Jaggedness and limitations

Every single-pass decision model has a jagged capability frontier: it outperforms slow generative LLMs on narrow, well-scoped judgment calls, and underperforms on tasks that require multi-step scratchpad computation or exact counting. Designing reliable agent harnesses means calling mouse-1 where a single forward pass is strong and keeping deterministic code or reasoning models where it is not.

Where mouse-1 excels

Use case	Why a single forward pass wins	Recommended primitive
Zero-shot tool and intent routing	Dynamic option sets (criteria changes per turn or user), natural-language descriptions, ~40–100 ms latency	choice
Pre-execution action gating	Binary safety/policy check on a proposed command, URL, or diff before execution	noul
Prompt-injection resistance	Untrusted content lives inside state and is never decoded as instructions; mouse-1 only scores your fixed option keys	noul + choice
Ordinal patch and rubric grading	Evaluating whether a diff or test log satisfies an objective along 2–10 concrete levels	score
Speculative multi-question fan-out	Answering 5–50 independent questions over one state in a single pass (state_tokens billed once)	choice + noul + score

Known failure modes (do not use mouse-1 for these)

Because mouse-1 scores options in one non-autoregressive forward pass with zero output tokens (output_tokens = 0), it has no chain-of-thought scratchpad. Avoid delegating these tasks to the model:

1. Exact multi-step arithmetic and financial totals: Do not ask mouse-1 whether invoice.line_items sum to invoice.total. Compute the sum in application code and pass the boolean or discrepancy into state if a semantic judgment is still needed.
2. Exact counting over large arrays or long strings: Do not ask "Are there more than 15 failed tests in test_log?" over raw log text. Parse or count in code (failed_count: 17) and pass the structured count in state.
3. Calendar, timezone, and date math: Do not ask "Is 2026-10-14 more than 14 business days after 2026-09-22?". Compute the delta in code using your runtime's date library.
4. ** Cryptographic, regex, or syntax validation:** Checksums, UUID formats, JSON schema validity, and carrier opt-out keywords (STOP, UNSUBSCRIBE) belong in deterministic validators, never a probabilistic model.
5. States exceeding 65,536 tokens: mouse-1 refuses states above 65,536 tokens with 422 state_too_long ($0 billed) rather than silently truncating. Pre-filter relevant files or fields before calling /v1/decisions.

When a fixed-label supervised classifier (DeBERTa / ModernBERT) is a better fit

A general-purpose decision model is not always the right tool. Consider a fine-tuned encoder (such as DeBERTa-v3 or ModernBERT) instead of mouse-1 when all four of these hold:

• Your label set is fixed and small (for example, a permanent 3-way sentiment or spam tagger that never changes per customer or per turn).
• You already have 10,000+ clean labeled examples for that exact schema and retrain whenever the rubric changes.
• Your input is a short single text (< 512 tokens) rather than a multi-field agent state or 10k-token diff.
• You only ask one question per input rather than fanning out multiple choice, noul, and score questions in a single pass.

Conversely, use mouse-1 when your options or rubrics are dynamic per tenant or turn, when you have 0 to 1,000 labeled examples (using POST /v1/calibrations once you reach 200 examples), or when you evaluate multiple heterogeneous questions against a shared agent state.