Every Model Signed Off. Sweep Found 5 P1s.
What happens when you stop asking one model to review its own work.
The problem nobody talks about
I spent 30+ iterations on an auth module. Claude, GPT, Grok, Gemini, Kimi. Paid tiers, long sessions, hours of back-and-forth. Every model agreed the code was clean.
Then I ran it through Sweep.
Five P1 bugs. An infinite login loop. Stale auth roles surviving a refresh failure. A token acquisition call missing a parameter that would silently break under specific redirect conditions.
Two passes. $0.098. Under a dime.
Why individual models miss things
When you iterate with a model, even across separate conversations, you're inside what I call convergence anchoring. The model sees code it or its peers helped shape. It pattern-matches against the solution structure it already endorsed. It verifies rather than discovers.
Every model has blind spots. The critical insight: they're not the same blind spots.
What Sweep does differently
Sweep sends your code to multiple LLM reviewers in parallel. Each one sees the codebase cold. No prior reviews, no conversation history, no awareness of who else is reviewing.
A lead model triages the combined findings: adopt, dismiss, or flag as duplicate. Iterative passes run until convergence; when new passes stop surfacing new findings.
The architecture is simple. The results aren't.
Real runs, real numbers
Run 1: Next.js frontend (6,200 lines)
Auth module, CSP headers, accessibility, error handling.
| Passes | 2 |
| Adopted findings | 47 |
| P1 bugs | 5 |
| Total cost | $0.098 |
Run 2: Fish shell system (custom CLI)
A codebase that had defeated every frontier model individually. Claude, GPT, Grok, Gemini, Kimi, each tried in isolation. None could fully resolve the issues.
| Passes | 2 |
| Adopted findings | 86 |
| P1 bugs | 24 |
| Total cost | $0.073 |
No single model found all 24 P1s. One caught regex quoting bugs. Another caught command substitution splits. A third found path traversal. The ensemble covered territory none of them covered alone.
Seven cents.
Then reality checked
An external reviewer validated Run 1. Verdict: 0/5 P1s were actual bugs. The tool's biggest weakness was misreading existing code by claiming a fix was missing when it was already there.
47 findings. Five P1s. Zero real.
What broke and what I fixed
I ran a meta-review: six models independently evaluating Sweep's failure modes, none seeing each other's output. 6/6 agreed the P1 precision was trust-destroying. 5/6 independently proposed the same fix: a post-triage verification gate.
The core problem: the triage lead assigned severity based on reviewer claims, not code verification. “Real bug” isn't a property of the finding; it's a property of consequences in production, which the system couldn't observe.
I built the verification gate, restructured triage to require evidence on high-severity findings, and added pre-triage deduplication.
Run 3: Security platform (32,000 lines)
Auth, 2FA, token management, role-based access control.
| Passes | 3 |
| Adopted findings | 69 |
| P0 bugs | 3 |
| P1 precision | ~70% |
| Total cost | $0.10 |
P1 precision: 0% → 70%.
A dime. Ten minutes. Thirty-two thousand lines.
The cost of not using it
I ran the same 32k-line codebase through Claude Code manually. Opus 4.6, agentic mode, 117+ tool calls. Twenty-five minutes. 42% of my daily usage quota.
| Claude Code | Sweep | |
|---|---|---|
| Time | 25 min | ~10 min |
| API cost equivalent | $5–15 | $0.10 |
| Plan quota burned | 42% | 0% |
| Output | 1 issue draft | 69 findings · 3 P0s |
Sweep calls the API directly. No agentic middleware, no tool-call overhead, no quota burn. Models are stateless disposable functions and all memory lives on the filesystem. Persist the state, not the conversation.
What's still broken
70% is not 100%. The triage lead can still anchor on its own output. Convergence bounds cost but doesn't guarantee coverage. A different reviewer mix finds different bugs.
But now I have proof it works. And it costs a dime.
Looking for codebases to test
Three runs isn't enough. I need more real-world codebases to validate and stress-test the pipeline. If you maintain an open source project and want a free multi-model review, reach out. I'll publish the results.