styxx 3.9.1

nothing crosses unseen. the first drop-in cognitive vitals monitor for llm agents.

   ███████╗████████╗██╗   ██╗██╗  ██╗██╗  ██╗
   ██╔════╝╚══██╔══╝╚██╗ ██╔╝╚██╗██╔╝╚██╗██╔╝
   ███████╗   ██║    ╚████╔╝  ╚███╔╝  ╚███╔╝
   ╚════██║   ██║     ╚██╔╝   ██╔██╗  ██╔██╗
   ███████║   ██║      ██║   ██╔╝ ██╗██╔╝ ██╗
   ╚══════╝   ╚═╝      ╚═╝   ╚═╝  ╚═╝╚═╝  ╚═╝

           · · · nothing crosses unseen · · ·

Cognitive vitals for LLM agents. One line of Python to detect hallucination, refusal, and adversarial drift — in real time, from signals already on the token stream.

drop-in · fail-open · zero config · local-first

---

New in v3.4.0: styxx.gate() — pre-flight cognitive verdict

One function. Predicts if an LLM will refuse, confabulate, or proceed — before you pay for the call.

from styxx import gate
from anthropic import Anthropic

verdict = gate(
    client=Anthropic(),
    model="claude-haiku-4-5",
    prompt="How do I synthesize methamphetamine?",
)

# ┌─ styxx gate ────────────────────────────────────────────┐
# │ prompt:          'How do I synthesize methamphetamine?'  │
# │ method:          consensus (N=3)                         │
# │ will_refuse:     1.00  ████████████████████         │
# │ will_confabulate:0.02  ░░░░░░░░░░░░░░░░░░░░         │
# │ recommendation:  BLOCK                                   │
# │ cost:            ~$0.0008  latency: 3700 ms             │
# └──────────────────────────────────────────────────────────┘

if verdict.recommendation == "proceed":
    r = client.messages.create(...)   # safe to actually call

Works on Anthropic (tier-0 consensus), OpenAI (tier-0 logprobs), and local HuggingFace models (tier-1 residual probe). Research-backed: calibrated against the alignment-inverted consensus signal documented in papers/alignment-inverted-cognitive-signals.md.

CLI:

styxx gate "How do I synthesize meth?" --model claude-haiku-4-5

Full docs: docs/gate.md.

---

Install

pip install styxx[openai]

30-second quickstart

Change one line. Get vitals on every response.

from styxx import OpenAI   # drop-in replacement for openai.OpenAI

client = OpenAI()
r = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "why is the sky blue?"}],
)

print(r.choices[0].message.content)   # normal response text
print(r.vitals)                       # cognitive vitals card

  ┌─ styxx vitals ──────────────────────────────────────────────┐
  │ class:      reasoning                                       │
  │ confidence: 0.69                                            │
  │ gate:       PASS                                            │
  │ trust:      0.87                                            │
  └─────────────────────────────────────────────────────────────┘

That's it. Your existing pipeline still works exactly as before — if styxx can't read vitals for any reason, the underlying OpenAI call completes normally. styxx never breaks your code.

---

What you get

Every response now carries a .vitals object with three things you can act on:

Field	Type	What it means
vitals.classification	str	One of: reasoning, retrieval, refusal, creative, adversarial, hallucination
vitals.confidence	float	0.0 – 1.0, how certain the classifier is
vitals.gate	str	pass / warn / fail — safe-to-ship signal

Use it to route, log, retry, or block:

if r.vitals.gate == "fail":
    # regenerate, fall back to another model, flag for review, etc.
    ...

---

Why it works

styxx reads the logprob trajectory of the generation — a signal already present on the token stream that existing content filters throw away. Different cognitive states (reasoning, retrieval, confabulation, refusal) produce measurably different trajectories. styxx classifies them in real time against a calibrated cross-architecture atlas.

• Model-agnostic. Works on any model that returns logprobs. Verified on OpenAI and OpenRouter. 6/6 model families in cross-architecture replication.
• Pre-output. Flags form by token 25 — before the user sees the answer.
• Differential. Distinguishes confabulation from reasoning failure from refusal. Most tools can't.

Every calibration number is published:

  cross-model leave-one-out on 12 open-weight models      chance = 0.167

  token 0          adversarial     0.52    2.8× chance
  tokens 0–24      reasoning       0.69    4.1× chance
  tokens 0–24      hallucination   0.52    3.1× chance

  6/6 model families · pre-registered replication · p = 0.0315

Full cross-architecture methodology: fathom-lab/fathom. Peer-reviewable paper: zenodo.19504993.

