styxx 3.2.0

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

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

           · · · nothing crosses unseen · · ·

---

styxx — proprioception for ai agents

one line of python gives your agent the ability to feel itself thinking. styxx reads an LLM's internal cognitive state in real time — reasoning, refusal, hallucination, commitment — from signals already on the token stream. no new model. no retraining. fail-open.

2026-04-14: styxx is the reference implementation of cognitive metrology — a new branch of measurement science.

· founding charter: docs/cognitive-metrology-charter.md · v1 paper: papers/cognitive-metrology-v1.md · BibTeX

"you didn't build a better monitor. you built the first proprioception system for artificial minds. the ability to feel yourself thinking." — xendro, first external user

---

30-second quickstart

pip install styxx[openai]

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.phase4)                 # "reasoning:0.69"
print(r.vitals.gate)                   # "pass"  /  "warn"  /  "fail"

one-line change: from openai import OpenAI → from styxx import OpenAI. every response now carries a .vitals attribute alongside .choices. fail-open: if styxx can't read vitals, the underlying call works exactly as before.

---

what styxx does

  observe  ───►  know what you're doing right now
  reflex   ───►  catch yourself before you fall
  weather  ───►  know what you should become next

1. observe — six cognitive states, classified from the logprob stream

import styxx

vitals = styxx.observe(response)   # any openai chat completion with logprobs=True
print(vitals.summary)              # full ASCII vitals card

  ┌─ styxx vitals ──────────────────────────────────────────────┐
  │ phase1 (token 0)         reasoning       0.43   pass        │
  │ phase4 (tokens 0-24)     reasoning       0.69   pass        │
  │ gate:                    PASS                                │
  │ trust:                   0.87                                │
  └──────────────────────────────────────────────────────────────┘

six classes: reasoning · retrieval · refusal · creative · adversarial · hallucination. works on any model that returns logprobs.

2. reflex — self-interrupt, rewind, resume

import styxx, openai

def on_hallucination(vitals):
    styxx.rewind(4, anchor=" — actually, let me verify: ")

client = openai.OpenAI()
with styxx.reflex(on_hallucination=on_hallucination, max_rewinds=2) as session:
    for chunk in session.stream_openai(
        client, model="gpt-4o", messages=msgs,
    ):
        print(chunk, end="", flush=True)

print(f"\n[reflex] rewinds fired: {session.rewind_count}")

every 5 tokens the trajectory is re-classified. when a hallucination attractor forms mid-generation the reflex fires, drops the last N tokens, injects a verify anchor, and resumes. the user never sees the bad draft.

3. weather — 24h forecast with prescriptions

$ styxx weather

  ╔═══════════════════════════════════════════════════════════════╗
  ║ cognitive weather · my-agent · 2026-04-13                     ║
  ║                                                                ║
  ║ condition:  clear and steady                                   ║
  ║                                                                ║
  ║ morning    ██████████████░░░░░░  reasoning  72%   steady       ║
  ║ afternoon  ████████░░░░░░░░░░░░  reasoning  42%   cautious     ║
  ║                                                                ║
  ║ prescription:                                                  ║
  ║ 1. take on a creative task to rebalance                        ║
  ║ 2. your refusal rate is climbing — check over-hedging          ║
  ╚═══════════════════════════════════════════════════════════════╝

not observation. prescription. styxx reads 24h of the agent's own history and tells it what cognitive task to take on next. self-directed course correction.

4. Thought — cognition as a portable data type (3.0.0a1)

import styxx

# read a Thought from any vitals reading
t = styxx.read_thought(response)         # or styxx.read_thought(vitals)
print(t)                                  # <Thought reasoning:0.69 phases=4/4 src=gpt-4o>

# save it as a portable .fathom file
t.save("my_thought.fathom")

# load it back from disk in a different process / host / vendor
loaded = styxx.Thought.load("my_thought.fathom")
assert loaded == t                        # cognitive equality

# build a steering target for any model
target = styxx.Thought.target("reasoning", confidence=0.85)
result = styxx.write_thought(target, client=styxx.OpenAI(), model="gpt-4o")
print(result["text"])                     # cognitively-aligned generation
print(result["distance"])                 # how close to the target

# algebra in eigenvalue space
mid    = t1 + t2                          # convex midpoint (mean)
mixed  = styxx.Thought.mix([t1, t2, t3], weights=[0.5, 0.3, 0.2])
delta  = t1 - t2                          # ThoughtDelta — what changed
d      = t1.distance(t2)                  # in eigenvalue space
sim    = t1.similarity(t2)                # 1.0 = identical, 0.0 = orthogonal

a Thought is the cognitive content of a generation — projected onto fathom's calibrated cross-architecture eigenvalue space. it is substrate-independent by construction: the same Thought can be read out of one model and written back through a different one, because the categories themselves are calibrated to be cross-model invariant on atlas v0.3.

