Progress

What has actually been done, in order, with its status. Work began on 17 November 2026. Results are not shown here until they have been checked and cleared for release.

  1. 17-18 Nov 2026

    Adaptive Biological Discovery (R-001): analysis workflow built and quality-checked

    A written protocol, a reproducible computational workflow on public perturbation screens, and negative controls were built. An internal quality review led to corrections, which were recorded openly, and a replication scope was drafted. No result is public yet.

    In progress
  2. 17-18 Nov 2026

    Perturbation prediction baselines checked for leakage

    Simple baseline predictors were compared with a published benchmark model under leakage checks. A reproduction defect and a script error were found, disclosed and fixed. Further model training is on hold by pre-set rule.

    In progress
  3. 17-18 Nov 2026

    Biomedical foundation-model evaluation: gates frozen, one error corrected

    A preregistered evaluation was frozen before inference. An earlier claim about possible training-data overlap was found to be wrong and was formally corrected. Further benchmarking is on hold.

    Held
  4. 17-18 Nov 2026

    Virtual Tissue Laboratory (R-002): simulation environment verified; biological validation blocked

    The simulation software environment was integrity-checked and its run times measured. An eligibility review concluded that the current model cannot yet be tested against a matching public experiment, so the main bottleneck is now biological testability, not software. This is recorded as blocked, not as a result.

    Blocked
  5. 17-18 Nov 2026

    Research software layer built and hardened, test providers only

    A private software layer for routing computational requests was built with automated tests and repeated security reviews. Ten hardening changes were merged into one integration candidate without conflicts, and an independent smoke test of the merged result passed. Five defects from the final review remain open. It uses simulated providers only and is not approved for real-provider or paid use, which would first require enforced upstream spending limits.

    Held
  6. 18 Nov 2026

    Independent AI-assisted review of every programme

    Reviewers (AI systems, not human peer reviewers) were asked to find reasons each programme is wrong, unoriginal or not worth continuing. They found substantial overlap with published work. The programmes were consolidated, several were put on hold, and a draft of a single combined study was rewritten after a hostile statistical review.

    Verified
  7. 18-22 Nov 2026

    Public website rebuilt around the evidence ledger

    The site now states the real status of the work, uses the approved identity, and labels its illustrations as conceptual. Public release is held until institutional and licensing checks are complete.

    Verified
  8. 22 Nov 2026

    First software release archived with a DOI

    perturb-audit v0.1.0, a prototype tool that checks whether the evaluation of a genetic-perturbation prediction release is valid for its declared task, was archived on Zenodo under an MIT licence. It is a prototype with stated limits, has not been peer reviewed, and overlaps existing tools.

    Verified

Dates are repository commit dates. The reviews described here were carried out by AI systems and are not human peer review. Holds and blocked items are listed because they are part of the record.