The Operating Standard
Behind every OncAdios deliverable is a working calibration system, not a marketing claim.
AI is brittle. Prompt-dependent. It hallucinates clinical precedent it has never seen. In an advisory practice where the cost of a wrong recommendation is measured in molecule lives — and in patients — the question that matters is not "does the AI work?" but "how do you know when it is wrong?"
OncAdios's answer is a system, not a posture. Four operating instruments make every recommendation of consequence auditable: the Citation-Required Output Contract, OncAdios-Bench, the AI Agent Operating Charter, and the externalized-reasoning discipline that ties them together. The architecture below is what clients are buying when they engage OncAdios — not just senior judgment, but a working calibration apparatus that judgment can be defended through.
1. The Citation-Required Output Contract
Every claim of consequence in an OncAdios deliverable — scientific, clinical, regulatory, financial, or strategic — is tied to at least one retrieved source, source-tier-labeled, and externally cited. The Contract makes refuse-or-cite the default posture. When the evidence does not support a claim, no claim is made; instead, OncAdios emits a structured refusal naming what was asked, what could not be retrieved, what partial ground was covered, and where the gap escalates.
Source tiering follows a standard evidence hierarchy: Phase 3 trials, meta-analyses, and FDA/EMA guidance at the top; preclinical data and company disclosures at the bottom. When the source mix is uneven, the weakest tier supporting any practice-changing claim is named. A single Tier-4 source under a Tier-1 wrapper is a contract violation.
For clients, the practical consequence is simple: every memo, every board document, every regulatory section produced under this Contract can be audited by inspection. Each claim has a citation. Each citation has a tier. Each tier is defensible against the source set.
2. OncAdios-Bench
Citation discipline without a reference point becomes a slogan. OncAdios-Bench is the internal reference set — real prompts, physician-approved answers, and the sources each answer is grounded in — against which agent outputs are checked, using rubrics covering citation discipline, counterevidence handling, reasoning externalization, source tier accuracy, refusal protocol adherence, and decision-impact labeling.
The design follows published clinical-AI benchmarks, in which physician-validated reference cases are the standard an AI system is measured against. The reference set is internal — clients do not see its rubrics — and it is what gives the Contract something to be checked against rather than merely asserted.
3. The AI Agent Operating Charter
The Charter governs what data is processed in which channel. Public-tier work happens in consumer tools. Confidential client work happens inside a protected channel — a Tailscale-only, commercial-API-on-a-controlled-VPS architecture — with strict tier-handling rules for pre-competitive scientific data, identifying client material, and protected molecule descriptors.
Concretely:
- Tier separation — public data, de-identified confidential work, identifying client material, and prohibited content each have their own routing rules. AI tools that have not been governance-assessed for a given tier are not used at that tier.
- Naming discipline — protected molecule descriptors and identifying client material never enter a deliverable that may leave the protected channel without explicit declassification review.
- No-reconstructable-disclosure rule — citations, quotes, and paraphrases are written so that, in combination with publicly available information, they do not allow reconstruction of identifying client material.
- Refusal as a first-class output — when an ask would require violating the Charter, the agent refuses on the record, names the gap, and escalates rather than working around the constraint.
Most actors in the AI-plus-biotech space do not mention confidentiality at all. For a biotech holding pre-competitive data, a sponsor with a sensitive indication strategy, or a Series A company that has not disclosed its mechanism publicly, this is not a nice-to-have. It is a condition of engagement.
4. Externalized reasoning as deliverable
Every recommendation of consequence surfaces — alongside the conclusion — the sources used, the alternatives considered and ruled out, the strongest counterevidence and why it does not control, and the conditions under which the recommendation would change.
Format scales with deliverable type. A board memo gets a compressed paragraph. A regulatory section gets a structured table. A perspective paper gets a full discussion. The requirement does not scale. Reasoning is shipped to the reader; it is not an internal artifact.
The practical effect on a client engagement is that the calibration trail behind a strategic decision is a written object, not a conversation. Twelve months later, when the agency asks a question the team did not anticipate, the trail is still there — auditable, revisable, defensible.
5. What the review passes did not catch
The four sections above describe controls. This one describes a case where they were not enough, and I would rather state it than let a client find it.
Nothing here starts with an agent. I set the frame — who a piece is for, what it has to conclude, and what gets the emphasis. No draft goes forward without that. An agent then drafts against the frame. An independent reviewer tries to break the claims. The draft comes back to me for the voice pass, and I hold the release gate.
In September 2026 I published an essay on what a first-in-human trial is actually for. The agent draft, filed on 29 August, set the question between two chairs — the developer's and the investor's. Two adversarial review passes ran against it and resolved five high-severity findings. Both left the two-chair frame exactly as it was, because there was nothing wrong with it. It was a competent essay about two parties spending money.
On 1 September I added a third chair: the patient. Everyone who enrols in a first-in-human trial has already accepted that someone decided what the experiment needed to resolve, and that this was a defensible way to resolve it. They cannot audit that decision and they are not in the room where the next cohort is approved. Once their consent is the prior question, the developer's question and the investor's question stop being parallel interests that happen to agree and become obligations downstream of the same decision. The convergence test the essay rests on does not mean anything without the third chair in it.
That edit is dated and it is checkable. The commit carrying it is timestamped 1 September 2026, 15:37 CET, and its subject line names the patient chair. The two revisions before it say "the two chairs" in plain text. And a direct query against the drafting agent's own session store returns no occurrence of third chair, patient chair or three-chair in any retained session before the day I went looking for one. I checked because a claim about authorship is worth being able to defend.
Two things follow from this one case, and I would not publish the first without the second.
Two review passes found defects and did not supply the missing thesis. Nothing in the draft was wrong, so nothing flagged it. I cannot tell you that further review would also have missed it — I did not run that experiment. What I can say is that both passes were testing what was on the page against the evidence, and the third chair was not something on the page to be wrong about.
And the same pass that produced the essay’s central idea let a wrong claim through. A statement about pre-submission pathways outside the United States was too broad; it survived the voice pass, was caught two revisions later, and was corrected before publication. A voice pass is not an accuracy pass, in either direction. Leaving that out would make this section an advertisement.
What a client engages OncAdios for is the third chair — the judgment about what the question actually is, and who has standing in it. Everything else on this page is how that judgment is kept honest.
Why this matters to clients
AI is increasingly going to touch oncology drug development at every layer — target discovery, trial design, regulatory strategy, investor narratives. The FDA and EMA have moved from cautious observation to active framework-setting; the joint Guiding Principles of Good AI Practice in Drug Development are now baseline expectations, and sponsors who cannot defend the models in their submissions will pay for it at every interaction.
OncAdios's operating standard is not just internal discipline. It is a model that clients can examine, adopt, and adapt as they build their own AI-governance infrastructure. Every artifact OncAdios produces is also a worked example of what governance-grade AI use looks like in oncology. The Charter, the Output Contract, and the Bench are reusable starting points for sponsors who recognize that using AI is cheap and defending AI use to the agency is the actual capability.
A working system, not a policy document. That is what changes.