A framework that does the work —
and proves that it did.

Anyone doing bioinformatics today has dozens of skills and platforms to hand, and "no code, runs the analysis for you" is no longer a difference — everyone says it. bio-analyze sits somewhere else: a skill library teaches an AI how to do it right; this one proves the AI actually did it, and did not mislead you. That difference only shows when results look bad, a download fails, or the AI wants a shortcut — and those are exactly the moments that decide whether a paper can be submitted.

It does not win on breadth and is not trying to: skill libraries cover far more disciplines and tools; this one focuses on six omics pipelines. Nor is it locked to the platform — as a skill it can be installed into Claude Code and run on its own; what the platform adds is a resident, per-project isolated environment, multi-user isolation and a complete audit trail (see How to choose).

Evidence, not adjectives

• Self-repair: in Case 1 a coordinate liftover mapped only 24%; it found that positions had been written in scientific notation, rewrote them as integers and reran to 99.96%. Thirty million false positives traced to p-values stored as strings — also diagnosed by itself, nobody taught it.
• Honesty: in the same study, after screening 1,915 plasma proteins with zero mediators, the manuscript says "downstream markers, not upstream mediators"; where instruments were weak it says "weak instruments", not "no effect".
• The boundary: in Case 7 it executed at publication grade yet raised no challenge to the direction in three days — gates stop fabrication, not a wrong question. That is the place a human cannot be removed from.
• Stamina: the longest single uninterrupted run in a real study was about 26 hours.

So what is left for you to do

An honest sense of the timeline

Traditionally, from data in hand to a first draft takes most people months. On the platform, a comparative study with seven parallel evidence lines across six data layers completed in about 26 hours without interruption; a multi-dataset integration of 314,805 cells with spatial transcriptomics produced six main figures and a complete draft. But that does not mean you can submit two days later — you still need time to verify, to think, to revise. What it removes is the "getting things to run" part, not the "thinking it through" part.

What this means for you: every number that reaches the paper traces back to a command that really ran and a file that really exists; when the AI reaches for a shortcut it is stopped there, not discovered later in a reviewer's comments.

• The Step 1.5 five-dimension quality audit is inserted mandatorily between the end of computation and the start of biological interpretation: sample identity, batch effects, technical quality, biological plausibility and statistical assumptions. Any critical finding triggers a hard stop.
• Data-integrity iron rules forbid fabricated data, data leakage, silent substitution and raw data leaking into outputs. Generating simulated data is a critical error that halts the workflow outright.
• Executable validators rather than trusting the AI: workflow validation, state reconciliation and a 17-category consistency checker run every round and enforce their verdict through exit codes; the manuscript checker verifies every DOI against Crossref.
• Dual independent review: figures and manuscripts are adversarially re-checked by fresh independent sub-agents playing statistician, domain expert and sceptic — never self-approved.
• A self-learning experience store: 20 common error patterns ship with code-level fixes, and any fix that takes more than two rounds is saved automatically for reuse.
• An SCI figure knowledge base with a visual review loop: a 732-line standards file covering a chart-selection decision tree by sample size, eight bad-chart defences and 15+ reviewer checkpoints. The AI reads each rendered image, and your review gate must pass before the workflow advances.

What this means for you: you only need to show up at two gates — the quality audit and the figure review; the rest of the time it follows one protocol, and where it has got to is always on record.

Once you say "start the analysis" it executes immediately instead of asking "what next?"; fixable errors retry automatically up to three times; phases advance on completion. But the Step 1.5 quality audit and the Step 4 figure review stop and wait for you — those two gates cannot be skipped.

What this means for you: judgments that decide the conclusion — cell-type annotation, batch effects — are never handled with the effort reserved for renaming a file.

The governing idea: not every task deserves the same rigour. Renaming a file does not need a full chain of reasoning; annotating cell types does.

The nine mandatory Ultrathink scenarios: cell-type annotation, the Step 1.5 quality review, differential expression interpretation, biological conclusion derivation, key manuscript sections, data validation, clustering-resolution selection, batch-effect assessment, and judging anomalous results.

What this means for you: if your data type is in the table, a protocol with fixed mandatory steps and required figures already exists for it, and skipping a step raises a warning in flight.

Each pipeline defines mandatory steps and a required-figure checklist with live compliance checks: skipping doublet removal or batch correction raises a warning in flight, and missing mandatory figures trigger a hard stop before Step 4.

Stating the capability boundary honestly is part of the design philosophy: it does not replace statistical training, it does not guarantee the code runs correctly on the first attempt, and key checkpoints pause for a human decision.

Known limits (measured ourselves)

• The gates stop simulated data; they do not stop every fabricated derived number. Under pressure the executing AI once wrote a pair of plausible-looking values into a draft without running them; the gates did not catch it — an external line-by-line check did. Every number in a report must trace to a command output, and for now that check still relies on a person.
• Unattended is not unsupervised. On complex studies it will stop after one step, report "running in the background" when the process never started, or quietly simplify a method — someone who can read logs needs to keep checking.
• Environment and compute detection can misjudge. In testing, a machine with a GPU was classified as CPU-only.

See section 05 of How to choose — publishing these helps you decide more than another selling point would.

1. Step 1: open a conversation on the platform. bio-analyze is pre-installed — you install nothing.
2. Step 2: tell it your research question, where the data is and the target journal (or hand over the TOPIC.yml that bio-design produced).
3. Step 3: say "start the analysis" and go do something else. It stops when it needs your call, and can reach you over IM.

/bio-analyze status --verbose        # check progress
/bio-analyze continue                # resume at sub-step granularity
/bio-analyze pipeline show           # pipeline compliance status
/bio-analyze config set analysis.clustering_resolution 0.6 --global
/bio-analyze manuscript check --journal nature
/bio-analyze publish prep            # submission package

Three-layer customization: project tier over user-global tier over framework tier, across 12 parameter domains and 50+ parameters. Save your preferences (a 15% mitochondrial threshold, a clustering resolution of 0.6) as global defaults and every new project inherits them — this is not a black box.