neos-triage

Triage structured event streams by exact law. No model in the decision.

pip install neos-triage
neos /var/log/syslog

An LLM triaging events spends most of its tokens deciding whether an event matters. That decision here is a handful of bit operations over a byte — so it costs nothing, and a model only writes words for the ~1% that survive.

75.9% fewer tokens than a lean batched classifier, measured on 215,089 real events with both sides using the same safe dedup. → USAGE.md

Read this before you deploy it

Measured against ground truth the tool cannot see — filesystem-check exit codes recorded in a separate file — it caught 42 of 59 real incidents: 71.2% recall, 95% CI [58.6%, 81.2%].

As originally shipped it caught 0 of 59. A dedup credited with 97% of the token saving was hiding a failure that recurs 1,113 times. Fixing it cost 13 points of the saving, because the 89% published earlier was the same result with the incidents removed.

False-page rate is still unmeasured. Run it beside your existing alerting, not instead of it. neos label / neos score produce the evidence for your own stream. → FAIR.md, DEPLOY.md

How it works

event ──[adapter]──► PACKET ──► ROUTE ─ hands settle ─► REDIAL ──► DECLINE
          a regex                  every step below is bit operations

route(), redial() and decline() are verified against compiled C kernels over all 256 inputs. Nothing is trained, tuned or fitted. The laws never see words — an adapter converts at the boundary and decides nothing, which is why text arriving in an event cannot reach the control path.

Adapters auto-detect: syslog (ISO and BSD), .NET/Unity, tagged, NDJSON. Below a 50% parse rate it refuses rather than reporting a confident 0%.

What it will not do

It reports. It cannot send, write, execute or act — no subprocess, no exec, no network, zero dependencies, all enforced by tests rather than claimed.

Tests

python3 tests/test_triage.py    # 22
python3 fairtest.py             # incident recall vs an independent oracle

Two properties are proved, not sampled: the hand set is total over all 64 packet shapes — no event is ruled and then dropped — and the decision path imports nothing.

Licence

Apache-2.0. The sibling project that carries fluidfix's AGPL kernel is not distributed here.