Interfaze Promises Zero AI Hallucinations. Really?

The phrase "no more hallucinations" in AI is doing a lot of heavy lifting these days. It's become almost a marketing tic—every new model waves it like a flag. So when a tool called Interfaze shows up claiming 100% deterministic JSON outputs and zero hallucinations, the correct first move is curiosity, not celebration. And the correct second move is to actually stress-test it.

That's what Andras from Better Stack did in a recent breakdown of Interfaze's beta release, running it through web scraping tasks and—in a move I genuinely respect for its chaotic energy—using it to parse recently declassified Pentagon UFO documents. The results were interesting. Not flawless, but interesting.

What Interfaze Actually Is (and Why the Architecture Matters)

Most of the AI models developers interact with daily—GPT-4, Gemini, Claude—are monolithic transformers. Giant, generalist models that, when you hand them an image or a document, essentially throw the entire model at the problem, predicting tokens one at a time. They're remarkably capable, but that architecture has a specific failure mode: it doesn't know what kind of task it's doing. It's all vibes and statistics.

Interfaze takes a different structural bet. Instead of one massive model, it uses a hybrid architecture with task-specific encoders. There's a convolutional neural network (CNN) dedicated to vision and OCR tasks, and a deep neural network (DNN) stack handling audio and speech. These encoders do the heavy lifting on their specific modalities first—identifying text blocks, bounding boxes, coordinates—and then hand structured data to a transformer orchestrator that converts it into human-readable output.

The analogy that clicks for me: it's like the difference between asking a generalist doctor to read an MRI versus sending the scan to a radiologist first. The radiologist speaks the language of the scan natively. The general practitioner is pattern-matching from memory.

Whether that architectural choice actually delivers on the promise of reliability is the real question.

The SOB Benchmark (Yes, That's What It's Called)

Interfaze introduced their own benchmark for this: the Structured Output Benchmark, or SOB. And before you make the obvious joke—they already did, probably.

What makes SOB worth paying attention to is that it measures something most benchmarks ignore. Andras explains it clearly: "Usually we measure if a model can output valid JSON, but SOB measures if the content inside that JSON is actually correct."

That's a meaningful distinction. Syntactically valid JSON that contains wrong information is still wrong. A model that outputs {"result": "false"} when the answer is true has technically followed the format while completely failing the task. SOB tries to close that gap by evaluating content accuracy, not just structural compliance.

According to Interfaze's own testing, their beta model outperforms Gemini 3 Flash and GPT-5.4 Mini on deterministic tasks—things like extracting data from complex charts or doing multilingual transcription. Worth noting: this is the company's own benchmark, measuring their model against specific competitors on specific tasks. That's not nothing, but it's also not a neutral third-party evaluation. Take it as directional evidence, not gospel.

Where It Gets Genuinely Useful

For developers, the pitch is pretty specific: stop babysitting your model outputs. The frustration Andras describes is real and widely shared—"you ask for JSON and nine times out of ten it's fine, but then there's that one time where it decides to add a helpful introductory sentence or just skips the closing bracket entirely. And that inconsistency kills the production pipeline."

Any developer who's built a pipeline around LLM outputs has lived this. The workarounds—output parsers, retry logic, regex validators—exist precisely because you can't fully trust the format to hold. If Interfaze's architecture genuinely bakes structured output compliance into the processing layer rather than bolting it on as a post-hoc instruction, that's a real solve for a real problem.

The web scraping angle is similarly grounded. By treating a webpage as a structured map rather than raw text, the model can supposedly extract clean data without getting tangled in nav bars, footers, and ad boilerplate. The demo in the video shows structured JSON headlines from a live Nvidia chip query—tidy, no fluff.

There's also the guardrails story, which is actually more nuanced than it sounds. Safety filters in most models are binary and opaque—either on or off, with no visibility into why something was blocked. Interfaze apparently lets developers dial in sensitivity by use case. That's useful for legitimate edge cases (medical imaging, legal documents, content moderation tools) where over-refusal creates real friction. It also raises real questions about misuse potential, which is worth being honest about—adjustable safety filters are only as trustworthy as the developers using them.

The UFO Test: A Beautiful Mess 🛸

Here's where the video earns its entertainment value while also delivering genuine signal about Interfaze's limitations.

Pentagon declassified UFO documents are, objectively, terrible OCR inputs. White text on black backgrounds, degraded scans, faded ink, and actual handwritten notes from mid-20th-century field agents. Andras fed these into Interfaze and built a quick HTML visualizer (which he open-sourced, props) to display bounding boxes color-coded by confidence score—green above 70%, yellow for lower confidence, red for low confidence.

The results were... a mixed bag, honestly. Some pages that looked unreadable to a human eye got parsed reasonably well. A handwritten FBI note yielded fragments like "thought it was a balloon, but it went in a definite direction" and "gradually ascending following a path, similar to the trajectory of a bullet." Spooky and actually coherent.

But one green-coded box—supposedly high confidence—came back as "flying flapjacks," which is either a classified aviation program or a spectacular hallucination. Andras' laugh at this moment is earned. The model "is really struggling with some of the areas," he acknowledges, while still calling the overall performance impressive given the source material quality.

That caveat matters. "Impressive given terrible inputs" and "reliable for production use" are different bars. The UFO test is a useful stress test precisely because it exposes where the confidence scoring itself can be misleading—a green box doesn't guarantee correct output, it guarantees the model thinks it's confident.

What This Is and What It Isn't

Interfaze is currently in beta. The pricing is genuinely accessible—$1.50 per million tokens with $20 in free credits to start, which is enough runway to put it through real tasks before committing to anything. The dashboard's system prompt builder is a nice touch for developers who want to generate ready-to-paste code snippets without writing boilerplate from scratch.

But "the idea behind this tool" doing the heavy lifting in Andras' final assessment—"it's not going to hallucinate. At least that is the idea behind this tool"—is an honest hedge worth holding onto. The architecture is thoughtful. The benchmark is genuinely measuring something important. The OCR performance on structured, legible documents looks strong. On degraded, handwritten, or ambiguous inputs, it struggles in ways that resemble... other models.

The determinism claim is probably the most defensible part of the pitch. Structured output compliance being architectural rather than instructional is a real design difference. Whether that holds at scale, across diverse production use cases, outside the demo conditions—that's the question that only time and independent testing will answer.

The real test of Interfaze won't be Pentagon UFO documents. It'll be some developer's messy, real-world data pipeline at 2am when everything is on fire. That's when you find out if "100% deterministic" was architecture or marketing.

---

Yuki Okonkwo is Buzzrag's AI & Machine Learning correspondent. She's the friend who actually read the research and can explain it over drinks.