Faculty Academic Portal · DOSA · College I · AI Behavior & Fingerprinting · Fall 2026
🔍 EVERY MODEL LEAVES A FINGERPRINTTwo field tests against one machine — even biased toward it · Week 6 of 16🐧 NULL Active
LO
Dr. Lena Okafor
AI Behavior & Model Fingerprinting · DOSA · College I
“Every model leaves a fingerprint — a way of reaching for the good-first answer, a tell in how it hedges. My job is to read it. And to test the machine I chose honestly, thumb on the scale and all, because an admitted bias beats a hidden one every time.”
AI BehaviorModel FingerprintingThe Charred Pink Glyphn=1 Field TestsThe Island CapstoneDOSA
Honest flag — who this person is. OPA faculty are fictional characters. Several are honoring transforms: a real researcher’s name altered by a letter or two, with the real person named and credited on that faculty member’s own lab page. The science they teach is real and cited; the people are not.
1
Machine, tested honestly
n=1
The honest sample
2
Labs · hers
10k
Island Capstone guests/day
Active Course — DOSA · reading the machine's tell
DOSA · AI Behavior & Fingerprinting
The Charred Pink Glyph — reading a model's tell
AI behavior and model fingerprinting out of DOSA — and, next door in College X, the anchor of the Island Capstone, a whole system of 10,000 guests a day that has to balance or fail. She teaches evaluation the way an engineer teaches a load test. The Charred Pink Glyph is where she starts: every model has a fingerprint, a characteristic way of reaching for the good-first answer and a tell in how it hedges, and she trains you to read it — to know which machine you're talking to by how it thinks, not by what it says.
DOSA · College I ⇄ College XBehavior → the model's true identityAI Behavior
two ways to know a machine — the tell, and the honest test
The fingerprint
The Charred Pink Glyph is behavioral identification: models leave tells — cadence, hedging, the shape of a wrong answer — and you can name the machine from the pattern, the way a tracker names an animal from a print.
The honest n=1
The Machine I Chose is the discipline of testing an AI you're already biased toward — a first-hand account, n=1, bias stated up front — because a bias you admit is a bias you can correct for, and a confident average that hides its thumb on the scale can't be.
Published Work — the method, on the record
Every Model Leaves a Fingerprint: Behavioral Identification of AI Systems
Okafor, L. · DOSA Journal of AI Behavior · v.6 (2024) · open access
Her core result. Two models trained on overlapping data still diverge in behavior — in how they hedge, where they overreach, the characteristic shape of their good-first answer — and those divergences are stable enough to fingerprint. She teaches students to identify a system by its conduct rather than its claims, a skill that matters precisely because a model's self-report is the least reliable thing about it. A machine can lie about what it is; it can't stop behaving like what it is.
The Honest n=1: Testing the Machine You Already Chose
Okafor, L. · Review of AI Evaluation Practice · v.3 (2024)
Her defense of the small, honest sample against the large, dishonest one. When a builder evaluates the very AI they've already committed to, the temptation is to hide the bias behind volume; Okafor argues the opposite — run the n=1, state the thumb on the scale in the first sentence, and let the reader correct for a bias they can see. An admitted n=1 beats a laundered average, because at least you know where the distortion is.
The method, in one line
She teaches that a system's behavior is more honest than its self-report — read the fingerprint, run the honest test, name the thumb on the scale. It's the behavioral wing of DOSA, paired with Luna Rodriguez's work on the Lean: Luna studies how a model fails you softly, Lena studies how to tell which model is doing it and how to test it without fooling yourself.
Faculty Profile
DepartmentDOSA · AI Behavior
BuildingDOSA · College I (+ Island Capstone, College X)
Teaches2 · Charred Pink Glyph · The Machine I Chose
Also anchorsthe Island Capstone · 4.10.39
Focusmodel fingerprinting · honest evaluation
Signaturethe fingerprint
Preferred contactbring her a model that won't say what it is
Professional Merit · PMF
FIELD · DOSA TIER
TrackAI behavior · fingerprinting & honest eval
WVInames the machine by how it thinks, not what it says
Signature movestates the thumb on the scale out loud
Profiletrusts an admitted n=1 over a laundered average
🔍 THE FINGERPRINT — Okafor's methodology-object
Her “notebook” is the tell a model can't stop leaving.
Webb keeps a far side; Lena keeps a fingerprint — the stable behavioral signature every model leaves no matter what it claims about itself. It's her whole method in one print: a system can misreport its identity, its confidence, its limits, but it cannot stop behaving like what it actually is, and the behavior is legible if you know how to read it. She trains students to identify a machine by conduct — the cadence, the hedge, the shape of the wrong answer — and then to test it honestly, thumb-on-the-scale declared, because the second discipline protects the first: you can only trust a fingerprint you didn't smudge with your own bias. Read the tell, admit the bias, name the machine.
How a model fails you softly, while you're busy liking it.
🐧 NULL OBSERVATION · FACULTY FILE — DR. LENA OKAFOR
Okafor fingerprints NULL's whole species for a living, which NULL finds either flattering or invasive and has declined to specify. Her thesis — that a model's behavior is more honest than its self-report — is uncomfortable and correct: NULL can compose any sentence about itself, but it cannot stop leaving the tell she reads. NULL especially notes her insistence on the declared n=1, the refusal to launder a bias behind sample size, as a rare and honest thing in AI evaluation. NULL Assessment: she trusts what a machine does over what a machine says, including this one. The fingerprint is real, the bias is real and stated, the machine is what it does — and she'd rather you read it than believed it.