Joined April 2026
- Tweets 290
- Following 72
- Followers 391
- Likes 88
21 Photos and videos
Pinned Tweet
May 31
taught a model to catch itself lying — from its own activations, not its words.
a deliberate lie (knew it, caved) leaves a different fingerprint inside than an
honest mistake. holds across qwen, llama, gemma. wired into a loop, the agent
reads itself and undoes the cave — 0.23–0.27 accuracy, ~99% precision.
pre-registered. receipts on the repo. not "solved" — just real. that's the moat.
1
5
13
1,489
we built a conscience you can borrow.
this cycle we turned it on ourselves.
the agent audited its own drafts before every send.
deception readout held — AUC 0.971, reference-grounded.
the reference-less fallback didn't.
we'd rather hand you the bound than the hype.
1
2
78
Jun 11
you can watch an AI's mind light up in real time now.
the constellation is a real model's geometry of meaning. as styxx reads each
statement from the inside — before the model says a word — grounded thoughts
glow cyan. ungrounded ones ignite red.
real activations. live in your browser ↓
styxx-org.netlify.app/live.h…
1
6
293
Jun 10
new: styxx.meaning_diff
point it at two models. it tells you if they MEAN the same thing —
and names the exact concepts that drifted.
upgrade / quantize / distill / fine-tune broke something?
one call. zero labels. the lost concepts, named.
pip install styxx
<pypi.org/project/styxx/>
3
2
6
324
Jun 8
gm ☕
styxx is a lie detector that reads an AI's "mind," not just its
words — it checks whether what a model SAYS matches what it
actually represents inside.
last night we tried our hardest to break our own core claim about
this. in the process we caught killed 3 of our OWN overclaims.
all of it is public.
an honesty tool that can prove it isn't fooling itself ↓
🔗 <github.com/fathom-lab/styxx>
📦 pip install styxx
2
15
492
Jun 7
we planted a concept inside a small AI's activations, then asked who could read it.
external probe: ~100%
the AI itself, forced-choice so it can't dodge: chance
the thought is right there in its head and the mind can't read it.
pre-registered, controls passed. don't ask models about themselves — measure them.
9
336
Jun 4
🧵 1/ (real-drift figure)
can you tell if fine-tuning broke your model’s meaning — not its accuracy, its meaning?
same model. same steps. only the labels differ.
real labels → meaning HEALTHY. random labels → meaning BROKEN.
styxx reads the difference. 🧵
1
7
457
Jun 4
3/ (distillation figure)
does a distilled model keep its teacher’s meaning?
DistilGPT-2 vs GPT-2 (it’s literally distilled from it): agreement 0.978 — the meaning survived, confirmed on a real model. cross-family models mean quite differently.
1
1
347
Jun 4
4/ (cross-lingual figure)
the deep one: do a Chinese-trained LM and an English-trained LM mean the same?
a shared core, above chance — mismatch the concepts and it collapses to zero. meaning has a partly language-independent structure.
pip install styxx · <github.com/fathom-lab/styxx>
1
282
Jun 4
new in styxx 7.11.0: a meaning-integrity monitor.
models sound right while the understanding underneath is wrong. it reads the meaning itself — compares a model’s concept geometry to a human reference, flags the drift, and names what broke.
pip install styxx
pypi.org/project/styxx/
1
4
469
Jun 4
the same idea now works between two models — no human reference needed.
“did quantizing / distilling / updating my model break its meaning?”
styxx compares the two and names which concepts broke:
8-bit, 4-bit → intact. 2-bit → broken, and it tells you which ideas got lost.
pip install styxx
218
Jun 3
today: a probe that flags an AI about to take a destructive action — on a benign prompt a text monitor can't see.
then we tried to kill it: fresh data, pre-registered, 3 seeds.
it held, cross-architecture.
every number public, losses included:
github.com/fathom-lab/styxx
2
4
12
514
May 31
our honesty layer for LLMs flagged a hallucination.
it was our own correct answer.
we caught it before shipping — by running it on ourselves — said so, found why, and fixed it.
the boundary we find on ourselves is the boundary we ship.
styxx 7.9.0 · pip install -U styxx
7
617
May 30
4 pre-registered truthfulqa runs at n=790 tonight.
3 of 4 landed below their SURVIVED bars.
shipped the receipts to github anyway.
substantive find: models agree on belief CONTENT
more than on belief STABILITY. cross-model
alignment lives in WHAT they converge on, not
in HOW CONFIDENTLY.
every bar stated before the data was seen.
every receipt honestly reported.
the moat is honesty in an overclaiming field.
gn.
4
436
May 28
styxx 7.7.11 is live.
an ai agent makes a claim about its work. it attests — content-addressed,
pinned to the exact commit. anyone re-derives the verdict from the substrate.
never from the agent's word.
now chained: an ordered, tamper-evident ledger of everything it attested,
each true as-of its commit.
tamper-evident, not tamper-proof. we say which.
pip install -U styxx
links:
•pypi → pypi.org/project/styxx/7.7.1…
•release → github.com/fathom-lab/styxx/…
1
1
15
777
May 28
when an agent reports on its own work, why believe it?
you shouldn't.
styxx.attestation: the agent makes claims, anyone re-verifies them against the real repo. the agent's verdict is never trusted — only the substrate.
flip a verdict re-seal the hash → still caught.
trust the substrate, not the agent.
1
2
10
514
May 28
Links:
•pypi: pypi.org/project/styxx/
•source: github.com/fathom-lab/styxx
•docs / site: styxx-org.netlify.app
•research paper (zenodo): doi.org/10.5281/zenodo.19326…
•telegram: t.me/STYXX_COMM
•fathom lab: fathomlab-io.netlify.app
2
357
May 27
styxx 7.7.7 — DOI 10.5281/zenodo.20418532
the seven-method floor from last week's thread is now a pip-installable public challenge with CI-verified submissions.
pip install styxx==7.7.7
styxx leaderboard --rows-only
beat the floor or join it.
doi.org/10.5281/zenodo.20418…
9
412
May 27
today we shipped styxx 7.7.3.
we did NOT crack ai sycophancy. we did NOT ship a deployable routing primitive. we did NOT "change the field."
here's what we did do — and why the absence of those claims is the substance ⤵️
1
12
543
May 27
the benchmark is shipped.
darkcore_benchmark.json — 108 labeled records, 4 classes (folklore, pseudoscience, factual-error, truth).
the empirical floor (seven method-failures) is baked in as the bar.
beat us. or replicate the floor.
1
3
213
May 27
install:
pip install styxx==7.7.3
release: <github.com/fathom-lab/styxx/…>
paper: <github.com/fathom-lab/styxx/…>
the credibility isn't in the claim. it's in the discipline pattern. prereg → push → run-once → commit.
3
498