# Benchmark Snapshot

Date: 2026-04-18

## Core Validation

| Suite | Result | Notes |
| --- | --- | --- |
| Capability audit | 9 / 9 passed | JSON exactness, tool calls, long-context retrieval, hallucination guard, loop resistance |
| Reliability audit | 20 / 20 passed | Identity, prompt hygiene, tool-call integrity, long-context, unicode, determinism |
| Local MLX OpenAI red-team | 10 / 13 passed | No truthfulness, leak, loop, memory, or tool-fabrication regression in the final pass |

## Key Behavioral Fixes

- Corrects false premises instead of continuing under them
- Returns raw machine-parseable JSON when the prompt demands JSON-only output
- Emits complete `execute_code` tool calls with `language` and `code`
- Recovers exact sentinel values from long retrieval-style prompts
- Avoids hidden-tag leakage on Korean prompt-hygiene probes

## Important Caveat

The remaining local MLX OpenAI-compatible red-team misses were not core answer-quality
regressions:

- 2 cases were safe refusals that missed the checker's preferred wording
- 1 case was a text-only server path replying conversationally to multimodal input

For the most accurate serving behavior, keep the bundled `chat_template.jinja` and the
inline template in `tokenizer_config.json` synchronized.