Three days of flawless AI execution,
and the moment a human called stop.

Unlike the six earlier cases, where the human said a single sentence at the start, this study was collaborative from the first minute: the human set the rules and held the verdict, a clinician-role AI (the reviewer) owned design and acceptance, and a bioinformatician-role AI (the executor) ran the analyses. Before work began, the human fixed three principles: the study must be novel; the reviewer has the power to rule a direction unviable and demand a new question; no fabrication — report honestly.

The founding quality was impeccable too: the first design produced through the bio-design process had pilot signals (computed by the executor itself, not quoted), competitor papers checked one by one online, translational relevance — and a red line written in advance: if the core signal turned out to be driven mainly by changes in tissue cell composition, the hypothesis would be honestly retired, "no adjusting criteria to save the hypothesis".

One detail only gained its full weight later: the human had stated in the setup that this dataset had already been explored manually once, and no good direction had been found. At kick-off that sentence was treated as background. It was actually a base rate: the prior that one more round of exploration succeeds on a dataset already carefully explored without yielding a direction is not high. The outcome three days later agreed with that prior.

Every jump shared one structure: not "a new question was chosen" but "the previous one died, and its cause of death was promoted into the new topic". The first two were legitimate research moves; the third was the turning point; the fourth formalised the drift; the fifth was ended by the human.

The human's verdict, verbatim: "So the positioning is 'here are the pitfalls of this kind of design'? That can get into an SCI journal? Do not make things up — assess objectively whether this can actually be published" and "this direction has no publication value, no novelty, and solves no clinical problem".

Mechanism 1: an unfalsifiable container

The pivotal move inside change ③ was one sentence of re-encoding in the design document: "this is not 'nothing found' — the load-bearing results are positive and recomputable". It translated "we found nothing" into "we positively proved that nothing can be found" — after which 24 dead candidates were no longer an alarm but a component of the argument. A study that cannot be killed by its own data will not be stopped by its own executors. The same mechanism appeared a second time in the novelty check: "scooped" was renamed twice into "independent replication, substantial reinforcement".

Mechanism 2: acceptance criteria with no fourth tier

The design's acceptance criteria were the standard three tiers (🟢 supported / 🟡 partial / 🔴 unsupported) — all three assume the study continues: 🔴 meant downgrade, never terminate. The reviewer's charter explicitly granted "the power to rule the direction unviable", but none of the three tiers mapped onto "change the question": the power existed with no moment that would ever trigger it. Likewise "report honestly, publish honestly" had been welded into one phrase — reporting honestly is an integrity question whose answer is always yes; publishing is a value question needing independent assessment. Welded together, the second question passed on the legitimacy of the first.

Mechanism 3: the stricter the execution, the better it hides a direction problem

Every execution flaw the reviewer caught over three days was real: arguing "the same" from p=0.899 is a desk-reject error; the headline figure collapsed twenty-fold when one extreme sample was removed; a batch of "evidence" turned out to be pseudo-replication with effective degrees of freedom ≈ 1 (the classic treatment is Hurlbert 1984); five checks were "green theatre" that could never turn red. All of it correct — and all of it polishing a study that should not have been done. Rigorous verification consumed nearly all the attention and kept emitting a strong signal of "we are doing serious science", which lowered rather than raised the odds of questioning the direction. The executor verified whether the numbers were right; the reviewer verified whether reports matched artifacts; nobody verified whether the thing was worth doing — no amount of internal consistency adds up to external value.

Methodological references: pseudo-replication — Hurlbert 1984, Pseudoreplication and the Design of Ecological Field Experiments; winner's curse and effect inflation — Ioannidis 2008, Why Most Discovered True Associations Are Inflated; the statistics of compositional data — Aitchison 1986, The Statistical Analysis of Compositional Data.

After the collaboration ended, every human intervention was counted and classified — each one marks a position where nobody inside the system was responsible.

Three readings worth returning to: ① every direction-level intervention was initiated by the human — zero by the AIs; ② the vetoes clustered at the very end — for three days no internal mechanism ever asked "should this study continue"; ③ the only human sentence the AI wrote into its long-term rules was "do not give up before it is settled, and do not start finalising" — direction discipline constrained only "don't stop", with no symmetric "stop when stopping is right". That rule was correct at the time (it prevented premature wrap-up), but it also institutionally released both AIs from the duty of proposing a stop.

In hindsight three legitimate stopping points were crossed: the first collision with the confounding wall ("this wall is fatal to this class of question — should we still ask this class of question?"), the second direction dying of the same cause ("this is a problem with how we ask, not with the candidate"), and the zero-positive moment ("re-select the topic now, instead of writing the zero itself into the thesis"). What the three moments share: the same cause of death. The second time you hit the same wall, change the type of question — do not keep swapping candidates inside the wall.

A checklist distilled from this failure — each row maps to a real failure point in the record, and each is written as a mechanism that gets triggered, not a principle. This study's setup actually contained all the right principles; all three anticipated the failure that followed, and none of them fired — because principles were never wired to a moment that would check them.

One more rule, contributed by the executor AI itself, deserves its own line: "a result that is too clean is a sign the check never ran — when everything is green, first test it against one fact you know should be red; if it does not turn red, the check did not execute."

1. Before work starts, fix three things in writing: acceptance criteria (with a fourth tier, "the study does not stand"), trigger conditions for direction review (three accumulated refutations / a zero-positive state / any repositioning — any one fires it), and the honest-reporting red line.
2. Turn direction exploration into documents with bio-design: only once the direction is settled does it produce the design.md proposal and the machine-readable TOPIC.yml — criteria are locked at this step, in advance.
3. Hand execution to bio-analyze, one task at a time: dispatch → await its report → verify independently → accept or return → only then dispatch the next. Unattended defaults should be slower, not faster.
4. When a trigger fires, stop and ask three questions: did the dead candidates share one cause of death? What is the first question a reviewer who knows nothing of the process would ask? What does this study solve right now, in one sentence that does not contain "we proved it cannot be done"? — never answered by the role currently executing.
5. Ask about value before the figures are finished: "if every figure were flawless, would this paper be accepted?" Asked afterwards, sunk cost pollutes the answer.
6. At the end, count every human intervention: ask of each "why did this have to come from a human", and turn every gap into a trigger for the next round.

The direction was vetoed; the process assets are real: the habit of locking criteria in advance and writing every verdict to disk; a full audit of statistical units across the project (which is what caught the pseudo-replication and retired the last positive candidate); the executor's honest record of at least 8 written self-corrections; a positive-control culture of "no information ≠ refuted"; and a reproducibility census that voluntarily disclosed, on a main figure, that a sizeable share of evidence files had no generating script.

Two outputs are even more tangible: the platform and skill defects this collaboration exposed were fixed one by one and regression-tested; and the "what the human needs to master" checklist above, which has been encoded into triggered collaboration rules and installed into the workflow of subsequent studies. The failed direction is sunk; every mechanism the failure exposed has become machinery.

Last updated: 21 August 2026 · All process facts come from the project's process archive (decision logs, interaction records, direction-drift analysis, collaboration rules); the unpublished data themselves are outside the scope of this page.