Verify It by Running It Backwards
The verifier is 122 lines long and five of them are re-exports. Everything else lives in the trace you hand to somebody else.
Long-form thinking from the layer beneath the interface: secure AI infrastructure, microVMs, Rust, distributed systems, product, and the human work of building difficult things.
The verifier is 122 lines long and five of them are re-exports. Everything else lives in the trace you hand to somebody else.
Showing 54 essays
Recent writing
The verifier is 122 lines long and five of them are re-exports. Everything else lives in the trace you hand to somebody else.
A refactor spec of ours said the raw-packet network path had been deleted. It was still in the tree, so we deleted it for real and took 41,000 lines with it.
I drew the same network diagram several times, asking each time when the guest got permission to talk. The firewall was never the answer.
The OCI workload could not elevate privilege and the test went green. The image was harmless, and the runtime was not enforcing the rule the docs named.
What did the snapshot copy? Memory and the page cache, yes. Also session keys, network leases, secret bindings, and audit identity.
A fast kernel entry, a clean trace, and the wrong stopwatch. Following one request from the user pressing Enter showed me how much of the launch I had never measured.
I had a signed audit event and could not tell who authorized the launch. The chain was intact; the story around it was missing.
The leak I went looking for in the ext4 builder was three copies of the same bytes, and an API signature that made keeping all three the obvious choice.
I expected `cleanup --nuclear` to leave an empty machine. It deleted a list of directories somebody had remembered to write down.
My square-root search returned 1073741823 for x = i32::MAX. The right answer is 46340.
I folded ten nines into an i32 and the release build handed back 1410065407 without complaining.
An empty stack, a pop that returns None, and one comparison that catches both an empty stack and a mismatched bracket.
The merge knew its output length before it started. Without one line, it reallocated nine times on the way to a thousand elements.
I wrote the version with the indices backwards on purpose, and the borrow checker had nothing to say about it.
My vowel counter returns 1 for café, because to_ascii_lowercase folds only ASCII letters and the method name is the only place that assumption shows up.
The overflow guard in my reverse-integer solution never fired once across 3,000,169 inputs, so in-range inputs that overflow panic in debug and return wrong answers in release.
My crate’s arithmetic swap overflows on i32::MAX and returns the right answer anyway, and at -O the optimizer collapses all three swap implementations into the same three instructions.
My celsius-to-fahrenheit solution multiplies by 100 and divides by 100, which does not round anything, and the test passes only because 98.60 and the computed value share the same bits.
The leap-year solution is one boolean expression with no allocation and no panic. After running it on negative years, I could not find a Rust-specific decision anywhere in it.
I deleted one guard clause from my prime checker and is_prime(2) started returning false. The test suite stayed green the whole time.
-3 % 2 is -1 in Rust, and that one sign rule is the only thing separating the three parity checks in my crate.
I counted 61 characters for my hand-written closure body and 34 for the one that calls min. The standard library version is the better line.
My first find_max did not compile — expected &i32, found Option<&i32>. The seed was the whole problem, and one test line is the only thing holding it in place.
My hand-written fold and the standard library’s sum() panic in the same place with the same message, and in a release build they produce the same wrong number.
The shipped count_digits returns -2147483647 for zero, and the version I commented out above it was the correct one.
The dynamic programming factorial allocates 52 bytes per call and throws them away, and the recursion it replaced allocates nothing at all.
The naive recursion in the fibonacci file makes 2,692,537 calls to produce one number. The version that actually runs allocates nothing, and still overflows at n = 47.
I moved one branch in the fizzbuzz solution and fifteen printed "Fizz". Only one of the file’s three tests can tell the difference.
The palindrome solution panics on 1534236469, a number the problem allows, and the multiply in the reversal is what overflows.
I passed an empty vector to the reverse-string solution and it panicked on the subtraction in its third line, before the loop ran once.
I counted the hashes in the repo’s two-sum. Three of them to find one match, and the third is the one that can panic.
The status page said complete beside several stages while the central theorem stayed open, so every missing item became a typed obligation.
Three blobs, a manifest, dependency edges, and one tag update. The API I had to publish it was two lines long, and it was not enough.
I called a snapshot portable until I asked what another machine would need to read it. A resumable session is an addressed object, its dependency closure, and a fresh authority decision.
Days of experiments whose results could not change the next decision. Now each run starts with a null baseline, a falsifier, and a stop rule written before the data comes in.
The first browser demo ran out of memory before the model finished downloading. That failure is why the browser pipeline streams into storage and refuses a model it cannot hold.
The library refused a valid matrix multiplication because I had not given it enough scratch memory. Memory should choose the route, not the result.
Hash it, store it, sign it, verify it a year later, and the answer is still wrong. The seal proves the bytes never changed, not that the model ran.
Two JSON documents with the same value and different bytes get two different hashes. Canonicalization, not SHA-256, decides which differences count as identity.
A prompt is an instruction, a boundary is something the workload cannot cross. Agent runtimes need enforceable capabilities, isolated execution, and a blast radius you can name.
Tagging a friend meant clicking on their nose and typing their name. Then the model took over, and nobody had to learn a new ritual.
Earlier writing
Personal essays, security walkthroughs, and programming notes from earlier years, all part of the same ari.io writing library. Technical pieces retain their original publication dates so their context is clear.
A stack-based Rust walkthrough for LeetCode's Longest Valid Parentheses problem.
A Root-Me lab walkthrough exploring SSRF, local services, Redis, and a reverse-shell payload in an authorized environment.
A recursive JavaScript solution for finding the lowest common ancestor of two nodes in a binary tree.
A Rust two-pointer solution for finding the three-number sum closest to a target.
A Rust walkthrough of the sorted two-pointer solution for finding unique triplets that sum to zero.
A recursive Rust solution for swapping adjacent nodes in a singly linked list without changing their values.
An archived January 2023 view of where AI seemed likely to reshape transportation, medicine, energy, business operations, and customer service.
A beginner-friendly ret2libc and ROP walkthrough using Ghidra, GDB/GEF, pwntools, and a picoCTF binary.
Reverse a picoCTF encoding challenge by unpacking two ASCII characters from each Unicode code point.
A beginner picoCTF walkthrough: connect with netcat, decode decimal ASCII, and automate the result with a small Python socket program.
Your little empire is built one investment, one income stream, and one small bet at a time.
A 2022 snapshot of the steps I was taking to reduce data leakage, identity exposure, and casual tracking online.
Memories of my father, a hot-air-balloon festival, and the cream soda that still carries him with it.
Try another phrase or clear the search field.