What happens to your reviews?
Every accepted review is stored as a structured observation in our research database and a separate, hash-checked contribution archive. Completing a paid entry is separate from research quality: payment never makes an answer scientifically correct.
These are crowd-reviewed observations—not medically verified cells. We do not diagnose patients, discover treatments, or claim that a laboratory currently uses the results. No active laboratory partner is involved.
- Record: shape, relative size, uncertainty, confidence and image-quality concerns, linked to the source image and highlighted nucleus.
- Screen: hidden reference questions estimate observer reliability. Ambiguous answers remain useful and never disqualify someone from Sunday Pick.
- Aggregate: combine eligible observations to measure agreement, uncertainty and differences from source reference labels.
- Preserve and share: researchers with approved access can export current results. Public downloads are versioned snapshots of frozen datasets, released separately to avoid revealing answers to active players.
Public research downloads
No wallet, payment or login is required for published downloads. Each release includes a SHA-256 checksum, source provenance, image identifiers, target coordinates and aggregate observations.
Checking published releases…
Empty downloads are not fabricated. A completed ten-review game does not mean a research target or entire dataset is ready for public release.
Controlled access for researchers
Checking researcher access…
Authorized researchers receive an access token privately. Do not post that token or put it in a URL.
- Aggregate JSON
GET /api/research/aggregates?limit=100&offset=0with theX-Researcher-Tokenheader. Followpage.next_offsetuntil it is null—even when a page contains no eligible items.- Detailed pseudonymous JSON
GET /api/research/export?limit=1000&offset=0with the same header. This controlled export contains per-observation records, quality snapshots and integrity hashes. Calibration records are explicitly flagged. It must not be treated as an anonymous public dataset.
Access must be arranged with the project operator, including the intended analysis and safeguards for pseudonymous data. Published snapshots remain freely downloadable without this process.
Methodology and data dictionary
Public and controlled aggregates require at least three currently calibrated accounts for a target. Calibration requires at least eight scored answers, at least three examples per broad reference class, smoothed balanced accuracy of 0.70 or more, and a 95% Wilson lower bound of 0.55 or more. These are quality screens, not clinical validation.
Accounts are not verified unique people. Multiple wallets are allowed in the prize program. Correlated answers, repeat participants, incentives and self-selection can bias the data. Researchers must account for these limitations; a wallet count is not a count of independent human observers.
| Fields | Meaning |
|---|---|
| task_id, dataset_id, image_sha256, target | Reproducible link to the source image and highlighted target. Image links may be relative to this website. |
| source_url, license, reference_label | Dataset provenance, source licensing and original reference label—not a new diagnosis. Player-derived observations need their own release terms before external redistribution. |
| votes, observed_shapes, observed_relative_sizes | Counts of observations from currently calibrated accounts, not a verified cell identity. |
| uncertain_rate, split_disagreement, reference_disagreement | Fractions describing uncertainty and disagreement. Disagreement does not prove the reference wrong. |
| quality_flags | Counts for blur, marker misalignment, cutoff/artifacts and unusual appearance. |
| difficulty_score, priority | Heuristic prioritization: 0.35 × uncertainty + 0.30 × vote split + 0.25 × reference disagreement + 0.10 × capped flag rate. |
| consensus_status, consensus_probability_tumor | Model-derived aggregation, not a calibrated clinical probability. “Expert-review candidate” does not mean an expert has reviewed it. |
Public aggregates exclude player identifiers, wallet/payment details, session credentials and hidden calibration answers. A target is retired after strong consensus from at least seven calibrated accounts, or marked for potential expert review after fifteen without strong consensus. Fully completed batches are frozen before public publication.
Potential uses include studying non-expert uncertainty, identifying image-quality concerns and choosing difficult examples for future studies. No research benefit or medical discovery is guaranteed. Formal studies need their own protocol, ethics determination where applicable, and expert validation.