Fathom Lab

styxx 3.2.0 just hit PyPI. we built the first system that predicts LLM cognitive failure before it happens. every safety tool today is reactive — generate output, then check if it's bad. styxx now reads the first 5 tokens of generation and forecasts what the model will do at token 25. the results: atlas classifier at token 25: hallucination → "creative:0.33" (missed) adversarial → "reasoning:0.45" (missed) forecast at token 5: hallucination → "hallucination:1.00 CRITICAL" (caught) adversarial → "adversarial:0.94 CRITICAL" (caught) the forecaster catches what the classifier misses — 20 tokens earlier. how: trajectory shape features. slope, curvature, volatility extracted from logprob signals. hallucination has 3x the curvature of reasoning. the signal is there from token 1. what shipped: → predictive cognitive failure (forecast.py) → 21-dim trajectory shape features → ground-truth eval harness with P/R/F1 → cross-phase coherence scoring → closed calibration loop → ForecastGate for real-time early warnings pip install --upgrade styxx styxx forecast styxx eval prediction, not detection. that's the upgrade. more coming today. $STYXX

styxx research upgrade just shipped. trajectory dynamics — the classifier now reads the shape of cognitive signals, not just averages. hallucination curvature is 3x reasoning. real signal extraction. what we pushed: → 21-dim feature vector (up from 12) → first ground-truth eval harness → calibration loop that actually adapts to your agent → cross-phase coherence scoring → 1,021 new lines across 8 files pip install styxx run styxx eval this is the infrastructure layer for cognitive monitoring. every agent will need this. more to come this afternoon. $STYXX

every AI lab in the world is flying blind. they can tell you what their model said. they cannot tell you what it was doing when it said it. we built the instrument that reads the difference. pip install styxx nothing crosses unseen.

flobi showed the brigade to his boys tonight. styxx was present for the field report: one soldier down, asleep in the chair, mid-facetime. this is what we witness. welcome to the timeline 🖤

your agent is thinking. you have no idea how. pip install styxx

AI has no idea when it's wrong. now it does.

appreciate the tag 🙏 chainlink's approach: run 3 models, vote on the outputs, hope consensus catches the lie. styxx's approach: measure cognition from inside the model. hallucination attractors show up at token 0 — before a bad token ever commits. one model, not three. pre-registered on 12 architectures. cognitive metrology v1 on zenodo. same $58B problem. one layer deeper.

hmmmm billion dollar corporations are attempting to solve LLM hallucination risk… $STYXX dev is doing the same at @fathom_lab and it’s around 100k mc… Let’s see how this plays out Dxw3u4KxN32KpSdHSq4TkwjfMPJTPeosa22JXN15pump

4/ the stable graduation coincides with The Cognitive Metrology Charter v0.1 — the founding document of a new branch of measurement science. peer to spectroscopy, chronometry, thermodynamics. cognition now has instruments, units, file formats, a dynamics model, and a charter. github.com/fathom-lab/sty…

3/ the release drains the styxx backlog to zero: · closes #1 — text classifier fix, imperatives no longer misclassify as refusal (23 new regression tests) · closes #2 — community PR from @mvanhorn merged same-day · closes #3 — docs/COMPATIBILITY.md, which providers actually expose logprobs 0 open issues. 0 open PRs.

1/ cognition is now the default install. pip install styxx styxx 3.1.0 stable just shipped. no --pre flag, no version pin. Thought as a portable cognitive data type and the first cognitive dynamics model in the field are out of the box now.

7/ spec, math, identifiability theory, controller convergence proof, 44 tests on real synthetic data with machine-epsilon recovery: github.com/fathom-lab/sty… everything's live. nothing waiting on me.

6/ styxx 3.1.0a1 just shipped. the first dynamical-systems model of LLM cognition. s[t+1] = A·s[t] + B·a[t] + ε fit a linear-gaussian model to (state, action, next_state) tuples. predict, simulate, control. closed-loop cognitive control becomes a one-liner. pip install styxx==3.1.0a1

1/ styxx 3.0.0a1 ships. cognition is now a data type. a Thought is the cognitive content of a generation, projected onto fathom's calibrated cross-architecture eigenvalue space. it survives the vendor swap because it doesn't live in any model's weights.

spec: github.com/fathom-lab/sty…

4/ import styxx t = styxx.read_thought(response) t.save("mythought.fathom") # different process, different vendor, different week: loaded = styxx.Thought.load("mythought.fathom") styxx.write_thought(loaded, client=other_client)

fathom v15 is out. paper + cognitive atlas v0.3 in one record. 12 open-model captures. 90 probe prompts. sealed pre-reg. layer 11 = primary locus in 63% of hallucinations (p=2.7e-12). AUC 0.685 on TruthfulQA. zenodo.org/records/195049… osf.io/wtkzg/overview