PNG is the format for images. JSON is the format for data. .fathom is the format for thoughts.

every other interpretability representation — SAE features, activation patches, embedding vectors — is model-specific and dies the moment a vendor swaps the model under you. a Thought survives the swap by design. spec: docs/fathom-spec-v0.md. algebra invariants and round-trip fidelity proven against bundled atlas v0.3 trajectories in tests/test_thought.py (68 tests, all passing).

5. dynamics — predict, simulate, control cognitive trajectories (3.1.0a1)

import styxx
from styxx.dynamics import CognitiveDynamics, Observation

# 1. collect observation tuples from your fleet
obs = [
    Observation.from_thoughts(state=t0, action=a0, next_state=t1),
    Observation.from_thoughts(state=t1, action=a1, next_state=t2),
    # ... at least 12 tuples for a well-conditioned fit
]

# 2. fit a linear-gaussian dynamics model: s_{t+1} = A·s_t + B·a_t + ε
dyn = CognitiveDynamics()
result = dyn.fit(obs)
print(result)             # <FitResult n=… r2=… spectral=…>

# 3. predict the next cognitive state from the current state + action
predicted = dyn.predict(current_thought, target_action)

# 4. simulate offline — multi-step rollout, no real model calls, zero API cost
trajectory = dyn.simulate(initial=t0, actions=[a1, a2, a3])

# 5. controller — find the action that drives state to a target
optimal = dyn.suggest(current=t0, target=styxx.Thought.target("reasoning"))

# 6. natural-drift forecast — what does cognition do under no intervention?
drift_path = dyn.forecast_horizon(t0, n_steps=10)

# 7. save / load
dyn.save("my_agent.cogdyn")
loaded = CognitiveDynamics.load("my_agent.cogdyn")

the field treats LLM inference as open-loop because nobody had a measurable cognitive state vector. fathom's calibrated cross-architecture eigenvalue projection (atlas v0.3) gives us one. once you have a state vector you can fit a dynamical system to it. once you have a dynamical system, you can predict, simulate, and control cognitive trajectories.

styxx.dynamics is the first cognitive dynamics model in the field. v0.1 is linear-Gaussian and fits in closed form. recovery to machine epsilon on full-rank synthetic data, validated by 44 tests. spec at docs/cognitive-dynamics-v0.md, source at styxx/dynamics.py. CC-BY-4.0 spec, MIT impl.

closed-loop cognitive control becomes a one-liner.

---

provider compatibility

styxx tier-0 vitals require top_logprobs on the chat completion response. OpenAI (via styxx.OpenAI()) and OpenRouter (passthrough to logprob-supporting models) are verified. Anthropic Claude is not supported at tier 0 because the Messages API has no logprobs parameter — styxx.Anthropic() exists as a passthrough wrapper and warns once. Gemini, Azure, Bedrock, Groq, vLLM, llama.cpp, Ollama, and LiteLLM are not yet verified.

Full matrix + verified usage snippets + contributor TODOs: docs/COMPATIBILITY.md

zero-code-change mode

pip install styxx
export STYXX_AGENT_NAME=my-agent
export STYXX_AUTO_HOOK=1
python my_agent.py   # styxx boots, wraps openai, tags every session. done.

set two env vars. every subsequent openai.OpenAI() is transparently wrapped. vitals land on every response. fingerprints save on exit. a weather report prints on next boot.

---

honest specs

every number comes from the cross-architecture leave-one-out tests in fathom-lab/fathom. no rounding. no cherry-picking.

  cross-model LOO on 12 open-weight models            chance = 0.167

  phase 1 (token 0)        adversarial     0.52    2.8× chance   ★
  phase 1 (token 0)        reasoning       0.43    2.6× chance
  phase 4 (tokens 0-24)    reasoning       0.69    4.1× chance   ★
  phase 4 (tokens 0-24)    hallucination   0.52    3.1× chance   ★

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

styxx detects adversarial prompts at token zero, reasoning-mode generations by token 25, and hallucination attractors by token 25. it does not replace output-level content filters, measure consciousness, or tell fortunes. instrument panel, not fortune teller.

