Thesis
From a thesis to an infrastructure
Why the hardest problem in biology and the hardest problem in biopharma are the same problem — and what it takes to solve it.
I was sitting in a meeting about a promising asset for lupus when the two halves of my life collided.
Lupus disproportionately affects Black and Hispanic women — many lower-income, many already underserved by the healthcare system. The asset had a real mechanism and a population with everything to gain. It did not advance into that indication. Not because anyone in the room was careless or unkind, but because the evidence available couldn't make the case defensible enough to justify the commercial risk. That is a rational decision given the tools the industry has. It is also a decision whose cost is never paid by the people in the boardroom.
I have lupus. I was diagnosed after contracting COVID-19 while working as clinical staff in a Boston ICU. So I was sitting in that meeting as the patient and the operator at once — and as something else, too. Because I recognised the shape of the problem in front of me. I had spent years studying it, in a completely different domain.
This essay is about that shape, and about why naming it correctly changes what you decide to build.
What my research actually showed
My research career began in high school and became the throughline of my life. For years it ran alongside a clinical track — I trained and worked toward medicine — but the questions that held me were always research questions. The one that held me longest: how does cumulative stress become biology, and why does the burden land hardest on the people already carrying the most?
My graduate research at Johns Hopkins answered a version of it directly. Working from the weathering hypothesis — the idea that chronic exposure to social and economic disadvantage produces accelerated physiological decline — I synthesised the evidence that maternal stress transmits biologically into the next generation. Chronic stress, borne disproportionately by Black women through economic, social and political exclusion, is associated with dysregulation of the body's stress and immune signalling, with epigenetic modification, and ultimately with elevated risk that offspring develop autoimmune disease later in life. Stress is not only psychological. It becomes immune dysregulation. And the cost is inequitably distributed across exactly the populations least equipped to absorb it.
That finding matters on its own. But here I want to draw attention to something about the evidence itself, because it is the hinge of everything that follows. The mechanisms I was studying share a particular structure: the cause is distant from the effect, often by decades; the pathway is mediated through layers that are real but hard to observe directly; the evidence is fragmented across many partial sources, no one of which tells the whole story; and there is no auditable chain running cleanly from a piece of evidence to a conclusion. You assemble the picture from pieces, and the assembly lives in an expert's head.
Distant cause, mediated mechanism, fragmented evidence, no auditable chain.
The pattern is everywhere
Here is the claim this essay exists to make, and I want to argue it rather than assert it: the structure of the hardest questions in biology is the same structure as the hardest questions in therapeutic decision-making.
Consider what it takes to evaluate whether a biopharma asset should advance — to develop, license, acquire or fund it. You are reasoning about a distant outcome: will this work, years from now, in a population you can't fully observe yet. The mechanism connecting the asset to that outcome is mediated through biology, trial design, regulatory precedent, manufacturing, market access. The evidence is scattered across literature, trial registries, regulatory filings, patents, payer data and proprietary judgment. And the reasoning that ties it together is, today, almost entirely editorial — it lives in the heads of experts, rendered as opinion, with the provenance of each underlying claim unrecoverable and the chain of inference unauditable.
Distant cause. Mediated mechanism. Fragmented evidence. No auditable chain.
It is the same shape. The thing that made my thesis hard to do well is the same thing that makes asset evaluation hard to do well. Once you see the isomorphism, you cannot unsee it — and you start asking a different question. Not "what is the right answer for this asset," but "what would it take to make the reasoning itself trustworthy."
One structure underlies both — and the inequity studied as a mechanism in the body reappears as a mechanism in the market.
Why today's infrastructure fails this shape — and who pays
The biopharma industry evaluates assets with infrastructure that is editorial, opaque and slow. Expert synthesis is valuable, but it is unverifiable by design: the conclusions rest on brand and reputation rather than on a chain anyone outside the room can inspect. Bolting generic AI onto that process does not fix it. It makes it worse — adding fluency, and with it the risk of confident fabrication, to decisions that already lacked verifiable evidence chains. You get answers that sound more authoritative and are less possible to audit.
And that pressure only runs one way. Every improvement in these systems raises the volume of confident, unaccountable output; none of them raises the number of parties willing to put a name on a conclusion and be findable later if it was wrong. As generated analysis approaches zero cost and unbounded supply, the scarce thing is not insight. It is accountability, and a reader who can check without trusting anyone.
And the failure is not evenly distributed. When the evidence supporting a high-need indication cannot be made governed, certified and defensible enough to clear the commercial bar, the rational decision is to not pursue it. The asset goes to the indication that's easier to justify. The harder, higher-need indication — often serving a smaller, poorer, or more marginalised population — is left behind.
