Automated, immutable and auditable.
No human edits, overrides, or hand-picks the model's output. The scan and ranking run end to end without human input. A human only steps in if an automated data-integrity check flags a problem with the input data - in that case they fix the input and the same pipeline re-runs from scratch; they never touch the ranking logic or the result itself. Once a run is clean, results export directly to a public GitHub repository the website reads from - every data point carries an immutable commit timestamp, independently verifiable by anyone.
How integrity is enforced
Each mechanism below closes a specific way the process could otherwise be gamed - from re-running a bad week to editing a result after seeing how it performed. Together they make the claims above structural, not just stated.
Scheduled, unattended
The scan and ML ranking run on a fixed weekly schedule with no human input into the model or its output. Publication is gated by an automated data-integrity check; if it flags a problem, a human corrects the upstream input and the same pipeline re-runs - nobody edits, reorders, or overrides a result.
No partials
Each week's output is either fully published or not published at all. If the pipeline fails partway through, the previous week's data stays intact - there is no in-between state.
Locked, then revealed
Candidates are encrypted and committed the moment a scan completes - proof of timing without revealing them to free users yet. The key is published alongside the plaintext 12 weeks later, so anyone can verify nothing changed in between.
Outcome lockout
Outcome files are written exactly once at 12 weeks. The system has no modification path after initial write. What the model ranked at scan time is permanently what the archive shows.
Corrected, not hidden
If an error is ever found in already-published data, the original file stays exactly as committed - it is never edited or removed. A separate, clearly-labeled correction is published alongside it instead, with a plain explanation of what was wrong. That has to hold even when the mistake is ours.
Verify it yourself
The public data repository contains every exported file with its full commit history. Inspect when any scan was published, compare any two exports, and confirm no file was modified after its initial commit.