/// WRITING

Writing.

26 pieces · every one declares what kind of claim it is making · RSS

Reported: This happened.
Designed: This is what I designed, and how I would build it.
Argued: This is what I think is true.

The night the doctrine failed

Seven-rule batch protocol. Five rounds of adversarial audit. An independent cross-verifier as the final gate. Every gate returned green. None of them were real. A 75% false-close rate.

#agents#verification#postmortem#mnemix#moll

OCT 20
Reported
Reviewing coding-agent output: attention is the bottleneck
Reviewing coding-agent output is the bottleneck, not the stuck-agent ping. I read one brief from 8:00 to 8:30 AM and say what that window does not claim.
#Agents#Attention#Review
4 MIN
OCT 15
Reported
RSS is a follow button for your website
What RSS does, how to subscribe to a website, and why a feed is separate from an email newsletter or publishing on X. A short visual guide.
#RSS#Independent publishing#Writing
2 MIN
OCT 15
Reported
I priced 30 days of Claude Code usage: $34,829
A measured comparison of Claude Code subscriptions and usage credits: the $34,829 replacement-cost estimate, trade-offs, and how to calculate your own break-even.
#Claude Code#AI costs#Agent workflows
7 MIN
OCT 13
ArguedMeasured on one engine I run, 2026-10-11; design lesson, not a benchmark.
Your memory ranker is starving, not broken
If every score sits at the floor, stop editing the formula. Count uses, not serves, and feed the ranker the one signal it can learn from.
#agent-memory#ranking#feedback-loops#retrieval
8 MIN
OCT 13
Reported
min-width: 0 on the item in a grid's middle track can let centred, non-wrapping links overlap their neighbours.
One merged change in Northsun's records describes links that overlapped their neighbours in a navigation bar, and its pull request description gives min-width: 0 on a grid item in the middle track as the cause. This post builds a pattern with the same CSS ingredients on a stand-in page and measures it. It says nothing about deployment or production.
#northsun#css#layout#testing
16 MIN
OCT 11
Reported
A CI time limit cancelled a job whose steps had all passed.
Northsun's records show a CI job whose time limit sat just above its slowest passing runs. A slower run was cancelled after every step had passed, and the deploy workflow read the run as not green.
#northsun#ci#deployment#nodejs#testing
18 MIN
OCT 07
Reported
A model parser returned the same empty array for four different things.
A step that reads a model's JSON answer returned an empty array when the model found nothing and when it could not read the answer. Northsun's fix kept the leniency and named the outcomes.
#northsun#testing#reliability#nodejs#llm#observability
22 MIN
OCT 07
Reported
Four more ways a check goes green without running.
Northsun's records describe four checks that could report success without doing their job. Each is reproduced here in a few lines, with the fix and how to catch the next one.
#northsun#ci#testing#reliability#shell#nodejs
20 MIN
OCT 07
Reported
A flaky timing test blocked a deploy. The fix bought stability with a bigger blind spot.
Northsun's timing test failed main's push CI on a loaded runner and blocked a deploy. The record reports the replacement stable under heavy synthetic load, and the blind spot it bought is one you can measure.
#northsun#testing#ci#performance#reliability#flaky-tests
21 MIN
OCT 07
Reported
A load test counted its neighbours and never saw its own burst.
Northsun's connection-ownership test went red on two pull requests in one night and passed when run alone. It was counting the neighbours and, run alone, reading none of its own burst.
#northsun#testing#postgres#concurrency#ci#reliability
19 MIN
OCT 06
Reported
The same email gave two different hashes in Node and Python. Nothing raised an error.
A review of a draft Northsun encoding spec showed that one email address can hash two ways depending on the runtime. The comparison was re-run on five runtime builds.
#northsun#unicode#hashing#interoperability#testing#nodejs#python
13 MIN
OCT 04
Reported
Our memory read kept missing its time budget. The last cause we found was one line of config.
Northsun's voice route reads a caller's stored facts under a hard time budget and fails open. The budget kept firing. We suspected the database first, then found a Worker running far from it.
#northsun#cloudflare-workers#postgres#latency#observability#voice-ai#reliability
15 MIN
OCT 03
Reported
What does a 200 from your newsletter signup endpoint actually prove?
I had an AI coding agent audit this site's newsletter signup on a local build, against a stand-in for the provider. The success path returned {ok:true}. Around it: a 312-character address forwarded, a dropped connection answered with an empty 500, and a field the provider doesn't document.
#newsletter#resend#nextjs#testing#email#verification
8 MIN
SEP 29
ArguedArgument and definition. The Northsun paragraph describes a design, not a measurement: no benchmark, no customer, no measured latency.
What is an agent memory layer?
An agent memory layer is the policy between an AI agent and its storage: what gets written, whose it is, when it was true, and what comes back at the next decision. Here is the definition, the six jobs it has to do, who needs one, and who doesn't.
#ai-agents#agent-memory#memory#architecture
7 MIN
SEP 26
Reported
