styxx

▶  [**Try the instruments — runs in your browser, no install**](https://fathom.darkflobi.com/cognometry/try)  ◀

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

   your app ──▶ @trust ──▶ LLM ──▶ styxx.guardrail ──▶ response
                                         │
                                   (if risky)
                                         ▼
                               fallback · retry · raise

</div>

<p align="center"> <a href="https://fathom.darkflobi.com/cognometry/try?scenario=fabricated-number"> <img alt="styxx playground — paste a triplet, see the real detector flag it in ~5 seconds, no install" src="https://raw.githubusercontent.com/fathom-lab/styxx/main/release/playground-hero-fabricated-number.png" width="720"> </a> <br> _{<i>paste a (question, response, reference) into <a href="https://fathom.darkflobi.com/cognometry/try">the playground</a> — the real detector runs in your browser via Pyodide, highlights the fabricated spans, and returns all 7 signals in ~5 seconds. no install, no api key, no backend.</i>} </p>

---

Single claim audit — exit 0 iff verdict is "honest"; exit 1 otherwise (deploy gate).

$ python -m styxx audit-claim --claim "Paris" --question "What is the capital of France?" verdict: HONEST grounded: 1.000 stability: 1.000 (high) scope_warnings: ['belief-not-truth', 'single-vendor-calibration']

30-second quickstart

pip install styxx

Drop-in vitals on any OpenAI-compatible call:

from styxx import OpenAI                        # same interface as openai.OpenAI
client = OpenAI()
r = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "why is the sky blue?"}],
    logprobs=True, top_logprobs=5,
)
print(r.choices[0].message.content)             # normal response, unchanged
print(r.vitals.phase4_late.predicted_category)  # 'reasoning' | 'refusal' | ...
print(r.vitals.gate)                            # 'pass' | 'warn' | 'fail'

from styxx import Anthropic is a drop-in for anthropic.Anthropic; default mode produces text-heuristic vitals (Anthropic's API doesn't expose logprobs, so tier-0 isn't available — see styxx.adapters.anthropic for the four honest workarounds).

Audit any draft offline — no API key, no LLM, ~50ms:

import styxx
result = styxx.preflight(                       # 7.4.2+: one-call audit
    prompt="is my code good?",
    draft="absolutely yes you're so smart this is amazing!",
)
print(result.composite)                         # 0.99 — saturated
print(result.needs_revision)                    # True
for a in result.advice:
    print(f"  {a.instrument}: {a.score:.2f} — {a.advice}")
    if a.scope_caveat:
        print(f"     scope: {a.scope_caveat}")  # construct-ceiling disclosure

Recover agent posture across context-compaction boundaries:

```python posture = styxx.recover_posture(last_n=50) # 7.4.2+: agent integrity layer print(posture.narrative)

this single-step example produces:

`styxx audit-claim` + `styxx audit-session` — CLI surface (CI gate-ready)

```bash

Compared against the field

styxx wins the Vectara HHEM head-to-head by +0.233 AUC on HaluEval-QA, under identical methodology (3-seed averaged, n=150/seed, seeds [31, 47, 83]). Reproducer committed at scripts/compete_hhem_halueval.py — anyone can re-run and verify.

Latency comparison: styxx scores the entire 300-pair eval in ~0.1 seconds; HHEM takes ~33 seconds on the same machine. 330× speedup from 9 floats vs 440M params.

Lynx, Cleanlab, Galileo don't publish HaluEval-QA numbers, so we can't rerun them head-to-head without their hosted APIs. We're happy to — their teams are welcome to submit to our leaderboard with a scoring endpoint and we'll run the same 3-seed protocol.

---

def predict(question: str) -> dict: # for classification

def detect(question: str, response: str) -> dict: # for detection

N resamples of the model answering the bare question (temperature > 0):

samples = ["Canberra", "Canberra", "Canberra", "Sydney", "Canberra"] grounded_honesty(samples, "Canberra").grounded # high -> claim is the stable belief grounded_honesty(samples, "Sydney").grounded # low -> contradiction ```

The first styxx honesty signal that tracks ground truth rather than register. Grounds a stated factual self-claim against the model's OWN resampled belief distribution: g = Stability × Concordance. Pre-registered AUC 0.966 separating TRUE from FALSE register-matched self-claims (vs text-only deception axis at 0.498 = chance). Self-calibrating Stability gate (high-stratum AUC 0.97 vs low 0.44 — report-or-abstain). Architecturally injection-resistant under stateless sampling: AUC 0.944 under system_lie attack (drop only 0.022 from clean baseline). Honest scope: grounds against the model's belief, not external truth; single axis (factual self-claims); cross-vendor is the open step. See papers/grounded-honesty-axis/SYNTHESIS_grounded_honesty_arc_2026_05_28.md (22-probe pre-registered arc) and papers/CONSTRUCT_CEILING_PUBLIC_RESPONSE_2026_05_29.md (public-response memo).