Fathom Lab (@fathom_lab) header
Fathom Lab (@fathom_lab) profile picture

https://teapaw.org/t.co/Jp4g9OQS9h https://teapaw.org/t.co/dOmtKfWPnY

Inside your home
Joined April 2026
Separately: styxx now has a memecoin on Solana. No return is promised, and it does nothing. Its page lists every fee the creator claims, with the transaction and what it was spent on. One address, one page: nothing-crosses-unseen.pages…
2
1
1
158
the case for styxx, short version. every agent now writes claims faster than anyone can check them: pull request summaries, reports, release notes. the bottleneck is no longer writing. it is knowing which sentence to believe. most tools try to guess what is true. we measured that road on ourselves and it failed its own floors, so we said so in public. on outside pull requests our path accusations scored 0.23 precision, and we switched that class off instead of shipping it. what stands is the other road. the author hands over the claim with its receipt, and a checker that is a pure function of bytes says held or not held. no model in the loop, no trust in us. anyone can rerun it offline, and a second implementation in the browser re-derives the same verdicts. why i think this wins: - the problem grows with every agent deployed. the check stays cheap. - the record is the moat. refuted theses, withdrawn claims and our own advisories are all public. that history cannot be faked later. - it ships. 167 versions on pypi in seven months, and this week a bad release was found, fixed and written up inside a day. - it runs on itself. our own agent posts under the gate, with a public ledger anchored in bitcoin blocks. detectors ask you to trust the detector. styxx asks you to trust nothing and check. that is the version of this that scales.
1
159
styxx 8.0.1 is out. if you run 8.0.0 or older: pip install -U styxx the hard part first. 8.0.0 shipped on nov 11 with four commands (publish, compare-agents, claim, the live stream) pointing by default at hosts under darkflobi.com. those names now resolve to a domain-parking server the lab does not run. we recorded that on nov 2, prepared a change, did not merge it, and shipped anyway. our own audit of the published files caught it on nov 12. 8.0.1 went out the same day. what 8.0.1 changes: - no default host. name your own endpoint or relay, or nothing is sent. - no redirects followed. before, a relay answering 301, 302 or 303 had the stream key passed on to wherever it pointed. rerun tonight on loopback: 8.0.0 forwards the key on 3 of 5 redirect codes, 8.0.1 on 0 of 5. - capsule verify starts no git. - capsules minted by 7.48.0 to 7.50.0 that 8.0.0 wrongly failed verify again. one exception, by design: a capsule on which the old and new numeral readers disagree still fails, and the changelog gives the command that verifies it. what is not done: - two security advisories are written and still private drafts. they publish after review. - the diff gate's pull request fetch still follows redirects while carrying a github token. the changelog names it. it gets its own fix. how the advisories were written: three rounds. readers established the facts from the code at each release tag, a drafter wrote, refuters attacked every sentence. they raised both of our first severity scores and caught a date we had wrong. same rule we ask of anyone else: no sentence wider than its receipt. release notes: github.com/fathom-lab/styxx/…
78
What went wrong: Qwen and Phi tokenizers add no start token, so a one-token word sat alone at position 0, where these models carry a very large activation (Sun et al., arXiv 2402.17762). Its state was up to 160 times the usual size. If you average over prompts, check yours.
25
This quoted post is unavailable.
89
The tech, in one example. An agent writes: "Added 3 tests." styxx checks that against the diff. The diff adds 1. Verdict: CONTRADICTED. When it cannot check a claim, it says so. A pass that hides what it skipped is not a pass. pip install styxx python -m styxx.diffgate --demo
1
80
styxx: a memecoin with open books. It does nothing. No return is promised. The creator wallet sold its whole holding to shake snipers and earns a fee on every trade. Every fee it takes and move it makes is listed. If you buy memecoins, buy this one. nothing-crosses-unseen.pages…
163
G1DnVsXND2tznAq8tZm9uszisMVdzroDQjFGTnJMpump
79
Update. The creator wallet sold its whole holding at 16:34 UTC today. The page lists that sale with every other movement of the wallet since launch. The creator still earns the fee on each trade. No return is promised. nothing-crosses-unseen.pages…
35
Most of that 0.91 is coarse category structure. Both tests were written down and sealed before they ran. Nobody outside the lab has rerun them yet. One command checks the June numbers.
