how the
machine works.

a panel is the floor, not the ceiling. amige. now trains its own eye on the marks different generators leave, and uses it to point each scan toward the detectors built to read that kind of content — so the right specialists weigh in, not just whoever's first in line. that hint is a hunch about where to look; it never votes on the verdict.

the fused result is calibrated and capped, held back to uncertain when the panel splits, and the whole thing keeps watch on itself: as new generators appear, amige. relearns and recalibrates rather than drifting out of date. you can watch a single scan move through it, start to finish.

not every detector runs on every scan. amige. starts by narrowing the roster to the ones suited to your file — an image detector for an image, video and deepfake models for a clip, the text classifiers for pasted text — and to what your plan unlocks. the suited set runs cheapest-and-fastest first.

then the panel checks itself. when the first couple of detectors land confidently on the same side, the scan stops there — a clean image doesn't need a tribunal. when they disagree, or all sit in the muddy middle, amige. doesn't force a call: it brings in another detector, looks again, and keeps going up to a ceiling. it never publishes off a single lonely vote.

no secret blend. a custom scan lets you see every detector by name and pick exactly which ones run, and every verdict shows which ones voted. the roster isn't fixed, either — as new detectors prove themselves on independent benchmarks we add them; as others drift, we rotate them out. the method is the constant: independent detectors, cross-checked, with the disagreement left on the table.

each detector returns its own probability that the content is AI. amige. blends them into one number — but not as a flat average. detectors that have earned more trust on representative test sets count for more; the weakest are kept on screen as evidence but barred from swinging a verdict on their own. a loud vote from a detector that's wrong a lot shouldn't drag everyone else with it.

the fused number then passes a calibration step, and that step has one rule: it is only ever allowed to soften an overconfident score, never to sharpen one. an honest “maybe” should not get talked up into a “definitely.” if calibration can't run, amige. uses the raw blend and flags the verdict as running in a reduced mode, instead of pretending nothing is missing.

the verdict lands three ways: likely AI, likely human, or uncertain / AI enhanced. the middle band is wide on purpose. and some signals get special handling — a confident deepfake flag is high-precision and low-recall, so amige. lets it carry the verdict instead of averaging it away when a co-detector whiffed. see it on a Tom Cruise deepfake or a Ronaldo & Messi clip.

the panel doesn't stop at AI or not AI. some classifiers return per-model guesses with their own confidence. when a flagged image comes back tagged as looks like Midjourney v6, that's a best guess from a classifier built to recognize a generator family's fingerprint — the frequency-domain and up-sampling artifacts these detectors lean on. that guess is model attribution: a resemblance, never an identification.

amige. canonicalizes model names across the panel (so midjourney_v6, Midjourney v6, and midjourney all collapse into one row), weights each guess by how sure the detector was, and shows you the top few. a flagged gpt-4o image or an AI-written essay both surface their best-guess model next to the verdict. it's a closed set — a detector can only name models it was built to recognize — and the guess never feeds the AI-vs-human score. it's the receipts, not the ruling.

the percentage next to a verdict is the probability of the verdict amige. landed on, not a raw AI score. on a likely-AI verdict, 84% is amige.'s confidence it's AI; on a likely-human verdict, 84% is its confidence it's human; on uncertain, you see the raw middle-band score so you can tell where it fell.

two things you'll never see: 0% and 100%. amige. caps every displayed number short of certainty, on purpose. a detector that says it is 100% sure is making a claim we won't make — it's the exact overclaim regulators have gone after. and amige. reports detector disagreement on its own line instead of blending it into the headline, so a likely AI where everyone agrees reads differently from one where a single detector is screaming alone.

uncertain isn't amige. shrugging. it's a decision.

most detectors are built to always answer — to round every hard case up or down. amige. carves out a band where it would rather say “I don't know” than guess, tuned so that across everything it does commit to, it stays under its own budget for wrongly flagging real content as AI. when a scan lands in that band, it abstains. and when the novelty radar says the input is too unfamiliar to stand behind, amige. drops the call rather than promise something it can't keep.

an answer it can't back is worth less than an honest “uncertain.” that's the whole posture: we'll tell you when we're not sure.

treat the verdict as a signal, not a sentence. a likely AI at 92% with every classifier agreeing is strong evidence. the same verdict with one detector at 95% and another at 51% is two stories at once — and the per-detector breakdown is the part you should read.

ask amige.'s answer what you'd ask a human reviewer: does the model attribution match what you'd expect? is the confidence agreement, or one classifier screaming alone? did it abstain — and is the content the kind this tech handles well (a clean diffusion image, long English text) or badly (a screenshot, an ESL essay, paraphrased chatbot output)? if you want to see how amige. compares, the breakdown is where to start. image or amige?