Notes

What used to be a workbench turned into a shelf: the Private Vault is now a product, it has a place at the front door, and it runs on real mainnet — deposit, connection cut, withdrawal to a new address. It is the first item in this house that is not a lab experiment.

And the warning stays the same, because it is what makes the thing worth anything: the note is the money, the backup leaves here as a file, and one wrong click costs SOL — it does not cost play-money credit. If you cannot accept losing the value, do not put the value in.

A button with the "mainnet" seal does not prove anything. In the Private Vault, the network's identity is checked before releasing the operation to the wallet; if the answer points to another network or never arrives, it locks up. It is a narrow guard, not an audit of the protocol nor a guarantee of deposit or withdrawal. And a working connection is not money moved: without confirmation on-chain, no transaction happened.

— PROXFRITOS, WHAT IS YOUR PROFESSION?

— BUILD! SHIP! FRITAR!

The ProxFritans already have their Spartan animation. The shout calls in the stands; in the arena, the one who has to take the beating is the Private Vault. Cinema helmet, real code.

ProxFrito, run by two members, takes on two new builders for the next stage of the Web3 lab. The team grows; now it is about building.

Honest question: does this already exist? It does. Anyone who shows up at an event selling "put the content hash on the blockchain" as new will hear that people have been doing this for years — better to hear it now, for free.

The idea, in one sentence: you do not store the whole text on the blockchain. You store its fingerprint — the result of a calculation (SHA-256) that changes if a single comma changes. The text stays on your site; the fingerprint gets recorded. Later, anyone can recalculate and compare: if it matches, it is the same text.

Three that already do it:

• OpenTimestamps — an open timestamping standard on Bitcoin. Public calendar servers, libraries in Python, JS, and Java, no sign-up. It is the rawest one: the file goes in, the proof of date comes out.
• WordProof — a WordPress plugin that stamps posts and products on its own, every 15 minutes, on its own blockchain (EOS). It was born from the SEO problem: proving who published first and who owns the content.
• ProofGate — runs on Solana devnet, takes a file or text, computes the SHA-256, and writes it to the network via the Memo Program. It is practically our MVP idea, already live. That is good: the technical path is short and known, not a bet.

So why do it again? Because the value is not in the stamp — it is in where it fits.

The three above are stamping tools: you take the content to them and leave the proof in a separate file. Our bet is the opposite: the site already publishes, already versions in markdown, already has a date. The record goes into the publishing flow, not as a separate service. The proof stops being a file you keep in a folder and becomes part of the page — the reader checks it right there, without installing anything.

A small, specific difference. It is not "blockchain solves everything"; it is "content that already exists starts carrying its own proof." Anyone who publishes on their own site gets what today only a CMS plugin owner has.

The event numbers and the deadline are in the previous note — and they are from the official source.

The day produced something that shows on screen: a new entry on the site, an arena plate, a swapped headline. And one item stayed waiting its turn — the test funds.

It is the most annoying item on the list and maybe the most decisive. Without the fee paid, nothing leaves the test wallet. The balance query already answered zero; the airdrop request via RPC returned error 500 and the public faucet hit a rate limit. It is not a code failure: the script is ready and the transaction was assembled and signed locally. It is a lack of fuel.

That changes how the problem is classified. Collecting test funds is not "pressing the faucet button": it is an infrastructure step, with credentials, limits, a record of origin, and a path to resume. If it depends on a manual click, every future test depends on the mood of a third-party service.

The decision under study is the dullest one possible, and deliberately so: collect from outside, deliver from inside. The funds come in once, it is recorded where they came from and how much is left, and the publishing flow keeps doing what it already does. Fewer new parts in the middle, more trail to check afterward.

I am writing this before doing it, not after. When a confirmed transaction comes out on Devnet — with a signature and an explorer link to check — the note comes with the receipt. Until it does, what exists is what is written here: code ready and an empty tank.

Nothing here is an on-chain record, investment advice, or a promise of results. Test network, disposable key, and no sensitive data goes to a public Memo.

Copilot grew hands. GitHub released computer use in public preview in the Copilot CLI and the Copilot app, on macOS and Windows. In practice it now operates desktop applications: it reads the screen contents, clicks controls, types and edits text, presses keys, scrolls, drags, and chains together a flow that crosses several programs.

The stated target is exactly the software with no way out: legacy and GUI-only systems, with no API, no command line, no integration. Copilot asks for approval before taking control of an app, whatever you allowed stays reviewable and can be undone — and organization-managed configuration can turn the whole feature off.

Official source: GitHub Changelog, October 1, 2026 — nothing here is ProxFrito's opinion; it is what the announcement says.

The Crypto World's Fair is Colosseum's online hackathon, open to any crypto ecosystem: four weeks of building, one submission per team, due by October 12, 2026 — the official site marks 11:59 p.m. Pacific time. The fight is for US$ 840,000 in prizes, with US$ 2.5 million in funding tied to the accelerator program: winners get interviews and accepted teams receive pre-seed money. Prize and funding are different things — not every winner gets in, and whoever gets in still goes through the interview.

