For researchers, clinicians and engineers
Tools
Small, open building blocks, each with its license, what it does and what it does not.
Three public repositories
-
cf-lab
The lab's shared instructions for AI agents, its skills (short guides an agent can load) and the security rules. Also an open task board, where every task records who proposed it, the evidence for it and how to check it. A model may propose a task. Only a person moves it forward.
-
cf-skills
Holds cf-evidence-loop, a command-line tool that answers a research question as an evidence record: the source, the date, the exact fields and the limits, with no summary in between. It is described below.
-
cf-research
Research notes; three guides to reading patient-registry reports and their genetics sections; a catalog of what each source permits; and small tools that check our own writing. See the evidence page.
Six interactive research workbenches
From single amino acid channel coordinates to continental registry cohorts, all running locally in your browser.
Four Cryo-EM structures (6MSM, 5UAK, 8EIQ, 8EJ1) with variant mapping, domain topology, and distance measurements.
Tri-modal rescue kinetics for G542X and W1282X: SMG1i NMD transcript salvage, suppressor-tRNA, and ERAD escape.
5-axis biophysical feasibility radar (BFI), mucus penetration, and non-destructive basal progenitor dedifferentiation.
PRISMA 2020 screening flow and surrogate-to-registrational phase shifts across 1,821 primary outcome measures.
Longitudinal 57,288 PwCF European surveillance alongside the ~6–12% multi-registry unserved denominator.
Interactive six-thread constellation map with 19 verified evidence nodes and four guided Pathfinder Quests.
The CFTR model, in the browser
Four published structures of the CFTR protein in one viewer. Look up a residue (one amino acid position in the chain) or measure a distance. UniProt, a public protein database, records which parts of the protein sit inside, across and outside the cell membrane (its topology) and where natural variants fall. Lay either or both over the structure. Save a picture, a CSV or a link. What you type or choose stays in your browser.
In silico delivery explorer
Explore biophysical transport feasibility, receptor transcytosis kinetics (FcRn, TfR1, pIgR), and 180-day epithelial lineage dedifferentiation trajectories to evaluate permanent stem cell delivery without tight junction disruption. Models run deterministically in your browser against cryptographically locked datasets.
Trial endpoint atlas showcase
Explore PRISMA 2020 trial screening flows, surrogate-to-registrational phase transitions, and primary outcome measure quotations across 940 included interventional cystic fibrosis clinical trials (screened from 1,053 registry records). Zero-build native ES modules with client-side faceted filtering.
Open the interactive trial endpoint explorer · Open the 23-class static endpoint census (/atlas/)
Interactive 66-Model Cross-Discipline Discovery Catalog & Curative Bottleneck Synthesis Explorer
In plain words: Discovering a permanent, curative therapy for Cystic Fibrosis requires solving a chain of biological and physical bottlenecks. It is not enough for a gene editing tool or mRNA construct to work brilliantly in a petri dish cell culture (≥12% WT rescue) if inhaled delivery particles cannot penetrate the thick, dehydrated CF airway mucus (a 317-fold diffusion barrier) or if treated cells are naturally shed and replaced every few months without reaching long-lived basal stem cells. This interactive catalog brings together 52 open-source biophysical models across six scientific disciplines—spanning structural protein physics, RNA and nonsense readthrough, aerosol delivery mechanics, phage microbiology, clinical trials, and multi-agent AI verification. Explore the mathematical equations, verified test suites, and PDB coordinate references grounding each discovery milestone. This is research, not medical advice.
Comprehensive cross-discipline catalog of 66 verified biophysical models registered in tools/discovery/catalog.json, featuring active triage status, governing equations, PDB/DOI coordinate grounding, and Engine C chained bottleneck synthesis radar.
Curative Bottleneck Synthesis Radar (Engine C)
Contrasting native in vitro molecular rescue against the multi-barrier pulmonary delivery gauntlet:
Cross-Discipline Grounding & Citation Provenance
Grounded in cf-research/tools/discovery/catalog.json (52 registered models) and discovery_council.py.
Suk JS et al. Nanoparticle transport in mucus. DOI: 10.1016/j.addr.2015.09.006 (317x diffusion collapse).
Danis-Wlodarczyk K et al. Alginate Lyase & Phage Biofilm Dispersal. DOI: 10.1038/s41591-025-03678-8.
Atomic structural coordinates grounded in PDB 6MSM (human CFTR), 6QW8, 6O2P, 8EIQ, 5OYB (TMEM16A), 5UAK, and 2PZE.
Preclinical in-silico biophysical modeling; research, not medical advice. No clinical efficacy or therapeutic safety claim is made.
