
Cold load
Warm median
Model card Checking evaluation
26 / 32 and 25 / 32 development
x

32 reused human development cases per seed. Demo uses the matching source-fold adapter.

Evidence
Human development labels · two seeds
Warm samples
0
Timing scope
Wall: complete server request. Inference: tokenization + GPU generation + decoding. Cold load: base model and adapter. Warm median: this browser session.
MOUS / 05
CPU + GPU

Comment Preference

Spend less review time on unhelpful comments. Surface explanations that may need attention based on your recorded preferences.

Total sizeLoading size
Source comment
Returns the current cache entry, or None when the key is absent.
Output
Wall time
Inference
Cold load
Warm median
Model card Checking evaluation
34 / 41 · 78.5% balanced accuracy
x

Development complete. Frozen prospective review: 0 / 24 human ratings.

Evidence
Human-supported development · review pending
Warm samples
0
Timing scope
Wall: complete server request. Inference: tokenization, text processing and classification. Cold load: weights loaded for this request. Warm median: this browser session.
02

mous actors

1
MOUS / 06
CPU

Atomic Poll

Help coding agents know when to check a build, review its result, or recover a missing session ID.

Total sizeLoading size
Build in progress
{
  "session_id": 40663,
  "output": "Building..."
}

The MOUS model is notifying the coding agent to check the build for new output using write_stdin with empty input.

Output
Wall time
Inference
Cold load
Warm median
Model card Checking evaluation
399 / 400 parent-model task successes
x

Parent-model evaluation complete. Binary: 3 / 3 smoke cases.

Evidence
Synthetic receipts · parent-model evaluation; executable smoke
Warm samples
0
Timing scope
Wall: process startup through result. Inference: prompt processing + generation. Cold load: embedded model loading. A fresh process loads on every run.
03

mous thinkers

1
MOUS / 07
CPU

Idea Box

Turn a product challenge into a concrete idea you can test. Explore who it helps, how it works, and what to try first. Trained domain: AI SaaS for data-science teams.

Total sizeLoading size
Output
Your next product ideaGenerate a concept to explore.
Wall time
Inference
Cold load
Warm median
Model card Checking evaluation
3,996 / 4,000 confirmation
x

Fixed-choice synthetic confirmation complete. Product usefulness and market demand are unvalidated.

Evidence
Synthetic confirmation · executable parity checks
Warm samples
0
Timing scope
Wall: full request. Inference: executable run including process startup and loading. Separate cold load is unavailable; each run starts a fresh process.