Bosun v1.1 — adds direction (B replaces A, B depends on A) and NLI-style judgment (paraphrase, causal, contradiction
v1.1 keeps the same architecture and inference contract as v1.0 — and adds directional/typed-edge judgment and public-benchmark generality, with no regression on what already worked.
Thank you — you asked for direction, here it is
The response to the launch shaped this release: the questions, the replications, the "can it do X?" The clearest, most-repeated ask was directional judgment — not just "are these two related?" but "does B replace A? does B depend on A?" Real agent memory is full of edges that have a direction, and a symmetric similarity score can't tell which way they point.
That was the immediate ask, and v1.1 delivers it. Mission accomplished on the thing you asked for first — thank you for the feedback and the interest, and keep it coming. We're listening for what's next.
The news: head-to-head with a frontier LLM
On three public benchmarks Bosun was never fine-tuned on — one instruction each, held-out test — we put Bosun-4B v1.1 against gemini-3.1-flash-lite on the exact same items, alongside the lexical/similarity baseline each benchmark is built to defeat:
| benchmark | similarity baseline | Bosun-4B v1.1 | gemini-3.1-flash-lite |
|---|---|---|---|
| PAWS (adversarial paraphrase) | ~chance (0.53 AUROC) | 0.91 | 0.81 |
| e-CARE (causal direction) | 0.60 | 0.85 | 0.86 |
| ANLI (adversarial NLI) | 0.33 | 0.57 | 0.74 |
Bosun-4B beats the frontier LLM on PAWS, ties on e-CARE, and trails on ANLI — as a general judge, no per-task fine-tuning. And on programmable judgment itself (WarrantBench steerability — does the score flip when the rule flips?) it isn't close: 0.945 vs 0.575. FollowIR holds flat vs v1.0.
New in v1.1
- Directional & typed edges — supersession ("B replaces A"), depends-on, supports / contradicts. Bosun now reads the ordered pair, not just symmetric similarity — the capability a knowledge graph needs to stay clean.
- General-benchmark coverage via training-blend slices, not per-task fine-tuning. Still one model, programmed by a sentence.
Models
Same links — main is v1.1; pin the launch model with revision="v1.0".
- 🟦 Bosun-XS (0.6B) → https://huggingface.co/Hanno-Labs/bosun-xs — at 0.6B it still beats gemini on PAWS (0.90)
- 🟩 Bosun-4B → https://huggingface.co/Hanno-Labs/bosun-4b — the strongest judge
GGUF builds (f16 / Q8_0 / Q4_K_M, runs local on CPU / Apple Silicon) are updated to v1.1 too: bosun-xs-GGUF · bosun-4b-GGUF. (Reminder: not --rerank — use the completion + logits path; the GGUF READMEs have a copy-paste example.)
Honest caveats
ANLI is a real gap — fine-tuned-encoder level, above zero-shot LLMs but below the frontier model. PAWS matches, not exceeds, the fine-tuned DeBERTa specialist (~0.95). The clean, consistent claim is "beats the frontier LLM on the same items."
Apache-2.0. The thesis — the missing judgment layer for agent memory — is in the launch post.