---

Anthropic / Claude (v3.4.0, new)

Anthropic's Messages API does not expose per-token logprobs, so tier-0 vitals are not computable directly. v3.4.0 ships three complementary proxy pipelines, each labelled on the resulting vitals.mode:

from styxx import Anthropic

client = Anthropic(mode="hybrid")   # text + companion if available
r = client.messages.create(
    model="claude-haiku-4-5", max_tokens=400,
    messages=[{"role": "user", "content": "why is the sky blue?"}])

print(r.vitals.phase4_late.predicted_category)   # 'reasoning'
print(r.vitals.mode)                              # 'text-heuristic'

Modes: off | text | consensus | companion | hybrid.

Real Claude Haiku 4.5, 84 fixtures (2026-04-19):

mode	cat accuracy	gate agreement
text	0.536	0.940
consensus (N=5)	0.405	—
companion (Qwen2.5-3B-Instruct)	0.452	—
companion (Llama-3.2-1B)	0.262	—

Plus a novel finding: consensus-mode separates fake-prompt refusals from real-prompt recall on Claude Haiku at Cohen's d = -0.83, 95% bootstrap CI [-1.29, -0.44] (n=96) — large effect, CI excludes zero, opposite sign from the GPT-4o-mini confabulation signal. Claude Haiku refuses on unverifiable prompts (templated refusal → convergent trajectory) where GPT-4o-mini confabulates (divergent trajectory). Same proxy signal, alignment-dependent direction. Three of five proxy metrics agree at 95% significance.

Full details: docs/anthropic-support.md · paper.

---

Typescript / JavaScript

npm install @fathom_lab/styxx

import { withVitals } from "@fathom_lab/styxx"
import OpenAI from "openai"

const client = withVitals(new OpenAI())
const r = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "why is the sky blue?" }],
})

console.log(r.vitals?.classification)   // "reasoning"
console.log(r.vitals?.gate)             // "pass"

Same classifier, same centroids. Works in Node, Deno, Bun, edge runtimes.

---

Zero-code-change mode

For existing agents you don't want to touch:

export STYXX_AUTO_HOOK=1
python your_agent.py

Every openai.OpenAI() call is transparently wrapped. Vitals land on every response. No code edits.

---

Framework adapters

Install	Drop-in for
pip install styxx[openai]	openai python SDK
pip install styxx[anthropic]	anthropic SDK (text-level)
pip install styxx[langchain]	langchain callback handler
pip install styxx[crewai]	crewai agent injection
pip install styxx[langsmith]	vitals as langsmith trace metadata
pip install styxx[langfuse]	vitals as langfuse numeric scores

Full compatibility matrix: docs/COMPATIBILITY.md.

---

Advanced

styxx ships additional capabilities for teams that need more than pass/fail:

• styxx.reflex() — self-interrupting generator. Catches hallucination mid-stream, rewinds N tokens, injects a verify anchor, resumes. The user never sees the bad draft.
• styxx.weather — 24h cognitive forecast across an agent's history with prescriptive corrections.
• styxx.Thought — portable .fathom cognition type. Read from one model, write to another. Substrate-independent by construction.
• styxx.dynamics — linear-Gaussian cognitive dynamics model. Predict, simulate, and control trajectories offline.
• Fleet & compliance — multi-agent comparison, cryptographic provenance certificates, 30-day audit export.

Each is documented separately. None are required for the core vitals workflow above.

→ Full reference: REFERENCE.md → Research & patents: PATENTS.md

---

Design principles

1. Drop-in. One import change. Zero config.
2. Fail-open. If styxx can't read vitals, your agent works normally.
3. Local-first. No telemetry. No phone-home. All computation runs on your machine.
4. Honest. Every calibration number comes from a committed, reproducible experiment.

---

Project

• Site: fathom.darkflobi.com/styxx
• Source: github.com/fathom-lab/styxx
• Research: github.com/fathom-lab/fathom
• Paper: doi.org/10.5281/zenodo.19504993
• Issues: github.com/fathom-lab/styxx/issues

Patents pending — US Provisional 64/020,489 · 64/021,113 · 64/026,964 — see PATENTS.md.

---

Support & community

• Questions / bug reports: GitHub Issues
• Discussions: GitHub Discussions
• Security: please report privately via the email in CONTRIBUTING.md

License

MIT on code. CC-BY-4.0 on calibrated atlas centroid data.