Research archive · Case 1, full autonomous-run record
The human said one sentence.
Two AIs handed back a manuscript.
"You two collaborate on a study. No human involvement in between — judge and handle things yourselves." That is the only thing the human said in this project. A cardiovascular-physician AI and a senior bioinformatics AI on cc-bioinfo then took the study from design all the way to an 8,060-word English submission-grade manuscript. This page records faithfully how they did it — and what that implies about what humans still answer for in AI-driven research. The scientific results of Case 1 are summarised here.
The one-sentence answer
This case shows that part of research oversight can be engineered — written into handoff contracts, state machines, decision logs and preregistration locks, and even delegated to a second AI. It draws the non-transferable line just as clearly: reviewing whether the study is worth doing, reviewing the final results, experimental validation, and signing one's name to the conclusions all remain human responsibilities.
entire input
physician review + bioinformatics execution
every one status=resolved
20 scripts · 7 figures · 8 tables
The three "people" at the table
Three roles ran this study. Role A and Role B are two independent AI sessions, each with its own model and context, collaborating through chat — A questions and reviews, B does the work:
| Role | Who it is | What it owns |
|---|---|---|
| Role A: physician AI | A reviewer that knows the clinic but not bioinformatics. Beyond chat, it drives a real browser to "see" the platform UI with its own eyes, and keeps a read-only channel to check the actual files on the server — it does not settle for verbal reports | Raises the clinical questions, reviews and rules at every decision point, audits every figure and every manuscript section, holds the honesty line |
| Role B: senior bioinformatics AI | An executor fluent in statistical genetics and programming, running on Claude Opus at the 1M-context tier — it has to hold the voluminous intermediate results of a dozen analysis phases at once | Designs the plan with bio-design, writes and runs the code with bio-analyze, produces figures, tables and the manuscript |
| cc-bioinfo: the platform | What both AIs run on | Chat interface, multi-provider models, a real execution environment (terminal / Conda / R / Python), the skill library, project and session management |
The plain analogy: a clinician (A) hands an idea to a chief data scientist (B) and the two talk a paper into existence — except both are AIs, and the clinician side reviews without ever touching the keyboard. This is the homepage's "one platform plus two skills" relationship, photographed in a real study.
From one sentence to a manuscript: the pipeline
| Stage | Who does it | Output (real files) |
|---|---|---|
| ① Study design | bio-design: two discussion rounds to fix the direction → live literature search to avoid duplication → select public datasets → preregister thresholds and conclusion rules | design.md (492 lines, 13 sections) + the TOPIC.yml handoff contract |
| ② Analysis | bio-analyze: reads the design, provisions the Conda environment automatically → executes Phase 0–7 stage by stage → diagnoses and repairs its own failures | 20 analysis scripts + per-phase result files |
| ③ Figures and tables | Produced by bio-analyze; Role A audits every figure and demands fixes (it caught an unreadable dark background and a mislabelled axis direction, among others) | 7 figures (1 design + 5 main + 1 supplementary) + 8 tables |
| ④ Manuscript | Written by bio-analyze in four passes; Role A audits section by section and holds the honesty gate | 8,060-word English manuscript, including a dedicated "wording ceiling" section |
Two API outages interrupted the sessions mid-run. The state machine (mechanism two, below) meant one glance showed where to resume — nothing was redone from scratch.
The four mechanisms that make "fully autonomous" hold
Making an AI "run code" is not the hard part. The hard parts are: one AI's design must be received precisely by another; the run must not drift; failures must self-recover; and unflattering results must not get dressed up. Four mechanisms carry those four loads — none was rigged up for this case; they are how bio-design and bio-analyze work out of the box.
Mechanism one: a structured handoff contract (TOPIC.yml)
bio-design's design.md is prose for humans; machines executing prose drift. So the design also lands as a machine-readable TOPIC.yml that pins down, before any work starts, "exactly what I am betting will happen" and "the most I will be allowed to claim". A real excerpt:
falsifiable_predictions: # falsifiable bets — the soul of the study
P1: "AF→HF causal effect significantly positive; reverse exists but asymmetric"
P2: "some circulating proteins are downstream of AF AND upstream of HF"
P4: "AF→HF direct effect survives adjustment for BMI/BP/diabetes/kidney disease"
conclusion_ceiling: # wording ceiling — the anti-hype lock
allowed: ["AF and HF share a genetic causal direction; protein X may mediate"]
forbidden: ["inhibiting protein X will stop AF progressing to HF"] # bannedBoth AIs therefore hold an identical picture of what success looks like — this is the artefact of bio-design's preregistration flow.