The evidence tool: its questions and commands
A researcher does not begin with an answer. They walk a fixed set of questions against the public record. The tool does the same, one public source at a time, and every answer is a record you can check. A drug label's indication text, for example (the section that says which condition, and in which patients, a drug is approved to treat), is a public document, not any person's eligibility.
| Question | Command | Source |
|---|---|---|
| What has been published? | pubmed, reviews, epmc | PubMed, Europe PMC |
| Was it retracted? How often is it cited? | retractions, cited, doi | Crossref, Europe PMC |
| When was a drug approved, and what does its label say? | approval, label | openFDA |
| Which trials list this condition? | trials, trial | ClinicalTrials.gov |
| What is this preprint? | preprint, preprints | bioRxiv, medRxiv |
| How is this variant classified? | variants | ClinVar |
| Who does NIH fund on this topic? | grants | NIH RePORTER |
| What do several sources say about one drug, variant or trial? | brief | The sources above, cross-checked |
| Does every paper id (DOI or PMID) and trial id in a note exist? | cite-check | Crossref, PubMed, ClinicalTrials.gov |
| What may the tool reach at all? | sources | The catalog |
Proof Receipt · Source Catalog & cf-evidence-loop Conduct Lock (SHA-256 verified)
Lockfile: data/evidence.lock.json · sources/catalog.yaml SHA-256: 63eab6bb789bfb8096c6bc456e8809e6c79c7010ab83439cc933c77505a02cde
Deterministic guardrails: cf-evidence-loop enforces network rules R1–R11 against the single source catalog before opening any connection.
Verify on GitHub: cf-skills/cf-evidence-loop · cf-research/sources/catalog.yaml · data/evidence.lock.json
How the tools treat other people's servers
Nothing here relies on a model being considerate. Every rule is code with a test, and the full text is in the skills repo (.claude/skills/cf-evidence-loop/references/NETWORK-RULES.md) so other skills can start from it. In plain words:
R1 · The catalog decides
A source that is not in the catalog, or is marked for a person to download, to request or as forbidden, is refused before any connection opens.
R2 · APIs need their terms
An API, a publisher's interface for programs, is called only if its usage policy and a request ceiling are recorded.
R3 · robots.txt is obeyed
For documents, the site's robots.txt (the file in which a site says what automated readers may fetch) is read once and followed. The group that names this tool applies, else the
*group. A robots.txt that cannot be read means no fetch, except a plain 404 or 410, which means no restrictions.R4 · Slowly
One request at a time, never faster than the source allows, never more than three a second, one a second by default.
R5 · A budget
A hard ceiling of 200 requests in one run, which can be lowered but never raised.
R6 · A circuit breaker
A host is left alone for the rest of the run after a refusal, a second rate-limit reply (a request to slow down), or three errors in a row.
R7 · One honest identity
A fixed project user agent (the name a program sends with each request) that says what the tool is. There is no way to pretend to be a browser.
R8 · Everything counted
Every request attempt is counted and can be logged.
R9 · A dry run
Any run can list every request it would make, robots.txt included, without making one.
R10 · A size cap
A response larger than 5 MB is refused, not read to the end.
R11 · A plain contact
The optional contact address in the user agent must be a plain email address.
Visual Sitemap & Biological Scale Constellation
A structured map of all 16 public pages, deep-dive atlases, and interactive workbenches across the Sensei Ewok CF Lab, ordered across 6 biological scales and 4 research pillars.
01 · CF & Community
02 · Biophysics & Trials
- 3D CFTR Channel Viewer Cryo-EM structures (6MSM/8EIQ)
- Nonsense PTC Readthrough Sim G542X & W1282X readthrough
- Airway LNP Delivery Explorer 150 nm mucus mesh & BFI radar
- 940-Trial Endpoint Explorer PRISMA 2020 flow & phase transitions
- 10-Yr ECFSPR & 5-Nation Cascade 57,288 European patients
- Trial Atlas: 33-Yr Matrix 25 classes across 33 years
- Endpoint Co-Occurrence Network Correlated measure pairs
03 · Evidence & AI Loop
What we use
A list of tools, with no claim about what any vendor does for us. We use Visual Studio Code, Git and GitHub, Python, PowerShell, Ollama for running models on our own machine, Playwright and Microsoft Edge for page observations and tests, and AI coding assistants including GitHub Copilot and Anthropic's Claude models. The site itself is hand-written HTML, CSS and a few small scripts, with no framework.
Sensei Ewok's CF Lab is one person's project, built on the maintainer's own time, hardware and paid accounts. No company has funded, reviewed or endorsed it, and the tool credits above are thanks, not recommendations. Product names belong to their owners.