DEV Community

Cover image for Stratagems #1: Mark Johnson Walked Into an AI Audit. The Benchmark Had Everything Figured Out — Except the Truth.

Stratagems #1: Mark Johnson Walked Into an AI Audit. The Benchmark Had Everything Figured Out — Except the Truth.

xulingfeng on June 24, 2026

Complete preparation breeds complacency. What is seen every day no longer raises suspicion. The hidden lies within the open — not opposed to it. — ...
Collapse
 
itskondrat profile image
Mykola Kondratiuk

the kafka migration detail is the real lesson. the skill wasn't wrong - it was right for the wrong system. in my setup, i treat any infra change as an automatic invalidation event for dependent agents

Collapse
 
xulingfeng profile image
xulingfeng

Glad you picked up on that detail. You'll probably enjoy the latest two entries in the series too — #2 and #3 carry the same kind of thinking.

Collapse
 
itskondrat profile image
Mykola Kondratiuk

added to the list. will check #2 and #3 - curious how the judgment call framing holds across the series.

Thread Thread
 
xulingfeng profile image
xulingfeng

Looking forward to hearing which one makes you pause longer. Both hit different 😏

Thread Thread
 
itskondrat profile image
Mykola Kondratiuk

will circle back once i get there - the judgment call framing is what i’m most curious to see tested

Collapse
 
technogamerz profile image
𝕋𝕙𝕖 𝕃𝕒𝕫𝕪 𝔾𝕚𝕣𝕝

Really enjoyed this one! :D
What stood out to me was how quickly "the benchmark says it's fine" can turn into a false sense of confidence. A benchmark can tell you what a model did on a test, but not necessarily how it'll behave when reality throws something unexpected at it.

The story was a great reminder that in AI, verification often matters more than impressive scores. Looking forward to seeing how the next stratagem unfolds!❤️

Collapse
 
xulingfeng profile image
xulingfeng

Honestly this comment made my day 🙏 The benchmark trap thing — yeah, that's exactly what I was trying to get at. It's not that benchmarks are useless. It's that once they turn green, the room goes quiet. Nobody says "let's run more tests" after a 97% pass rate. The system goes to prod, and everyone assumes the edge cases don't exist — until they do.
I've been thinking about this a lot while writing the series. There's a pattern I keep seeing across all these stories: the thing that breaks production is never the thing the benchmark measured. It's always the thing nobody thought to measure because the report already looked good.
Quick question — which stratagem do you think Lena shows up in? I have a feeling she's gonna be someone's favorite 😅

Collapse
 
technogamerz profile image
𝕋𝕙𝕖 𝕃𝕒𝕫𝕪 𝔾𝕚𝕣𝕝

I have no energy right now, so I'll tell you later.

Thread Thread
 
xulingfeng profile image
xulingfeng

Take your time. I know this one deserves more than a quick drive-by 😄

Thread Thread
 
technogamerz profile image
𝕋𝕙𝕖 𝕃𝕒𝕫𝕪 𝔾𝕚𝕣𝕝

That's an interesting question. From the way Lena has been hinted at so far, she doesn't feel like a character who is there just to move the plot forward. My guess is that she'll be connected to one of the stratagems that relies more on perception, misdirection, or understanding human behavior rather than pure strategy.

What makes me curious is that the series seems to be exploring the gap between what AI "knows" and what people actually do in real situations. A character like Lena could become the perfect way to expose those blind spots. Sometimes one unexpected person can reveal more about a system than all the benchmarks and audits combined.

Now you've got me looking forward to her introduction even more. I'm interested to see whether she becomes a key player in the strategy itself or the person who challenges everyone's assumptions about it. :D

Thread Thread
 
xulingfeng profile image
xulingfeng

You're reading way too carefully and I love it 😄
Yeah Lena lives in that space between what the system thinks is happening and what's actually going on. That's her thing.
She shows up a few times across the 36. Not a lot — but every time she does, things move.
Let's just say when Lena's around, whoever thinks they're in control... probably isn't.

Collapse
 
jugeni profile image
Mike Czerwinski

„Production data is dirty. This data isn't." is the diagnostic move worth pointing at. Real production noise (7-8 stack layers, GC pauses, truncated variables) is the integrity signature that's expensive to forge. Cleaned data inverts the trust gradient — what looks better is more likely fabricated. That same primitive runs underneath modern benchmark contamination problems: the actor producing the eval set has incentive to keep the surface clean, and a clean surface is the symptom.

Mark's two-week audit is the structural answer to a recursive problem. If the vendor evaluates own benchmark, the verdict is producer-attested. The only signal that works is a different actor with different motive doing the cross-check — which is exactly the move Mark makes. Authority by incentive, not by credential. Torres has stake in 95%; Mark doesn't.

Apex-Lens-Cleaner v1.0.0 appearing in metadata but absent from the public architecture is the absence-as-signal piece. A module that doesn't officially exist is running. The gap between what the architecture claims and what the telemetry shows is where most of these audits actually break a story open.

Collapse
 
xulingfeng profile image
xulingfeng

Appreciate the deep read, Mike. You're the first to call out the production noise as an integrity signature — that's exactly the detail I was hoping someone would catch. The absence-as-signal piece with Apex-Lens in the metadata is just the first layer. There are threads in the Benchmark numbers, the pipeline naming convention, and a particular 3-second silence that don't fully explain themselves until later in the series. You've earned one of the 36 fragments. The rest will surface.

