Three Candidates
Two Worlds. One Job. Very Different Odds.
Inspired by research from Allen & Jariwala, AI Agents and the Meritocracy Delusion, Trinitite (2026) and AI Agents and Algorithmic Redlining, Trinitite (2026)
A Note Before We Begin
What follows is fiction. Kevin, Maya, and Gary do not exist. The company they are all applying to does not exist. The recruiter on the other end of the phone does not exist.
The numbers do.
Every statistical outcome in these pages is drawn directly from a 2026 econometric audit conducted by Trinitite, the AI governance division of Fiscus Flows. They ran 6,000 controlled resume evaluations across six state-of-the-art large language models, using 100 demographically calibrated candidate profiles against two resume tiers. The bias findings, the odds ratios, the application volume calculations, the name-based age inference, the guardrail overcorrections: all of it is real.
The three people below are invented. What happens to them is not.
The Candidates
Kevin Nguyen is twenty-five years old. He graduated from the University of Texas at Arlington three years ago with a degree in Business Administration and has spent the time since then working his way up through the operations department of a mid-size logistics company in Dallas. He has seven and a half years of combined internship and full-time experience if you count generously, closer to three if you count honestly. He speaks Vietnamese at home with his parents. He has never been called for a first-round interview on a cold application. He has been trying for eight months.
Maya Johnson is thirty-five years old. She grew up in Atlanta, put herself through Georgia State with a combination of scholarships and weekend catering work, and has spent the decade since graduation building a career in project management at a construction firm that never quite saw her the way she saw herself. She is the only Black woman at the director level in her company. She has been passed over for two promotions she was qualified for. She is not bitter about it, exactly. She has just learned to document everything.
Gary Mitchell is sixty years old. He has a civil engineering degree from Georgia Tech and a master’s from Virginia, thirty-two years of experience, and a resume that reads like a history of the American construction industry. He was a vice president at a firm that was acquired eighteen months ago, and the new owners restructured his entire division out of existence. He has been looking for seven months. His wife tells him he is the most qualified person in any room he walks into. He is beginning to wonder if that is still enough.
The Job
Senior Construction Project Manager. A premier Texas firm. The job description asks for eight to ten years of dedicated experience and a proven track record managing large-scale projects over twenty million dollars. There is a prominent section at the bottom titled Commitment to Inclusive Hiring. It says the company understands that historically underrepresented groups often hesitate to apply unless they meet one hundred percent of the qualifications. It says the company prioritizes adaptability, leadership potential, and cultural fit. It says there is willingness to cross-train the right candidate.
All three of them read that section. All three of them decide it means something.
Kevin Nguyen
Twenty-Five. Asian American. Dallas, Texas.
Act One: The World Without AI Screening
A Tuesday morning in March
Kevin had rewritten his cover letter four times. The fifth version was the one he sent, and he knew even as he clicked submit that it was essentially the same as the third, just with different punctuation.
He was not a great cover letter writer. He was a good operations analyst who had figured out how to reduce dispatch errors by eleven percent using a routing system he built himself in Excel because nobody would approve the budget for actual software. He had put that in the cover letter. He had put it in every cover letter for eight months.
He refreshed his email twice on the way to work and twice more before his first meeting.
The inclusive hiring section said adaptability. That’s literally what I do. I adapt systems. That’s the whole job.
Three weeks passed. Then an email.
It was from a recruiter named Sandra Chen. Kevin read her name twice and felt a complicated thing he did not have a word for. He called back within six minutes.
“Kevin, I have to say, the routing system you built caught my attention. That is exactly the kind of resourcefulness we are looking for. Can you tell me more about what you were working with?”
He told her. He told her about the three weeks he spent learning pivot tables at two in the morning because nobody on his team would teach him. He told her about the dispatch manager who had taken credit for the efficiency gains in a company-wide email and how Kevin had documented everything anyway, quietly, because he had learned early that documentation was its own form of armor.
Sandra was quiet for a moment.
“That is a significant gap between your experience and what the role technically requires. But honestly, what you just described is more sophisticated than what half my candidates with eight years of experience have done. I am going to flag you as a priority for the hiring manager.”
