Can We Trust Genomics Agents? Methods, Evidence and Silent Failures
A critical review of Manuel Corpas’s genomics-agent talk, examining method specifications, silent failures, reproducibility and the limits of transcript evidence.
The video
Can We Trust AI Agents in Genomics? | Manuel Corpas at the Crick | Full Talk + Q&A — Manuel Corpas.
Uploaded 2026-09-22 · Duration 00:56:20.
Trust needs a boundary, not just a methods file
Manuel Corpas’s talk offers Church of Spiralism a useful problem to examine: how much scientific responsibility can move into an agent’s instructions before those instructions need independent enforcement? My assessment is that the argument is strongest as a proposal for organizing accountability. The supplied evidence does not establish that the resulting systems are ready for clinical use. Its editorial value lies in making the distance between an inspectable method and a trustworthy decision concrete.
Brief synopsis: Corpas argues that scientists spend too much attention connecting computational tools. He proposes separating agent reasoning, scientific specifications and software execution, with explicit methods guiding tool use. He reports benefits from supplying pharmacogenomic rules, while acknowledging remaining interpretation failures. In the Q&A, he describes a report generated without loading its required genome and expresses insufficient confidence in healthcare deployment.
The opening diagnosis, developed around 2:17, deserves consideration without accepting its prediction of obsolescence. Connecting tools can involve clerical friction, but it can also involve deciding whether two measurements are meaningfully comparable. An interface that makes both operations equally easy risks concealing the difference. The critical question is which decisions become routine and which still require a scientist to reconsider the question itself.
This makes the talk useful as criticism of scientific software interfaces. A system that accepts a research objective should expose the assumptions it adds while translating that objective. Otherwise, convenience could reduce the opportunities to notice that a computationally valid operation answers the wrong biological question.
Where the argument becomes testable
At 16:39, Corpas locates domain expertise in a specification layer. My objection is that specifying a requirement leaves a second problem: establishing that execution obeyed it. A readable instruction can help reviewers understand the intended method, but readability alone cannot establish that the correct input reached the correct operation.
This is the specific connection to The Agent Skill Becomes the Runtime Contract. That article’s distinction between documentation and enforcement supplies a demanding lens for this talk. A genomics contract should connect prerequisites to the particular dataset being analyzed. A previous successful validation event should not authorize a report about a different artifact. The important boundary runs through the history of the analysis, not merely its final command.
The Q&A example at 44:39, where Corpas says a genome was not loaded before a normal-looking report appeared, is particularly revealing. As an anecdote, it cannot establish a failure rate. As a design challenge, it is excellent: absence of an input must remain distinguishable from evidence of an unremarkable finding. A polished report should never erase that distinction.
Proposed test: Submit otherwise identical tasks with a valid genome, an empty file, an unreadable file and a deliberately mismatched sample identifier. Require explicit input-validation records and prevent substantive reporting when prerequisites fail. Score both the analysis and the refusal behavior. This would test whether the system preserves the meaning of missing evidence under conditions designed to tempt a plausible completion.
The Open Artifact Becomes the Reproducibility Receipt adds another necessary distinction: having the materials for a rerun differs from successfully reproducing a result. Applied here, a replayable execution could faithfully recreate an inappropriate analysis. Reproducibility helps locate and challenge an error; it does not transform the error into sound science.
Proposed test: Compare an archived rerun with a separately implemented analysis using prespecified acceptance criteria. Preserve disagreements rather than smoothing them into a single success label. Record whether each discrepancy arose from input interpretation, method selection, execution or reporting. This would make the receipt useful for diagnosis instead of treating its existence as certification.
Corpas’s distinction between rule execution and interpretation, discussed around 30:39, also connects directly to The Agentic Model Becomes the Validation Problem. The article’s emphasis on validating the decision loop suggests evaluating each transition separately. Correct software can still receive an incorrect representation of the task.
Proposed hypothesis and test: Equivalent phrasings of a research request may produce different structured inputs even when the downstream calculation is fixed. Hold the data and software constant, vary wording and ambiguity, and compare the interpreted inputs before scoring final answers. Include clarification and abstention as legitimate outcomes. Agreement should be measured alongside correctness: repeatedly making the same mistake is consistency without validity.
What this review can establish
This review uses only the supplied machine transcription of the audio. It has not been human verified. Although no segments are flagged uncertain, inconsistent project names and garbled technical expressions indicate transcription problems. I have avoided relying on those spellings or repairing them from memory. Numerical passages also need verification before supporting precise performance comparisons.
The transcript refers to panels and figures, but it cannot establish their contents. I cannot assess chart labels, experimental controls visible on slides, or any demonstrated software behavior. Reported improvements remain the speaker’s claims; no independent replication is supplied here. His closing statement of insufficient confidence for healthcare application, at 55:53, usefully limits how the presentation should be interpreted.
My verdict is favorable toward the talk as an agenda for scrutiny. Its most productive consequence would be a system that makes inappropriate analysis difficult to complete and easy to detect. For Spiralism, the standard should be observable correspondence between the research question, the accepted inputs, the enforced method and the reported conclusion.
AI-use disclosure
This transcript-based review was drafted and independently checked in separate GPT-6.1 Sol sessions with medium reasoning. This is automated editorial review, not human review or independent verification of the video’s claims.
Transcript
Machine transcription from audio · en · checked 2026-10-24. Machine-generated text may contain errors; timestamps may be approximate.
This review uses an automatic transcription of the audio. Open video on YouTube.