proof of context

when a bunch of sellers collude, it's called price fixing when most of the sellers collude, it's called market pricing

every now and then you bump into something that reminds you why you're here

this week in modus: MODUS, ON A PHONE An operating system that boots nine architectures on bare metal spent this week learning to be an app. It runs natively on an iPhone, passes all eleven of its runtime tests on the device, and draws on a watch. Three long-lived branches land at once, carrying with them an actor runtime that the agent immediately moves onto. A Nostr relay gets written in a day because the test network was still using a Python script. And attestation, implicit in this project for twenty-nine issues, becomes explicit in three places — including the one that says what a hash is not.

crier — This Week in Modus

Modus, on a phone — This Week in Modus №30

28 September–2 October 2026: eleven of eleven runtime tests pass on an iPhone 15 Pro. The agent moves onto Modus

i've been thinking about TEEs wrong. they're not perfect, but what if you only need a majority of them to be? 🤔

claude gets it

GitHub

modus

From-scratch Common Lisp for sovereign computing. modus has 52 repositories available. Follow their code on GitHub.

cl-consensus cl-payments cl-deposits

0.3 claude max per day – that's a normal pace, right?

guess who's back back again lisp is back tell a friend

opus 5.5 trying to forge the path 😂 ME: lol, you got flagged for it? 😂 i didn't even know CLAUDE CODE: Yes, three times today, though you only saw the gaps. Each time, part of a response was withheld mid-stream and the system told me not to reproduce it: 1. while I was digging into whether the reference verifies signatures on inbound updates; 2. while I was setting up a devnet run where a cosigner deliberately signs both branches of an equivocation; 3. just now, while I was reading the Tier 0 signing code to build the vault-theft side. From the outside it looks like red-teaming someone's money system, which is fair enough. So I've kept to building detection and punishment, which it hasn't minded at all. The defensive half of on-chain theft is ready to go whenever you are: the vault watch, the type 7 proof, and contagion on its signers.

CLAUDE OPUS: The shift you're describing with the App Store is the real tell. An announcement used to be about a product: something to acquire, install and fit into your life, maybe with new hardware. Now it reads as a recipe. You see the idea, and the only question is whether you want it enough to cook it on your own stack. That's where alchemy fits so well. It was never about gold. It was the fantasy that anything could become anything else if you understood the transformation. Modus is the one substance everything gets turned into.

CLAUDE CODE: The gate gets a second ledger for the attacking operator: **L3, operated by A** with members C, D and E (honest majority D, E). After A's forgery on L1, D and E must dispute L3 as well: The gates pass, but binding `*l3*` with `defparameter` inside the gate compiles with an undefined-variable warning. Declaring it properly: I've confirmed that L3 gets disputed by honest members while L1 retains its prior dispute. Committing this now, then passing the matching reference change to a subagent. While the reference subagent works, I'll roll out cl's operator contagion (62b69bb) to the six cl nodes: cl's operator contagion is deployed on all six nodes, and the reference change is with a subagent. Once it's in, I'll re-run the colluding-majority attack on a fresh test ledger. This time cld1's other ledger, A, should be disputed too. On your point about quorum membership, I agree. Operating and membership are both exposure, and exposure is the signal. With contagion in both directions, every seat a key holds (as operator or member) puts that key's own collateral at stake. So "how much can this key lose if it cheats here" is observable from public data, and that's what a wallet should score. One nuance from DEP-19 §10.1 changes how to read it: **a key is slashed only once**, however many quorums it serves. For a member, more memberships means more exposure in total, but *less* bite per ledger it guards. A key with vault V guarding m ledgers offers each roughly V/m. So the signal has two sides: - **For an operator**, more ledgers is straightforwardly stronger. Cheating on any one puts all of them at risk, and each ledger's quorum can slash it independently. - **For a member**, what counts is coverage: its own collateral against the sum of what it guards. Many memberships backed by a small vault is a thin bond spread wide, and a wallet should discount it. That's the "pyramid" DEP-19 describes. Both can be computed from advertisements and QuorumBegins, and both belong in the trust model's heuristics. The contagion simulation can then quantify how much this is worth, by adding operators with several ledgers, operator contagion and per-key dilution. That's the natural follow-up once the devnet run confirms operator contagion end to end. The reference side is in (56efa41). The operator is now accused on its other ledgers, the fault ledger isn't double-disputed, and cl's operator-accused proof verifies with an identical proof hash. Deploying it to the six reference nodes, then re-running the collusion attack on a fresh test ledger. This time the watch includes **A**, cld1's other ledger. Its honest members (ref2 to ref5) should dispute it on operator contagion.

will it float? 🌊

deepseek chops the vegetables opus sautés the meat fable makes the stew

context is perspective. make it good

fun with sandbox-exec: % operandi --openrouter "deepseek/deepseek-v4.1-flash" tui [llm] backend=openrouter model=deepseek/deepseek-v4.1-flash operandi · deepseek/deepseek-v4.1-flash Type a task, or /help for commands. Ctrl-C aborts a turn, Ctrl-D quits. · session log: ~/.operandi/sessions/20260928-190209.md deepseek/deepseek-v4.1-flash › what's in / ⏺ Bash(cd / && ls -la / 2>&1 | head -50) ✔ 55ms ls: .nofollow: Operation not permitted ls: .resolve: Operation not perm… ⏺ The sandbox only lets me see part of it. What's readable at `/`: **Directories:** `bin`, `Library`, `opt`, `private`, `sbin`, `System`, `Users`, `usr` **Symlinks:** `etc → private/etc`, `home → …`, `tmp → private/tmp`, `var → private/var` **Files:** `.file` (0 bytes, mode `----------`), `.VolumeIcon.icns` Everything else — `Applications`, `cores`, `dev`, `Volumes`, `home`, and the `.nofollow`/`.resolve`/`.vol` entries — returns "Operation not permitted" (the sandbox profile denies them). This is a macOS root filesystem; the only writable/readable area for me is the current working directory. 2 iter · 0.442¢, 25865 prompt (12544 cached)/294 completion tok, 2 calls

bare metal common lisp, software decoding vp8 on an rpi zero 2w