12 of my 15 meta descriptions were too long for search
An audit of this site found meta descriptions running past ~155 characters on 12 of 15 pages and four titles past ~60. All ten post descriptions were written to be read in full, not to survive a cut. Here is the budget, the fix, and the one change the audit got wrong.
#seo#aeo#nextjs#metadata#verification#postmortem
7 MIN
SEP 25
Reported
The analytics call that couldn't fail
Every CTA on this site fired an analytics event from the first commit. None of them went anywhere. No provider was installed, optional chaining swallowed the miss, and a type-checker directive kept the compiler quiet about it.
#analytics#nextjs#typescript#vercel#verification#postmortem
6 MIN
AUG 26
Reported
29 review rounds hardened a CI gate that nothing ran
The bypass that needed no edit at all — and the five-minute audit that finds it in your repo.
#northsun#github-actions#ai-agents#ci
7 MIN
AUG 23
ArguedArgument. Grounded in public vendor docs, not a measured incident.
The number is not the person
Voice agents that persist contact memory on the phone number will replay the last caller on a shared line. Public Retell and ElevenLabs docs describe the pattern. This is a failure class, not a customer story. No wipe, no price, no Northsun usage claim.
#voice#identity#memory#agents
5 MIN
AUG 22
Reported
Your pager is not your customer
A synthetic check that fatals inside its own warn budget will page for days without proving anyone is hurt. Keep signal, diagnosis, mitigation, and recovery as four evidence states.
#sre#incidents#verification#agents
2 MIN
JUL 25
Reported
The dashboard query RLS wouldn't let through
I revoked two columns from the dashboard's own role and broke every logged-in page load for three days. Postgres was right; my select('*') was wrong — and the retry screen I'd shipped earlier turned the crash into something quiet.
#postgres#rls#supabase#security#postmortem#mnemix
4 MIN
JUL 24
Reported
The meter that counted cache hits as cash
The ledger charged no one, the flag was off, and the migration never touched prod. Review still found three bugs — one that would have billed tenants for free cache hits, one that let a tenant rewrite its own billing history, and a table in prod that no migration created.
#metering#billing#postgres#rls#postmortem#mnemix
4 MIN
JUL 23
Reported
The health check that became a retry storm
The Redis retries on the Linear webhook's /health path existed to absorb slowness. Two designs in, they could amplify it: a shared abort fuse that fired inside unrelated writes, an unbounded claim that could stall past Linear's 5-second deadline, and a synthetic 200 from the SDK that read real first deliveries as duplicates. The fix was a factory and one rule — only null means duplicate.
#webhooks#redis#linear#retries#timeouts#postmortem
4 MIN
JUL 02
ArguedArgument. No artifact behind it.
Shipping solo with AI: who owns the architecture when the AI writes the code
Writing the code and owning the system are not the same job. The test that separates them: explain a decision, then change it. If the answer is 'the AI chose it,' you're a passenger.
#ai-native#engineering#agents#career
4 MIN
JUL 02
DesignedDesign. Not shipped.
Cross-video pattern detection: the YouTube signal everyone ignores
The audience tells you, second by second, where you lost them — but the signal is siloed one video at a time, so the same structural mistake repeats invisibly across a whole channel.
#retention-lab#analytics#youtube#ai-products
3 MIN
JUL 02
ArguedArgument. The latency figure is a design target, not a measurement.
The latency objection for voice-AI memory is a dead argument
Voice agents already wait through voice activity detection on every turn. Memory designed to return inside that window need not add a separate pause. The objection isn't a physics problem — it's an architecture problem.
#voice-ai#memory#latency#mnemix
3 MIN

Writing.

The night the doctrine failed

Reviewing coding-agent output: attention is the bottleneck

RSS is a follow button for your website

I priced 30 days of Claude Code usage: $34,829

Your memory ranker is starving, not broken

min-width: 0 on the item in a grid's middle track can let centred, non-wrapping links overlap their neighbours.

A CI time limit cancelled a job whose steps had all passed.

A model parser returned the same empty array for four different things.

Four more ways a check goes green without running.

A flaky timing test blocked a deploy. The fix bought stability with a bigger blind spot.

A load test counted its neighbours and never saw its own burst.

The same email gave two different hashes in Node and Python. Nothing raised an error.

Our memory read kept missing its time budget. The last cause we found was one line of config.

What does a 200 from your newsletter signup endpoint actually prove?

What is an agent memory layer?

12 of my 15 meta descriptions were too long for search

The analytics call that couldn't fail

29 review rounds hardened a CI gate that nothing ran

The number is not the person

Your pager is not your customer

The dashboard query RLS wouldn't let through

The meter that counted cache hits as cash

The health check that became a retry storm

Shipping solo with AI: who owns the architecture when the AI writes the code

Cross-video pattern detection: the YouTube signal everyone ignores

The latency objection for voice-AI memory is a dead argument