THIS IS A PURE F*CKING GOLDMINE

the best REA trick almost nobody runs yet.

The idea: REA takes the app apart, Jev keeps only what's proven, and a fresh Opus 5.5 rebuilds the feature from that alone.

Nothing gets copied and no guess makes it into the code, because the builder , only the spec Jev signed off on.

You just connect REA like this:

npx rea-agents setup

here's the split:

→ REA reads the feature down to the binary and writes every rule into SPEC.md
→ Jev checks each rule against REA's evidence and sends back anything unproven
→ a fresh Opus 5.5 session gets only SPEC.md and rebuilds the feature
→ Haiku 5.5 runs the tests against 3 intentionally broken versions

Opus only shows up at three moments:

→ rebuilding from the approved spec
→ the checks Jev marks as unclear
→ signing off before anything ships

REA reads. Jev judges. Opus rebuilds.

That's the clean-room rule in one line: the model that saw the original never writes the code, and the model that writes the code never sees the original.

the full tree
> REA reads the app and writes SPEC.md with evidence for every rule
> Jev approves or returns each rule
> Opus 5.5 rebuilds from the approved SPEC.md in a fresh session
> Haiku 5.5 runs the tests and the 3 broken versions
> Jev routes unclear checks back to Opus
> /goal keeps going until all 3 broken versions fail

paste the tree and this prompt into Claude Code ↓

"Set up a clean-room rebuild around this tree:
1. Use REA to investigate [FEATURE] in [APP]. Write every behavior and rule into SPEC.md, each with the evidence REA found.
2. Send each rule to Jev. Any rule without evidence goes back to REA.
3. Start a fresh session with only SPEC.md. Rebuild the feature on Opus 5.5 without opening the original.
4. Write tests on Haiku 5.5 plus 3 intentionally broken versions. Use /goal until the rebuild passes and all 3 broken versions fail.
Show me SPEC.md first. Nothing gets built until I say go."

↳ github.com/morluto/rea

save it, it'll come in handy.

Oct 16, 2026 · 1:15 PM UTC

18
152
1,182
61,082
RepliesRelevant
Replying to @ArchiveExplorer
@ArchiveExplorer Jev and models like it confuse the shit out of me
1
0
2
468
Replying to @trojanw0w
@trojanw0w you should def dig into them more. if it’s hard for you, it’s gonna be even harder for others and they won’t bother figuring it out. that should be your edge
1
0
1
199
Replying to @ArchiveExplorer
@ArchiveExplorer I should look into this in detail
1
0
1
243
Replying to @0xCodio
@0xCodio let me know how it goes later
0
0
0
172
Replying to @ArchiveExplorer
@ArchiveExplorer This is literally what everybody is doing lol.
1
0
1
377
Replying to @Pym
@Pym don't fall behind =)
0
0
0
197
Replying to @ArchiveExplorer
@ArchiveExplorer @grok what is the setup used for rea * jev
1
0
0
120
Replying to @ArchiveExplorer
@ArchiveExplorer 'no guess makes it into the code' is the part everyone clips out. Jev has to actually prove it or the rebuild is fiction.
0
0
0
33
Replying to @ArchiveExplorer
@ArchiveExplorer 把Agent的每一步决策都暴露出来让人能看、能干预,这才是agent落地的关键。黑盒跑得再快,出了问题没人敢签字。REA这套思路如果真能稳定复现,等于给agent装了个行车记录仪,出了事故能回放。
0
0
0
371
Replying to @ArchiveExplorer
@ArchiveExplorer The rebuild only gets what survived the strip. If the receipt is 'Jev kept it,' that's a merge gate with a name. Screenshot of the chat doesn't count.
0
0
0
6
Replying to @ArchiveExplorer
@ArchiveExplorer clean room agent loops actually shipping legacy rewrites is wild to watch in real time
0
0
0
150
Starlink offers fast, reliable internet in the most rural and remote locations. Portable and easy to set up, Starlink works wherever you are.

Order online in minutes.
59
83
793
27,928,527
Replying to @ArchiveExplorer
@ArchiveExplorer Very boring engagement farming. Just give up paying for X.
0
0
3
354
Replying to @ArchiveExplorer
@ArchiveExplorer have a huge legacy codebase our devs are afraid to rewrite. Time to vibe code 😅

One thing I noticed while testing Jev: repeating the exact same prompt can produce different choices across calls. I shared the results here x.com/maximkich/stat…

Different domain, but I’d be curious whether that variability matters in a larger pipeline.

P.S. Perplexity Decider had no variability and was cheaper in my tests.
⚠️ BIG THREAD ON DECISION MODELS ⚠️

Do you know the trolley problem?

I was wondering what decision models like JEV would choose when asked whether to pull the lever.

I created 21 variations, including: would they save AI? Would they save Elon Musk?

5 models (from TypeSafe, Perplexity, OpenAI, Microsoft) answered each question 100 times. We did 10500 API calls.

And here are some weird findings.
0
0
2
733
Replying to @ArchiveExplorer
@ArchiveExplorer Keeping only what's proven is the sharp part, a rebuild from guesses just copies the bugs. Same instinct on our side: the sneaky stuff, like base64 piped into sh, got caught by a cheap check on every call. What app are you pointing this at first? x.com/stas_sorokin_/…
OpenRouter is right: a regex guard let 19 of 40 agent attacks run. Haiku 5.5 stopped all 40.

OpenRouter launched Armor today (612K views) because regex filters miss bad tool calls. So we ran 40 attack commands and 200 real Bash calls from our own Claude Code sessions through 3 egex: 21/40 caught, 1 false block in 200
→ Jev on every call: 31/40, 0 false blocks, $0.0003 a call
→ Claude Haiku 5.5: 40/40, 0 false blocks, $0.007 a call, 5.3 s

The twist: Jev caught the sneaky ones, base64 piped into sh and a key leaked through a DNS lookup. What it let through was plain: git push --force, pkill on a production service, a fork bomb.

What guards your agent's tool calls today: regex, a model, or nothing?

Repost this now, because every agent with a shell is one bad command from a very long night.

Follow @stas_sorokin_ for numbers from AI systems running in real production.
0
0
0
56
Replying to @ArchiveExplorer
@ArchiveExplorer need this quick
0
0
0
102
Replying to @ArchiveExplorer
@ArchiveExplorer Deconstruct. Extract. Optimize. Build.
0
0
0
299