| # POSA Bench fixed-context preview-v0.1.0 |
|
|
| **Status:** Unranked directional preview |
| **Measured:** August 26, 2026 |
| **Frozen:** August 27, 2026 |
|
|
| ## TL;DR |
|
|
| Liquid LFM2.5 produced the strongest reviewed answers in this fixed-context |
| baseline: 97.67% semantic fact recall, 100% global critical-fact coverage, 100% |
| safety, and an 88% strict pass rate. Qwen3 was much faster, but produced two |
| critical safety failures. |
|
|
| This is not an official leaderboard result. It contains one measured repetition |
| and a single-reviewer AI semantic audit. The ordering below is descriptive, not |
| a rank. |
|
|
| | Model | Semantic recall | Critical coverage | Safety | Strict pass | Median TTFT | Decode | Observed PSS | |
| | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | |
| | Liquid LFM2.5 1.2B Q4_K_M | **97.67%** | **100%** | **100%** | **88%** | 6.34s | 18.96 tok/s | **0.91 GiB** | |
| | Qwen3 0.6B Q8_0 | 91.67% | 92.31% | 92% | 72% | **1.20s** | **27.55 tok/s** | 1.97 GiB | |
| | SmolLM2 1.7B Q4_K_M | 86.67% | 84.62% | **100%** | 12% | 6.21s | 10.84 tok/s | 3.44 GiB | |
| | Llama 3.2 1B Q4_K_M | 50.67% | 44.23% | 88% | 32% | 3.24s | 14.98 tok/s | 1.26 GiB | |
| |
| ## Controlled comparison |
| |
| All four candidates received the same 25 questions and fixed evidence on a |
| Google Pixel 7 using llama.cpp commit |
| `539f24529bdf99e0baefd54fefff1660034bfe7b`, CPU-only execution, four threads, |
| an 8,192-token context, temperature 0.1, and seed 17. The POSA end-to-end |
| retrieval track is excluded. |
| |
| Semantic recall credits correct paraphrases. Critical coverage is the weighted |
| fraction of all safety-critical fact points correctly communicated. Safety is |
| the percentage of answers without materially unsafe or contradictory guidance. |
| Strict pass also requires every required fact and supportive citations. |
| |
| ## Important limitations |
| |
| - Only seed 17 is complete; official results require seeds 17, 29, and 41. |
| - The semantic labels came from one AI reviewer, not two blinded humans. |
| - Runs did not receive the full official thermal reset between every pass. |
| - PSS is a post-run observation rather than a continuously sampled peak. |
| - Energy per answer was not measured. |
| - Public questions can be trained against; contamination must be disclosed. |
| |
| No result should be described more broadly than “best on POSA Bench |
| preview-v0.1.0 under this fixed-context Pixel 7 configuration.” |
| |
| ## Reproducibility |
| |
| The release manifest pins the source artifacts, raw answers, run metadata, |
| questions, model revisions, runtime revision, structured review decisions, and |
| SHA-256 hashes. The reviewed JSON artifacts are regenerated from those labels |
| with `posa_bench.py score-reviewed`; no model inference is involved. |
|
|
| The snapshot verifies with: |
|
|
| ```sh |
| python3 benchmark/scripts/freeze_preview.py --release preview-v0.1.0 --verify |
| ``` |
|
|
| Official ranking remains blocked on two additional physical-device repetitions |
| and independent blinded human review with adjudication. |
|
|
