Eliminating Bias Starts with Truth

Why deterministic verification removes what resumes and gut-feel can't.

Someone recently asked us a hard question: "Doesn't Alloy just create a new kind of bias?"

It's a fair concern. Every new system that scores humans carries the risk of encoding the prejudices of its creators. AI screening tools have earned that suspicion. They're black boxes that ingest resumes, apply opaque algorithms, and output rankings that no one can explain — not even the engineers who built them.

But here's where Alloy is fundamentally different. We don't score opinions. We verify facts. And facts, by definition, don't have a point of view.

The bias problem is structural

Hiring has always been a system of compounding biases. A recruiter spends an average of 7.4 seconds on a resume. In that window, confirmation bias takes over — they see a prestigious university and assume competence. They see an unfamiliar employer and assume risk. They anchor on the first data point and filter everything else through it.

These aren't character flaws. They're cognitive shortcuts that evolution optimized for survival, not for evaluating a cloud architect's Kubernetes experience. Daniel Kahneman documented this decades ago: System 1 thinking is fast, automatic, and riddled with heuristic traps. Hiring runs almost entirely on System 1.

Then we added technology and made it worse.

AI resume screeners promised to remove human bias from the equation. Instead, they encoded it. They train on historical hiring data — data that reflects every bias the organization has accumulated over decades. The algorithm doesn't eliminate the pattern. It scales it.

Truth as the antidote

There's a concept in philosophy called the correspondence theory of truth: a claim is true if and only if it corresponds to an actual state of affairs in the world. Your resume says you were VP of Engineering at a defense contractor from 2019 to 2023. That claim is either true or it isn't. It corresponds to reality, or it doesn't.

This is where Alloy starts. Not with interpretation. Not with prediction. With correspondence.

When a candidate submits their professional history, Alloy doesn't ask "does this look right?" It asks "can this be proven?" Every claim is tested against 14 independent public sources:

A resume is an interpretation. SEC EDGAR is a fact. A Credly badge is a fact. A patent filing with the USPTO is a fact. Alloy separates the two categories and only scores the second one.

The falsification principle

Karl Popper argued that real knowledge comes not from confirming what you believe, but from trying to disprove it. A theory that can't be falsified isn't science — it's faith. The same principle applies to candidate evaluation.

Traditional screening is confirmatory. You read a resume looking for reasons to say yes or no, and you find what you're looking for because that's how confirmation bias works. The information you encounter first anchors every judgment that follows.

Alloy works in the opposite direction. It doesn't try to confirm claims. It tries to disprove them. Every fact category — identity, employment, education, certifications, skills, contributions — is tested against independent sources that have no incentive to agree with the candidate. When a claim survives that adversarial process across multiple sources, you can trust it. Not because a recruiter felt good about it. Because it was tested and held.

Corroboration, not consensus

Intelligence analysts have known for decades that single-source intelligence is unreliable. A defector's testimony might be true, but it might also be fabricated. The standard for actionable intelligence is multi-source corroboration: the same fact confirmed independently by sources that have no connection to each other.

Alloy applies this methodology to professional identity. A claim confirmed by one source gets a baseline score. The same claim confirmed by three independent sources triggers a corroboration multiplier — because the probability of three unconnected sources all being wrong about the same fact is vanishingly small.

Consider a candidate who claims five years as a senior engineer at a defense contractor. If their LinkedIn shows the role, that's a data point. If the employer's SAM.gov registration confirms the company existed and held relevant contracts during that period, that's corroboration. If the candidate's GitHub contributions show activity consistent with the claimed tech stack, that's a third independent confirmation. Each source has no knowledge of the others. Their agreement is not coordinated. It's evidence.

A score built on three-source corroboration across identity, employment, and skills cannot be inflated by a polished resume. And it cannot be deflated by an interviewer's unconscious bias.

Measuring what you don't know

Most hiring tools pretend certainty. A resume screening score of 87 implies precision that doesn't exist. What does 87 mean? Compared to what? Based on what evidence?

Alloy takes the opposite approach. It measures uncertainty explicitly. Every Alloy ID falls into one of four verification tiers:

Verified — claims corroborated by multiple independent sources across all major fact categories.

Strongly Verified — strong corroboration in most categories, limited gaps in one or two areas.

Partially Verified — some claims confirmed, but significant categories lack independent corroboration.

Limited Evidence — few or no claims could be confirmed against independent sources.

Notice what's missing: there's no "Unverified" or "Failed" tier. Limited evidence doesn't mean the candidate is lying. It means the evidence isn't publicly available. That distinction matters. A candidate with legitimate classified work history may have limited public evidence for good reason. The tier system tells you what you know, not what you should assume.

Why deterministic scoring matters

Here's the technical detail that makes the bias argument collapse: there is no LLM in Alloy's verification path.

Large language models are powerful tools, but they're probabilistic. They generate likely outputs, not guaranteed ones. Two identical inputs can produce different scores on different runs. That's fine for writing marketing copy. It's not fine for evaluating a human being's professional credibility.

Alloy's scoring is deterministic. The same evidence always produces the same score. Every point maps to a specific source, a specific fact, a specific timestamp. You can audit it. You can challenge it. You can trace any score back to the public record that generated it.

This is what makes Alloy bias-resistant by design, not by intention. It's not that we tried hard to remove bias. It's that the architecture doesn't have a place for bias to enter. There's no human in the scoring loop. There's no model making judgment calls. There's evidence, and there's math.

The evidence grid

Every Alloy ID includes a visual evidence grid — a matrix showing exactly which claims are supported by which sources. It's not a score. It's a map of what's proven versus what's claimed.

Six fact categories across 14 sources. Each intersection is either confirmed, unconfirmed, or not applicable. No ambiguity. No interpretation. The hiring manager sees the same grid the candidate sees. Both can point to the same cell and ask the same question: what does this source say?

When combined with LinkedIn Verified — which confirms government-issued identity — the evidence grid creates a foundation of trust that no resume, no interview, and no AI screening tool can match. Because it isn't asking you to trust a person or an algorithm. It's asking you to trust public records.

Information systems amplify what you feed them

Every information system amplifies the properties of its inputs. Social media algorithms amplify engagement, which means they amplify outrage. Resume screening algorithms amplify pattern matching, which means they amplify conformity. If your input is biased, your output is biased at scale.

Alloy's input is public evidence. Its output is a verification of that evidence. The system amplifies the only thing worth amplifying: whether claims correspond to reality.

So does Alloy create bias? Only if public records are biased. And if your concern is that SEC filings, patent databases, and professional certification bodies are systematically discriminating against candidates — that's a problem worth solving, but it's not a problem Alloy created. It's a problem Alloy made visible.

Visibility is the first step toward accountability. You can't fix what you can't see.

The real question

The engineer who raised this concern was asking the right question. Every system that evaluates people should be interrogated for bias. But the answer to "does this system encode bias?" depends on what the system actually does.

If it reads resumes and predicts job performance, it will encode bias. If it scores keywords and ranks candidates, it will encode bias. If it uses historical hiring data to train a model, it will encode bias.

If it checks whether a claim corresponds to a public record, it doesn't encode bias. It encodes truth.