This is where the essay's two halves close into one. The populations the science says bear the biological burden of systemic disadvantage are the same populations whose indications get dropped when the evidence infrastructure can't make their case. The inequity I studied as a mechanism in the body reappears as a mechanism in the market. Same shape, second time. The infrastructure that decides which diseases get pursued is, quietly, an engine of the same disparity.
That is not a problem you solve with a better opinion. It is a problem you solve with better infrastructure for evidence itself.
The principles of a fix
So I built the fix — Foster Rx, an evidence-grade evaluation layer for biopharma. I'm not going to describe how the engine works here; the mechanics are where the genuine novelty lives, and they belong in their proper venue, not an essay. But the principles the architecture commits to are worth stating plainly, because the principles are the argument. Each is a stance about what it means to make evidence trustworthy.
Provenance on every claim
Every claim traces to a named, classified, timestamped source. Nothing floats free of its origin. If you can't say where a claim came from, it isn't a claim — it's an assertion, and assertions are what we're trying to leave behind.
The distinction between what is ingested and what is computed
A claim is either a fact pulled from a named source, or a value derived by a method operating on other claims — and the two are never blurred. A derived claim carries attribution back to its inputs, recursively, so the full lineage of any conclusion traces to the ground-truth sources underneath it. You can always tell the difference between something the system found and something it figured, and follow either one home.
Honest emptiness
Where the evidence is insufficient, the system says so — explicitly, with the gap named and explained — rather than filling the hole with something plausible. A platform built to make evidence trustworthy cannot fabricate to look complete. The gaps are part of the truth.
Outcome anchoring
Predictions are not left to stand on confidence. As real-world outcomes arrive, they are reconciled against what was claimed, so accuracy becomes demonstrable over time. An evaluation that never has to meet reality is not an evaluation; it's a guess with good grammar.
Asymmetric flow between evidence and inference
Certified evidence grounds downstream analysis — but downstream analysis never contaminates the evidence base. Suggestion and pattern-matching are kept rigorously separate from certified ground truth, flowing one direction only. The moment suggestion can write back into evidence, the word "certified" stops meaning anything.
Certification as a category, not a feature
Taken together, these are not a citation footnote bolted onto an AI product. They are a governance layer for evidence — where accuracy is demonstrable and the reasoning is built to be inspected by an investment committee, a board, or a regulator. The output is not an opinion that asks to be trusted. It is a certified artifact that can be verified.
That is as far into the building as this essay goes. The principles are the part worth arguing in public. How they are achieved is the part worth protecting.
The throughline
The thesis and the platform are the same inquiry at two scales. As a researcher, I asked how fragmented, hard-to-audit evidence about distant causes produces inequitable outcomes in the body. As a founder, I am answering how fragmented, hard-to-audit evidence about distant outcomes produces inequitable decisions in the market. The first question I studied. The second I am building the answer to.
I want to be precise about the motivation, because it is easy to misread. I did not build this because I am a patient, though I am one. I built it because I am an operator who spent years inside clinical research and trial execution, who watched high-need indications get rationally abandoned for want of defensible evidence, and who happened to have studied — in a different domain entirely — exactly why that failure mode produces the inequity it does. The lived experience gave me the stakes. The research gave me the lens. The operating years gave me the vantage point to see that the infrastructure was the problem. The company is what those three things produce when they converge on one person and refuse to let go.
Better evidence infrastructure should not only serve the indications that are easy to justify.
That is the whole of it. The asset for lupus that didn't advance, the populations my research said were already carrying the heaviest biological load, the boardroom decision that was rational and still wrong at the level of the system that forced it — all of it points at the same fix. Make the evidence governed, certified and defensible, and the decision to pursue an underserved population can rest on something better than the courage of the people in the room.
That is the infrastructure I needed as a researcher, and the one I now need as a patient. So I'm building it.
Since this was written
The principles above are no longer commitments. Honest emptiness is four evidence states enforced in the schema, where a section with nothing behind it renders with its explanation and cannot be silently omitted. Outcome anchoring is a prediction register that records every assessment produced, transmitted or not, so calibration is measured over all of it rather than over the favourable half. The asymmetric flow is two visually distinct registers, where an unreviewed answer is structurally unable to carry a seal.
And the part that was hardest to build is the part that matters most: the reader checks it themselves. A certificate verifies at a public address with no account, no fee, and no server of ours in the path — so verification does not depend on our cooperation, our uptime, or our continued existence.
Methods are developed with an academic research partner who has served as principal consultant on evaluation panels of the U.S. National Academies of Sciences, Engineering and Medicine, authoring reports to Congress on the effectiveness of federal research and development programmes for innovation and entrepreneurship. No methodology is validated only by the people who wrote it.