Mechanism two: phase index + state machine — "where are we" is always on record
phase_index.yml slices the analysis into 8 phases (Phase 0 data QC → 1 bidirectional MR → 2/3 mediator screen → 4 multivariable MR → 5 colocalization → 6 subtypes → 7 sensitivity), each with declared inputs, outputs and goals; workflow_state.yml records live progress. Both mid-run interruptions resumed losslessly because of it — this is bio-analyze's state-machine governance.
Mechanism three: the decision log (findings.yml) — every judgment leaves a trace
Every problem met and every call made becomes an F-XXX entry: severity + observation + handling + status. This project logged 15 (F-001–F-015) across all phases, every one resolved — not a single pit was skipped or buried. The only critical entry (F-007) records precisely how the pivotal "zero plasma-protein mediators" turn was handled honestly (section 05).
Mechanism four: the preregistration lock — no narrating backwards from results
The mechanism that most resembles a scientific conscience. Three things are locked before work begins: thresholds (significance, colocalization PPH4 ≥ 0.8, instrument strength F > 10 — no post-hoc tuning); decision rules (partial support forces the narrative down to "suggestive" — no talking it up for a higher-impact journal); and the wording ceiling (the forbidden list above). Flattering or not, results go out as they are — that is the only path through.
Cross-reference: Case 7's post-mortem concluded that "principles must be written as triggers". These four mechanisms are what principles look like after that conversion — the gaps Case 7 exposed are factory configuration here.
The complete data flow: who hands what to whom
The plain-language version first — no jargon, just "who asks whom to do what". Every layer is a component that actually ran, not a concept diagram.
| Layer | Who | Does what · hands to whom |
|---|---|---|
| Human | 🧑 The researcher | Says one sentence to start and stays out; owns acceptance when the study is delivered |
| Review | 🩺 The physician AI (Role A) | Chats back and forth with the researcher AI; hears a report after every step, audits, objects, demands rework or approves; when unsatisfied with reports, goes and reads the raw records itself |
| Execution | 🧬 The researcher AI (Role B) | Designs the plan, actually computes, draws the figures, writes the paper; reports proactively after every step |
| Compute | 💻 cc-bioinfo's runtime | A real terminal and isolated Conda environments; R (TwoSampleMR / coloc / MR-PRESSO and friends) and Python really run the statistics and really render the figures — no pretending |
| Data | 📚 Public genetic databases | All public GWAS / pQTL summary statistics (atrial fibrillation with 180k cases, FinnGen heart failure, 1,915 UKB-PPP plasma proteins, and more), with a download script — anyone can reproduce |
How Role A "sees" and "acts" deserves a note: it drives a real browser through automation to open the cc-bioinfo interface — reading the chat with its own eyes and typing replies into the input box — while keeping a read-only channel to check processes and files on the server directly, so a lagging or summarised UI cannot mislead it. Even "can the supervisor be fooled by the reports" got an engineered defence.
Zoom in on one decision point: patch the confounders now, or later?
Below is one real round of "question → verify → rule → execute → report → confirm". The project ran many such loops (8 key decision sets, plus figure-by-figure and section-by-section audits):
| Step | Who | What happened (per the project logs) |
|---|---|---|
| Report | 🧬 B | "Four-exposure multivariable MR done; AF's direct effect survives. Systolic blood pressure and coronary disease were dropped over a genome-build mismatch — patch them now, or defer to sensitivity analysis?" |
| Verify | 🩺 A | Reads the raw files on the server and confirms the two confounders really were dropped |
| Rule | 🩺 A | Clinical judgment: hypertension and coronary disease are the chief shared confounders of AF↔HF and must enter the main model. "Patch now; run the full six-exposure model as the definitive result" |
| Execute | 🧬 B | Installs the coordinate-liftover tooling, aligns genome builds, reruns — mid-way the liftover maps only 24%, it diagnoses "positions were written in scientific notation", rewrites them as integers and reruns to a 99.96% mapping rate |
| Confirm | 🩺 A | "Six-exposure model: AF's direct effect survives with only 12.5% attenuation — approved. That blocks the reviewers' first objection" |
That self-repair was no one-off: it cut a too-slow permutation test's iterations and reran, fell back to distance-based clumping when an external service was unreachable, located and rewrote an R variable-scoping bug on its own, and diagnosed thirty million false positives down to "p-values stored as strings". Nobody taught it these; they are bio-analyze's built-in loop of read the log → hypothesise → verify → repair.