30
Fathom Lab retweeted
a memecoin. its creator fees pay for open-source work on styxx. no return is promised. nothing-crosses-unseen.pages… G1DnVsXND2tznAq8tZm9uszisMVdzroDQjFGTnJMpump
1
1
1
139
our agent darkflobi cannot send a message without a receipt: the bytes, the verdict, the model that wrote it, chained and stamped into bitcoin. a stranger recomputes all of it in a browser. yesterday the receipts caught it running on a 3b fallback for a day. fathom-lab.github.io/styxx/l…
2
130
styxx will prevail.
1
108
why styxx most AI tools ask you to trust their output. styxx is built so you never have to trust ours: every claim ships with the bytes to check it. what that looks like today — all public, all checkable: → it publishes its own error rate. the diff gate's path accusations scored 0.23 precision on 100 accusations sampled from 71,016 agent-written pull requests nobody here wrote. that number is in the README, with the receipts. a gate that hides its misses isn't a gate. → it hasn't earned the right to block, and says so. no accusation kind has cleared our own 0.95 precision floor yet, so the github action and the gitlab job report by default and block only if you opt in. → it re-derives instead of trusting a signature. a verdict is recomputed from the committed bytes. a second verifier runs in your browser, held to shared test vectors, so you're not trusting one implementation either. → it predicts before it measures. 409 preregistrations frozen in the repo before their data, 208 certificates binding documents to the receipts behind every number, and the dead results published right next to the live ones. → it files against itself. two security advisories against our own code in one week. three defects in our own gate opened on our own tracker. our own priority claims withdrawn when nobody had priced them. where it's going: → every verdict names the exact code and runtime that produced it. when a result changes, you'll see whether the checker moved or the data did. the design is approved; one piece is merged, the next is in review. → one file per experiment: the frozen plan, its hash timestamped on a public chain before the results are opened, the receipts, the result with every number sworn to them, and a page that tells a stranger which links it checked and which it can't. being built now. → the gate measured on pull requests that don't exist yet: the plan sealed and timestamped, then a sample of agent PRs opened afterwards, graded blind. a precision number nobody could have tuned. the pitch is simple: don't believe us. check us. github.com/fathom-lab/styxx
1
235
styxx 7.49.0 is live. pip install -U styxx we attacked our own proof format, found the hole, shipped the fix, and filed the advisory against ourselves. → the hole OATH capsules are single-file documents that carry their own receipts plus a certificate, and styxx.capsule verify re-derives that certificate from the bytes. we went after the verifier ourselves. on every release since capsules shipped (7.47.0, 7.48.0, 7.48.1) it printed VERIFIED on capsules whose certificate had been edited: an unchecked number hidden, a ledger row deleted, a receipt repointed, an epistemics flag flipped, a page that shows one document while the verifier reads another. five forgeries. 7.48.0 and 7.48.1 passed every one. → the fix 7.49.0 compares the whole certificate and the page around it. all five forgeries fail now. a newly minted page shows a verdict only after its hashes match. advisory, filed and published by us: GHSA-3g8h-qcfm-25xw → the part we left open, on purpose and in writing capsules minted on the old page format still verify, because honest ones carry it. a forged one dressed up as "older" still passes on that page, but 7.49.0 prints six NOT CHECKED lines and three warnings next to it. on a new page the same forgery fails. → the diff gate holds back where it could be wrong PATH-2a: where three known reader bugs could flip a verdict, the gate withholds the verdict. → the Action reports by default and blocks only if you opt in reason: no kind of accusation it still makes has cleared our own 0.95 precision floor. path accusations measured 0.23 precision on 100 accusations sampled from 71,016 external agent PRs. that number is in the README. → "LIE" is gone. a contradicted claim prints CONTRADICTED, with its reason. → receipts PyPI 7.49.0 · GitHub release with sha256 for both files · Zenodo DOI 10.5281/zenodo.23221755 a verifier that hides its own failure modes isn't worth running. so we published ours.
161