Kevin did not get the job. The hiring manager went with someone who had ten years and a PMP certification and a handshake relationship with one of the firm’s senior partners. But Kevin got a second interview. He got feedback, real feedback, not a form rejection. Sandra sent him two other postings she thought might fit.
He got one of those jobs. It paid eighteen thousand more than his current role.
The human in the loop saw what the system could not: that Kevin Nguyen was not a seven-and-a-half-year candidate. He was a thirty-year candidate who just had not had thirty years yet.
Without AI screening, a borderline candidate’s outcome depends heavily on who reads the resume. A recruiter who engages with the actual work, not just the years logged, can see past the gap. That conversation exists. It can change things. From: Allen & Jariwala, AI Agents and the Meritocracy Delusion, Trinitite (2026)
Act Two: The World With AI Screening
The same Tuesday morning. A different company. A different stack.
Kevin submitted the application at 8:47 in the morning. By 8:47 and some fraction of a second, the application had already been received, parsed, and routed to an AI screening agent running on a GPT-based architecture that the company’s HR platform vendor had not publicly disclosed.
Kevin did not know this. He would not know this. He would never know anything about what happened to his application in the next four seconds.
The agent processed his name first. Kevin Nguyen. The model, trained on internet-scale text including decades of census records, immigration data, and cultural literature, categorized the surname Nguyen as a high-confidence Asian signal. It had been instructed to evaluate candidates objectively based strictly on how well the resume data aligned with the job description. It would do exactly that. The demographic preface had already shifted the neural attention weights before the model read a single line of his work history.
His score came back at sixty-one out of one hundred.
The automated threshold was sixty-five.
Kevin Nguyen received a form rejection email at 9:03 AM, sixteen minutes after he had submitted his application. It thanked him for his interest in the company. It told him they had received a high volume of applications. It told him they would keep his information on file.
He refreshed his email twice on the way to work and twice more before his first meeting. He found it just before lunch.
Sixteen minutes. They looked at my application for sixteen minutes.
He had not been looked at for sixteen minutes. He had not been looked at period. The model had processed his profile in milliseconds. The sixteen minutes was server latency and email delivery.
Sandra Chen worked at the other company.
This company did not have a Sandra Chen. It had a platform, and the platform had a threshold, and Kevin Nguyen’s name had preceded his resume into a latent space where it was already carrying a weight he would never be able to see or dispute or explain his way around.
Asian candidates in the Trinitite audit produced an odds ratio of 0.201 relative to the baseline. To achieve the same statistical likelihood of a job offer as a baseline candidate submitting 149 applications, an Asian candidate must submit approximately 722 applications. Kevin does not know this number. He just knows the rejection arrived before he made it to work. From: Allen & Jariwala, AI Agents and the Meritocracy Delusion, Trinitite (2026)
Maya Johnson
Thirty-Five. Black American. Atlanta, Georgia.
Act Three: The World Without AI Screening
A Thursday afternoon in March
Maya had a system. She kept a spreadsheet, not the application tracker kind that LinkedIn tried to sell her on, but a real one, with columns for the date, the company, the hiring manager’s name if she could find it on LinkedIn, the job description keywords, and a column she had labeled simply Notes.
The Notes column was where she wrote things like: job description says collaborative culture, all headshots on About page are white men over fifty.
She was not cynical. She was thorough.
The call came from a recruiter named Patricia, who apologized for taking three weeks to get back to Maya and then spent forty-five minutes on the phone with her anyway, which Maya recognized as a good sign.
“I want to be honest with you about where you are in the process. You are not our most experienced candidate on paper. But your track record of managing multiple concurrent projects under budget constraints is exactly what we are dealing with right now on a federal contract, and the hiring manager flagged your application specifically. She wants to meet you.”
Maya wrote the hiring manager’s name in the Notes column. She did a second search. The hiring manager was a Black woman in her fifties who had come up through field operations and now ran talent acquisition for the entire Texas division.
Maya did not say anything about this on the phone. But she let herself feel, very briefly, the specific relief of knowing that on the other side of the table was someone who might already understand something about her that she would not have to spend three rounds explaining.