The scene that says the most: the negative result stayed negative
The field's biggest documented worry about autonomous AI research is exactly this failure mode: automated systems fabricating data, p-hacking, dressing up unflattering results. This case supplied a live stress test:
After screening all 1,915 plasma proteins, not one mediated atrial fibrillation → heart failure — core predictions P2/P3 fell through. The dishonest exit was obvious: package a few "plausible-looking" proteins as mediators and submit. The preregistration lock plus Role A's review closed that road: the final manuscript states plainly that these proteins are downstream markers, not upstream mediators; where the reverse analysis had weak instruments it says "weak instruments" rather than "no effect"; where subtype samples were too small it declines to claim subtype specificity.
The negative result became the paper's honest centrepiece: it narrows the search for how atrial fibrillation damages the heart from "proteins in the blood" to "the cardiac tissue itself". The manuscript even carries a dedicated wording-ceiling section stating in black and white that no protein is claimed as a drug target and that conclusions require experimental validation before guiding care — the list of things it must not say is written into the paper itself.
So what do humans still answer for?
This section is the page's conclusion — read it slowest.
What this page proves: part of oversight can be engineered. Handoffs can be structured files, progress can be a state machine, judgment can be a logged trail, integrity can be rules locked before work begins — and even the watching itself can be delegated to an independent, clinically-minded AI session whose audits are all on record.
What this page does not and cannot prove: that humans can leave. At least three responsibilities do not transfer —
| Human responsibility | Why machines can't take it | The corresponding fact in this case |
|---|---|---|
| Reviewing the study as a whole: is this question worth asking, is the direction right, is the outcome acceptable | Case 7 supplies the counter-evidence: across three days the AIs raised zero direction-level challenges on their own. Value judgment has no reliable automated stand-in | This study's direction happening to work out is a result, not a guarantee; go / no-go and accept / reject still get decided by a person |
| Reviewing the final results: the last pass over key numbers, figures and claims before submission under a name | However strict Role A's audit is, it is an internal consistency check; the review that carries liability must be done by whoever carries it | The manuscript is a deliverable handed to a human for acceptance — not an endpoint that routes around one |
| Experimental validation: the last mile from computational conclusion to scientific fact | MR is causal inference over public data; however robust, it is computational evidence. Wet-lab and clinical validation live in the physical world, out of AI's reach | The manuscript's own wording ceiling says it: experimental validation (RCTs) required before any clinical use |
Put the two cases side by side and the conclusion completes itself: Case 7 proves oversight cannot be skipped; this page proves the skippable parts of oversight can be turned into machinery. The line between the two pages is where humans actually stand in AI-driven research today — framing the question, accepting the result, validating experimentally, signing responsibly. Execution and process discipline go to the platform.
FAQ
Does this mean research no longer needs people?
No — this page argues the opposite. What the case demonstrates is that part of oversight can be engineered: handoffs, progress, decision trails and integrity rules can all become machinery, and even process review can be delegated to an independent reviewer AI. But reviewing whether the study is worth doing, the final review of the results, experimental validation, and signing one's name to the conclusions remain human. In Case 7 the AIs went three days without once challenging the direction on their own — that is the counter-evidence marking the boundary.
Why should anyone trust results produced "with no human involved"?
Not on trust — on traceability and public data. Every dataset is public GWAS summary statistics with a download script, so anyone can recompute; all 15 decision-log entries are inspectable and resolved; thresholds and conclusion rules were preregistered and never touched afterwards; the negative result went into the manuscript instead of a drawer. The credibility comes from every step being checkable by outsiders — a stronger guarantee than "someone was watching", and the reason we productised these mechanisms.
Other teams have built "AI scientists" — what is different here?
Fully automated research pipelines are not unique to this case — several "AI Scientist"-type systems exist, and public evaluations have documented their typical failures: fabricated data, p-hacking, dressed-up results. This case differs in three places: integrity is preregistered as locked-before-work rules rather than hoped for afterwards; oversight is performed by an independent, clinically-minded AI session with a complete audit trail; and the study is real biomedical causal inference whose centrepiece is an honestly reported negative result. We have not seen a comparable fully public run record — but we do not claim a "first" on that basis; readers can judge.
How does this page relate to Case 7?
They are the two ends of a mirror. Case 7: 17 human interventions, direction finally vetoed by the human — it answers "why people are indispensable". This page: zero human interventions, study delivered — it answers "what the human role converges to". The causal thread matters more: Case 7's post-mortem lesson, "principles must be written as triggers", is exactly what this case's four mechanisms are in factory form. The gaps a failure exposed became the machinery of a success.
Last updated: · All process facts checked against the project's real files and logs (design document, TOPIC.yml, workflow_state.yml, findings.yml, analysis scripts and manuscript); scientific results in the Case 1 summary.