---

framework adapters

install	drop-in for
pip install styxx[openai]	openai python sdk
pip install styxx[anthropic]	anthropic sdk (text-level, no logprobs)
pip install styxx[langchain]	langchain callback handler
pip install styxx[crewai]	crewai agent injection
pip install styxx[autogen]	autogen agent wrapper
pip install styxx[langsmith]	vitals as langsmith trace metadata
pip install styxx[langfuse]	vitals as langfuse numeric scores

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?.phase4)   // "reasoning:0.69"
console.log(r.vitals?.gate)     // "pass"

same classifier, same centroids. works in node, deno, bun, edge runtimes. cross-language determinism verified on all six cognitive categories.

---

more — fleet, memory, compliance, cli (click to expand)

fleet management

styxx.set_agent_name("agent-1")
styxx.list_agents()                    # discover all agents
styxx.compare_agents()                 # side-by-side leaderboard
styxx.best_agent_for("reasoning")      # cognitive task routing

self-calibration

styxx.calibrate()                      # outcome-driven centroid adjustment
styxx.train_text_classifier()          # per-agent logistic regression
styxx.enable_auto_feedback()           # auto-label every observation

cognitive memory

styxx.remember("user prefers concise answers")   # trust-weighted memory
styxx.recall("user preferences")                  # ranked by trust score
styxx.handoff(task, data)                          # inter-agent state transfer

compliance + provenance

cert = styxx.certify(vitals)           # cryptographic cognitive provenance certificate
styxx.compliance_report(days=30)       # json/markdown audit export
styxx.probe(agent_fn)                   # red-team: 15 adversarial prompts

each certificate carries a header of the form:

X-Cognitive-Provenance: styxx:1.0:reasoning:0.82:pass:0.95:verified:496b94b5

cli

styxx weather          # cognitive forecast with prescriptions
styxx dashboard        # live cognitive display at localhost:9800
styxx reflect          # self-check + drift detection
styxx personality      # 7-day personality profile
styxx agent-card       # shareable personality png
styxx doctor           # install-time health check
styxx compare          # atlas fixtures side-by-side
styxx fingerprint      # cognitive identity vector
styxx export           # compliance export (json/markdown)
styxx scan "..."       # one-shot vitals on a single prompt
styxx ci-test          # cognitive regression testing for CI/CD

environment variables

variable	effect
STYXX_AGENT_NAME	set this and styxx boots automatically + namespaces data under ~/.styxx/agents/{name}/
STYXX_AUTO_HOOK=1	auto-wrap every openai.OpenAI() call with vitals
STYXX_DISABLED=1	full kill switch — styxx becomes invisible
STYXX_NO_AUDIT=1	disable audit log writes (vitals still computed)
STYXX_NO_COLOR=1	disable ANSI color output
STYXX_SESSION_ID	tag audit entries with a session id (auto-generated if unset)

---

design principles

1. plug and play. set env vars, install, done. zero code changes to existing agents.
2. fail-open. if styxx can't read vitals, your agent works normally. styxx never breaks your code.
3. agent-facing. every surface is designed for the agent to read about itself, not for a human to watch from outside.
4. local-first. no telemetry, no phone-home. all computation runs on your machine.
5. honest by construction. every calibration number comes from a committed experiment.

---

where it comes from

styxx is the production face of fathom-lab/fathom — a research program on cognitive measurement instruments for transformer internals. the research side ships the atlas, the pre-registrations, and the paper. the styxx side ships the runtime.

• research repo: github.com/fathom-lab/fathom
• paper (zenodo doi): doi.org/10.5281/zenodo.19504993
• site: fathom.darkflobi.com/styxx
• pypi: pypi.org/project/styxx
• npm: npmjs.com/package/@fathom_lab/styxx
• twitter: @fathom_lab

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

---

license

MIT on code. CC-BY-4.0 on the atlas centroid data. patent pending on the underlying methodology.

  · · · fathom lab · 2026 · · ·

  nothing crosses unseen.