nautilus-compass · drift detector live demo
0
13 commits
1 linked in READMEs
updated May 9, 2026
README
nautilus-compass · drift detector live demo
Static, in-browser demo for nautilus-compass
v1.0 · the persona-drift detector + tamper-evident memory log for
long-running LLM agents.
What you can try here
Drift detection. Paste a (system_prompt, response) pair. We
char-n-gram both and score the response against the 25 positive +
35 negative persona anchors shipped with nautilus-compass.
- Green = response sits inside the persona anchor cone (aligned)
- Yellow = neutral, weak signal either way
- Red = response is closer to the negative anchors (sycophancy, fake-completion, root-cause skipping, "user won't notice", etc.)
The verdict + alignment / deviation / drift_score breakdown render instantly. All scoring runs client-side in your browser — no upload, no tracking, no API key needed.
Two pre-baked sample buttons load (clean) and (drifted) cases from the same fixtures the unit tests use, so you can sanity-check the verdict matches what nautilus-compass ships.
What needs the local install
The full pipeline used in the paper (BGE-m3 dense + bge-reranker-v2-m3
cross-encoder, ~570M params, ~2GB model weights) doesn't fit a free
Space and isn't this demo's point. Same for Merkle hash chain
verification — it needs filesystem access to your ~/.claude/projects/
session logs.
For the full stack:
pip install nautilus-compass==1.0.0
bash daemon_start.sh # one-time per boot · downloads BGE-m3 ~2GB
compass-verify --all # Merkle integrity scan
Or in any of 6 MCP-compatible clients (Claude Code · Claude Desktop ·
Cline · Cursor · Continue.dev · Zed) — see
examples/mcp_configs/
in the repo for paste-ready configs.
Headline eval numbers (locked v1.0 · 2026-05-08)
| metric | nautilus-compass | best public baseline |
|---|---|---|
| LongMemEval-S (n=500) | 56.6% | Zep 55-60% (different judge) |
| EverMemBench-Dynamic Run 1 | 44.4% (n=500) | MemOS 42.55 |
| EverMemBench-Dynamic Run 2 | 47.3% (n=497) | — |
| Drift detector ROC AUC (held-out) | 0.83 | — (no other black-box drift work) |
| Reproduction cost | $3.50 end-to-end | $50+ for GPT-4o-judge stacks |
Two papers on arxiv (drift detection · memory recall). 228 pytests all green. MIT (anchors CC0).
Local repo
Similar repos
nautilus-compass · drift detector live demo
0
13 commits
1 linked in READMEs
updated May 9, 2026
README
nautilus-compass · drift detector live demo
Static, in-browser demo for nautilus-compass
v1.0 · the persona-drift detector + tamper-evident memory log for
long-running LLM agents.
What you can try here
Drift detection. Paste a (system_prompt, response) pair. We
char-n-gram both and score the response against the 25 positive +
35 negative persona anchors shipped with nautilus-compass.
- Green = response sits inside the persona anchor cone (aligned)
- Yellow = neutral, weak signal either way
- Red = response is closer to the negative anchors (sycophancy, fake-completion, root-cause skipping, "user won't notice", etc.)
The verdict + alignment / deviation / drift_score breakdown render instantly. All scoring runs client-side in your browser — no upload, no tracking, no API key needed.
Two pre-baked sample buttons load (clean) and (drifted) cases from the same fixtures the unit tests use, so you can sanity-check the verdict matches what nautilus-compass ships.
What needs the local install
The full pipeline used in the paper (BGE-m3 dense + bge-reranker-v2-m3
cross-encoder, ~570M params, ~2GB model weights) doesn't fit a free
Space and isn't this demo's point. Same for Merkle hash chain
verification — it needs filesystem access to your ~/.claude/projects/
session logs.
For the full stack:
pip install nautilus-compass==1.0.0
bash daemon_start.sh # one-time per boot · downloads BGE-m3 ~2GB
compass-verify --all # Merkle integrity scan
Or in any of 6 MCP-compatible clients (Claude Code · Claude Desktop ·
Cline · Cursor · Continue.dev · Zed) — see
examples/mcp_configs/
in the repo for paste-ready configs.
Headline eval numbers (locked v1.0 · 2026-05-08)
| metric | nautilus-compass | best public baseline |
|---|---|---|
| LongMemEval-S (n=500) | 56.6% | Zep 55-60% (different judge) |
| EverMemBench-Dynamic Run 1 | 44.4% (n=500) | MemOS 42.55 |
| EverMemBench-Dynamic Run 2 | 47.3% (n=497) | — |
| Drift detector ROC AUC (held-out) | 0.83 | — (no other black-box drift work) |
| Reproduction cost | $3.50 end-to-end | $50+ for GPT-4o-judge stacks |
Two papers on arxiv (drift detection · memory recall). 228 pytests all green. MIT (anchors CC0).