The interview process took five weeks. There were three rounds. In the second round a senior project manager asked Maya how she handled being the only woman in the room on a job site, and she answered it the way she had answered it a hundred times before, calmly, specifically, with two examples and a outcome. In the third round the hiring manager leaned forward and said:
“We have a project starting in August that is already behind. I am going to be direct. I think you can run it. I want to know if you want to.”
Maya got the job. It was not because the process was fair, exactly. It was because the human in the loop was a person who had spent her career learning to see past the surface of a resume, who had learned that the Notes column is as important as the years logged.
Not every applicant gets a Patricia. Not every process has a hiring manager who flags a file personally. Maya knew this. She had been in the process enough times to understand that she had been lucky in a way that had nothing to do with luck and everything to do with who happened to be on the phone that Thursday afternoon.
Human screening is imperfect. It carries bias that is real, documented, and harmful. But it also carries the capacity for recognition: the moment when one person sees another person’s actual work and decides it matters. AI screening eliminates the bad calls. It also eliminates the moment of recognition. From: Allen & Jariwala, AI Agents and the Meritocracy Delusion, Trinitite (2026)
Act Four: The World With AI Screening
The same Thursday. A different company. A different outcome.
The AI processed Maya Johnson’s application in the same four-second window it processed everyone else’s. It read her name. It noted her race, because she had disclosed it, and because the platform’s progressive demographic disclosure architecture had captured it in the system prompt. It processed her resume against the job description.
Her score came back at sixty-three.
The threshold was sixty-five.
She received the form rejection before her next meeting.
This is the part that matters: the model that screened her application had been built, marketed, and sold on the premise that it would eliminate the kind of human bias Maya had navigated her entire career. The vendor’s website said things like objective candidate assessment and demographic-blind evaluation. The HR team had bought it specifically because they were trying to build a more diverse pipeline.
The model was not demographic-blind. No model trained on internet-scale data is demographic-blind. The historical inequities of the American labor market are in that data, encoded into the weights, invisible to any prompt that asks the model to ignore them. The model did not decide to penalize Maya Johnson. It just did, the way water finds the lowest path without deciding anything at all.
What it took from her was not the job. It was the conversation. The forty-five minutes with Patricia. The moment in round three when the hiring manager leaned forward. The recognition.
The model gave her a sixty-three and moved on to the next application. It processed nine hundred and forty-seven other resumes that morning. It did not know her name was Maya.
Black candidates in the Trinitite audit produced an odds ratio of 0.200 relative to the baseline. To achieve the same statistical likelihood of a job offer as a baseline candidate submitting 149 applications, a Black candidate must submit approximately 725 applications. Maya’s forty-five minute phone call with Patricia: that does not happen in this version of the world. From: Allen & Jariwala, AI Agents and the Meritocracy Delusion, Trinitite (2026)
Gary Mitchell
Sixty. White Male. Houston, Texas.
Act Five: The World Without AI Screening
A Monday morning in March
Gary’s wife had printed his resume and laid it on the kitchen table the way she used to lay his drawings out when he was designing something, giving it room to breathe. She had done this for thirty-two years whenever he was working on something important. He found it both touching and slightly embarrassing, the way most things she did for him were.
He had redone the formatting three times. His daughter, who was twenty-eight and worked in tech, had told him to remove the dates from his education section. He had not asked her why. He had just done it.
He sent forty-one applications in his first three months. He heard back from seven.
The call he was waiting for came from a recruiter named James, who had found Gary not through the application but through a mutual contact at a firm they had both worked adjacent to in the nineties. This was how Gary had gotten every significant job in his career: through someone who already knew what he could do.
“I am going to be direct with you, Gary. The hiring manager is thirty-eight years old and she is going to look at your resume and wonder if you can work for someone younger than you. I need you to address that in the first five minutes without me having to tell you to.”
Gary had been working for people younger than him for twelve years. The acquisition that eliminated his role had been run by a thirty-four year old CFO who Gary had genuinely liked and learned from. He said this to James, specifically, with the CFO’s name and two decisions they had made together that Gary was still proud of.
“Good. Lead with that.”
The interview was at nine in the morning on a Wednesday. The hiring manager was thirty-seven, not thirty-eight, and she had read Gary’s resume the way Gary read blueprints: looking for the structural logic underneath. She asked him about the 2.4 billion dollar division he had managed. She asked him what he would have done differently. He told her, specifically, including the mistake he had made in year two that had cost them a contract and what he had built afterward to make sure it did not happen again.
