Skip to main content

Track Record · Public Ledger

We publish our predictions before the results exist.

Fresh videos, scored while nobody knows how they will perform. Every prediction is locked into this ledger, then measured against real 7 and 30 day results. Misses included.

344

Predictions logged

277

Outcomes measured

Jul 5, 2026

Logging since

Accuracy

No accuracy number gets quoted before the sample carries it. Live v2 predictions are accumulating in the ledger below; the accuracy figure unlocks once 50 of them have completed measurement.

0 of 50 live v2 outcomes measured so far. Every measured outcome sharpens the estimate, and every methodology change is versioned and public.

Calibration

Calibration statistics unlock once 50 predictions under the current methodology have completed their 30-day measurement window. No numbers get quoted before the sample carries them.

First 30-day results land on Aug 25, 2026. Every prediction below is already locked.

The ledger

Grouped by weekly cohort. Newest cohorts show every prediction; older cohorts are collapsed into their totals.

Rows from the v1 era additionally show a v2 score — the current method re-run over the inputs archived at prediction time, computed after the outcomes were already known and labeled as such. The original v1 score stays visible on every row.

Earlier cohorts

  • 2026-W30 · committed Jul 21, 202650 predictions
  • 2026-W29 · committed Jul 14, 202643 predictions
  • 2026-W28 · committed Jul 7, 202649 predictions
  • 2026-QA1 · committed Jul 5, 20268 predictions

How this ledger works

Committed before scoring.

Each week the cohort — niches, time window, and search queries — is frozen into the ledger before any video is scored. The selection rules are stored with the cohort, so the batch is reproducible and cherry-picking is structurally impossible.

Fresh videos, picked by date.

Candidates are less than 72 hours old and selected by publish date, never by views. Predicting videos that are already trending would prove nothing; predicting before the outcome exists is the point.

Scored from the video's content, in its own language.

Since methodology v2, system predictions read what the video actually says: the spoken transcript when one exists, otherwise the video file itself — plus title and tags — through the same scoring engine users get, prompted in the video's declared language. Each row records the exact model, prompt version, and input source used, so the series stays comparable over time.

Measured against the channel's own baseline.

After 7 and 30 days we record real views and compute the multiplier: views gained since the prediction was logged, divided by the median views of that channel's recent uploads. Views a video already had at prediction time are subtracted first, so a head start cannot count as predicted growth. Channels with too little history get no multiplier rather than a misleading one.

Methodology versions are public.

v1 (through 2026-08-08) scored from public metadata with an English-first prompt. v2 scores transcript or video content in the video's own language and measures net view growth. v1 rows keep their original frozen scores forever; where a retrospective v2 score exists it is shown alongside, clearly labeled, and the two versions are never pooled into one statistic. Retrospective v2 scores predate two live-v2 features — they ran with an en/US prompt locale and without the language policy — so they are stamped as their own variant and never presented as identical to live v2.

Nothing gets removed.

Outcomes are written next to the prediction — hits and misses alike. Deleted or private videos are flagged as unavailable, not dropped. The database itself rejects edits to logged predictions.

Scores are pattern-based estimates, not guarantees. How the score itself works

If you want the reasoning rather than the ledger, we wrote up what correlates with a breakout, and how that differs from the public stats a tracker can show you after the fact or what the established YouTube toolkits measure.

Questions

How does the PreViral track record work?

Every week PreViral commits a batch of freshly published YouTube videos, selected by publish date only, before anyone knows how they will perform. Each video gets a score from the same engine our users see, the prediction is locked into a public ledger, and after 7 and 30 days we record the real view counts next to it. Nothing is removed afterwards, including misses.

What does the channel multiplier mean?

The multiplier compares the views a video gained after the prediction was logged to the median views of that channel's own recent uploads. Views the video already had at prediction time are subtracted first, so a head start cannot masquerade as predicted growth. A 3x after 30 days means the video gained three times a normal upload's views for that channel after we called it.

Why do some entries say pending or unavailable?

Predictions are logged the moment they are made, so fresh entries wait for their 7 and 30 day measurement dates. If a video is deleted or set to private before measurement, we flag it as unavailable instead of quietly dropping it from the ledger.

Are the results cherry-picked?

No. The selection rules for each weekly batch, including the niches, the time window, and the search queries, are frozen in the ledger before scoring starts. Videos are picked by publish date, not by views, and every prediction stays visible whether it hit or missed.

Page methodology last updated 2026-08-08.