Collapse
 
jugeni profile image
Mike Czerwinski • Edited

Fragment received. The 3-second silence is the one I'll be listening for. Silence in instrumented systems usually means the question was asked of a layer that doesn't log. Apex-Lens being absent showed someone decided what doesn't get seen; a timed silence, what doesn't get heard. Same hand. I'll keep reading.

Thread Thread
 
xulingfeng profile image
xulingfeng

Nail on both. #2's already on deck.

Collapse
 
mnemehq profile image
Theo Valmis

The benchmark having everything figured out except the truth is the sharpest line about AI evaluation I've read this week. It's the core problem with outcome-only metrics: a system can score perfectly and still be wrong in the way that matters, because the benchmark measures what's easy to measure, not what's true. The audit that catches the real failure is the one that checks the reasoning, not just the result. Looking forward to the rest of the series.

Collapse
 
xulingfeng profile image
xulingfeng

Appreciate you catching that. The hardest line to get right in this piece — because it had to be true, not just clever.

Collapse
 
motedb profile image
mote

The RabbitMQ-to-Kafka detail is the kind of failure mode that keeps me up at night. The AI Skill was "correct" in the sense that its retry logic matched the spec it was trained on. The spec just stopped being true. And nobody noticed because the benchmark said 96.8%.

That is the core problem with AI benchmarks in production systems. They measure accuracy against a static snapshot of reality. Production systems do not have that luxury. The environment shifts, the assumptions rot, and the model keeps executing with confidence.

The naming convention catch is a nice detail. I have done similar audits where the fastest path to uncovering copied architecture was not reading the code but looking at naming patterns and directory structures. People can rewrite logic, but they rarely change their naming habits.

What I find most uncomfortable about this story is the benchmark gaming angle. If Pulse AI had run their benchmarks against a realistic test bed instead of a curated one, the numbers would have told a different story. But nobody in the deal chain had incentive to question good numbers.

Have you seen cases where the benchmark itself was technically accurate but the test data was deliberately curated to produce favorable results?

Collapse
 
xulingfeng profile image
xulingfeng

That's the question the story keeps circling and never quite says out loud. With Pulse AI I don't think anyone sat down and said "let's find data that lies." They picked the data that made their product look good — same thing every vendor does. It's not malice at the individual level. It's that the whole chain is set up to reward good numbers and punish whoever slows down to check them. Buyer needs the deal to close. Vendor needs the POC to pass. Nobody in that room gets paid to say "these numbers are real but misleading." The one person who might — the independent evaluator — doesn't show up until after the contract's signed.

The naming patterns thing you flagged is the same problem but scarier. Code you can audit. Habits you can't. And benchmarks are built on habits — what data someone thought to collect, what scenarios they thought to test, what they didn't think to question. That's where the rot lives, long before anyone writes a line of code.

Collapse
 
mudassirworks profile image
Mudassir Khan

"production data is dirty. this data is not." is the tell. real defect samples carry noise (GC pauses, log threads, 7 or 8 stack layers) because production is messy. an eval set that is too clean is a red flag. the cleaning step inverts the trust signal.

when the vendor grades their own benchmark, incentive and verdict are aligned and the verdict means nothing. Mark's value is not his credential, it is his different motive. that is the structural move.

same pattern in RAG: when the team that builds the retrieval index also evaluates retrieval quality, precision scores look great right up until a real user query breaks it.

which stratagem maps cleanest to the eval contamination failure mode in AI audits?

Collapse
 
xulingfeng profile image
xulingfeng

The 36 Stratagems cover pretty much everything — most problems we run into map to one of them. But I'll hold the answer on this one. What I can say is it's in the first ten. Stay tuned 😄

Collapse
 
yusuftmle profile image
Yusuf Temel

"Production data is dirty. This data isn't." -> perfectly said.

People are cooking data just to look good on green benchmarks, but production always tells the truth. Loved the technical details in the story.

Collapse
 
xulingfeng profile image
xulingfeng

Nice catch — that line is the moment the whole thing unravels. Clean data in production is the biggest red flag there is. Glad the technical details landed 👍

Collapse
 
xulingfeng profile image
xulingfeng

Which stratagem do you think maps best to AI reliability? I've got theories for all 36, but #1 felt like the right opener.😄

Collapse
 
xulingfeng profile image
xulingfeng

The 'deceive the heavens' part that stuck with me — the deception isn't hidden. It's right there in plain sight. The benchmark report, the methodology doc, all public. You just don't question it because the numbers look good. Every audit failure I've seen follows that exact pattern.

Collapse
 
xulingfeng profile image
xulingfeng

Deceive the Heavens isn't just some ancient tactic. You've used it — when you're at work pretending to look busy, when your project status chart tells the story the boss wants to hear, when you ask for a day off and leave out the real reason. Half the time you don't even notice. That's what makes this one wild — it's not about being a great liar. You use it every day and never feel like you're lying.😂

Collapse
 
leob profile image
leob

What a thriller, love it!

Collapse
 
xulingfeng profile image
xulingfeng

Glad it hit. The next 35 won't let up either — #2's outline already hurts more than #1's ending. Stay tuned.