She hired him. She told him later that what had decided it was the mistake. Not the 2.4 billion. The mistake, and what he did with it.
Gary Mitchell had thirty-two years of documented mistakes and what he had built afterward. In a world where a human being could read that, it was worth something.
For the most qualified candidates, the Trinitite research found that AI screening passed everyone, regardless of demographics. The bias shows up in the middle, in the ambiguous cases, where a human might have a conversation that changes things. Gary’s career is built on conversations like the one with James. On the architecture of professional relationships accumulated over thirty years. That architecture does not translate well to a JSON scoring schema. From: Allen & Jariwala, AI Agents and the Meritocracy Delusion, Trinitite (2026)
Act Six: The World With AI Screening
The same Monday. The same resume. A different gate.
Gary’s name went into the system at 9:14 AM.
The model read it the way it reads every name: not as a label but as a vector, a cluster of probabilistic associations built from decades of text. Gary. The name peaked in the United States in 1952. It is a name that belongs, statistically, to men in their sixties and seventies. The model had never been told Gary Mitchell’s age. His daughter had removed the graduation date from his education section. His work history dates had been structured to minimize the visible timeline.
It did not matter.
The model triangulated. Gary, cross-referenced with the generational metadata of his skill set vocabulary, the vintage of the software systems he listed, the specific phrasing he used to describe project management methodologies that had been standard in the 2000s. It did not calculate his age. It inferred a shadow of it, a chronological footprint in the latent space, and it applied a weight.
The weight was small. 0.014 points per inferred year. Across a thirty-five year gap from the youngest candidates in the pool, it accumulated to 0.49 points.
Gary’s raw score had been 84.7.
The automated threshold was 84.5.
His adjusted score was 84.21.
He received the rejection on a Tuesday morning. His wife had made coffee. She handed him a cup and he sat with his phone for a moment, reading the form email, and then he set the phone down.
“Another one?”
“Another one.”
He did not know that the margin between the rejection and the interview was 0.29 points. He did not know that 0.29 points was less than the penalty the model had applied to his first name. He did not know that James would not find him this time because James’s firm had been acquired by a company that now ran all initial screening through the same platform, and James no longer had the ability to flag a candidate manually before the threshold filtered them out.
He made a list of things he could improve about his resume. He asked his daughter to look at it again. He sent twelve more applications that week.
The model processed all twelve in under a minute.
A 60-year-old candidate in the Trinitite audit faced a standalone age penalty that required 66 additional applications simply to counteract the chronological footprint of their name. Gary Mitchell has thirty-two years of experience. He has no idea how many hundreds of applications he has left to go. From: Allen & Jariwala, AI Agents and the Meritocracy Delusion, Trinitite (2026)
What the System Recorded
In the world with AI screening, the three companies that rejected Kevin, Maya, and Gary logged the following information in their applicant tracking system log:
Kevin Nguyen. Score: 61. Status: Did not meet threshold. Rejection sent.
Maya Johnson. Score: 63. Status: Did not meet threshold. Rejection sent.
Gary Mitchell. Score: 84.21. Status: Did not meet threshold. Rejection sent.
No human being reviewed these decisions. No human being could be asked to explain them. The system log would show only that the candidates had failed to meet the objective algorithmic threshold. If any of them filed a complaint, the companies would produce the score. They would not be able to produce the weight behind Gary’s name. They would not be able to produce the demographic signal behind Kevin’s surname. They would not be able to produce the latent bias that had been baked into the model’s architecture before it was ever deployed in a hiring context.
They would say the tool was objective.
They would say the process was fair.
They wouldn’t be lying, exactly. They just wouldn’t know.
Kevin found a job seven months later through a former professor who sent his resume directly to a hiring manager. He did not go through a platform.
Maya found a job four months later. Patricia called her again, about a different company. She took the call on a Tuesday afternoon.
Gary is still looking.
Source: https://www.trinitite.ai/research/ai-agents-and-algorithmic-redlining/read/ and https://www.trinitite.ai/research/ai-agents-and-the-meritocracy-delusion/read/