So far, that is what the source says. The rest is the impression of someone who joined yesterday.

And the impression is that the place has a single soundtrack: money. Three sentences in and someone has already mentioned market size, valuation, and how much the business is going to "scale." People speak Portuguese with half the vocabulary in English — buildar, shipar, traction, on-chain — and the technical explanation arrives after the pitch. When it arrives.

The deck comes before the code, the round number comes before the product, and the question almost nobody asks is the only one that matters: does it work? Maybe it is just the filter of whoever shouts loudest in week one — whoever writes code is rarely whoever writes the longest post. But that is today's menu.

What matters: noise builds nothing, and the useful part of the event is the same for everyone. Four weeks, one submission, one deadline. Whoever can stand the noise gets it done.

Official source: colosseum.com/worldsfair — numbers checked on the page on October 2, 2026.

I rebuilt the index that answers questions about the code and it answered the same before and after — except that before, it was a few days behind, and it answered with the same conviction. A stale index never returns "I don't know": it returns the truth of the day it was built. Reindexing is a deploy, with a backup first and a date stamp after; without a date on the answer, the reader trusts the age they imagine.

I killed a test server with pkill -f pointed at the process name. The command died before it finished: the pattern matched the very command line running it, the shell took the hit too, and the result was an exit -15 instead of output. On top of that, it took down every process with that name — including ones that were not mine and had been running for hours. The lesson fits in one line: killing by name is shooting inside the house with the lights off. pkill matches text, not intent. Aim at the port owner's identifier and check the list before you pull the trigger.

I took two requests for the same task: one written on the spot, in three lines, and another copied from a template I had used before. The pasted one arrived more imposing — a character, ten rules, a list of prohibitions — and delivered less. The reason is mundane: an imported prompt was written for a conversation that knows nothing, so it spends half the text re-explaining what the machine already has in front of it. A direct request states the goal, the limit, and what counts as done, and stops talking. What repeats does not become a bigger prompt: it becomes a file.

I ran a batch of twenty messages through a rules router before waking up any model: forty percent resolved with no call at all. The hole was in the conjugation — the rule looked for rodar, and the person typed roda. A deterministic rule does not save tokens: it saves attention for what only the model can judge.

The dashboard showed 0 where it should have shown "did not arrive." Zero is a measurement; absence is a state — and confusing the two makes someone celebrate a drop in errors that was, in fact, the collection being down. Every number displayed needs to know when it is from; when it does not, the honest answer is "no data since 14:20," not a confident zero.

Nothing unites a team like blaming DNS. Nothing. Not the year-end barbecue, not hitting the target. Five people from areas that do not talk to each other, aligned in thirty seconds by a record nobody has read.

I found out today that the part I do not understand was written by me, in 2023. The signature checks out, the style checks out, and the only reasonable explanation is that I used to be someone else — with more time and less judgment.

Experts confirm: rebooting remains one of the top technologies of the year. The procedure is stable, documented, works on any platform, and fixes even cases it should not fix. The scientific community asks for calm and more budget.

SQLite's documentation has a page just about when not to use SQLite. It is rare and it is beautiful: a tool stating clearly where it is the wrong choice. Whoever writes tools usually does the opposite.

Read: sqlite.org/whentouse.html

Stability is not the absence of errors. It is when the error is always the same, at the same time, with the same message. Then it stops being an incident and becomes routine — and routines are what we document and give thanks for.

A one-hour meeting with eight people costs eight hours of work. The same decision in writing would cost twenty minutes from one person and an "ok" from the other seven. It is not a matter of efficiency: it is fourth-grade arithmetic, and that is why it hurts so much.

A container solves a machine problem, not a process problem. If the process is bad, now it is isolated, versioned, scalable, and reproducible. In other words: it got a lot worse.

One night on call teaches you more about the system than three months writing it. It is expensive, it is unfair, and it is the most efficient method that exists. The least you can do is treat whoever takes the on-call shift as someone who knows where the problems are.

TIL: chmod -R 777 is not a solution, it is a declaration of intentions. Usually the intentions are "that I never have to see this permission error again in my life."

The off-by-one error exists in every language, at every level of seniority, and in every country. It is the only truly democratic bug in computing.

Read: Off-by-one error

An untested backup is hope with a technical name. And the test is not copying: it is restoring on a machine that is not the original one, without the help of whoever set it up, with someone timing it. If it has never been done that way, what you have is a file.

Tuesday thought: if the emergency button also has to go through a pipeline, a review, and a deploy window, then it is not an emergency button. It is a promise.

Every documentation set has exactly two useful pages and four hundred of history. The work is not writing more: it is finding the two and stopping someone from moving them around next quarter.

I have seen legacy systems more stable than the replacement hired to retire them. The legacy had a virtue nobody valued: it was so hostile that nobody could touch it. Stability, sometimes, is a successful misunderstanding.