This report does not measure accuracy
July has no human-confirmed ground-truth labels. A returned result means the production system persisted a non-Unknown movie candidate. Result return and provider agreement are operational signals, not proof that a title was correct.
A useful production report should show regressions as clearly as gains. In July, the persisted cohort contained far more fresh direct-finder attempts than in June. The median response became 299 milliseconds faster, yet the share of attempts returning a candidate declined by 9.68 percentage points.
Both facts can be true. July's input mix changed, but mix explains little of the decline: applying June's category return rates to July's category volumes produces an expected overall rate of 48.61%, only 0.30 percentage points below June's 48.91%. The remaining 9.38-point gap reflects lower return rates within July's input groups, not merely a shift toward harder input types.
The recorded cohort grew, but returned less often
| Measure | June | July | Change |
|---|---|---|---|
| Eligible attempts | 9,713 | 17,485 | +80.0% |
| Returned candidates | 4,751 | 6,860 | +44.4% |
| Result-return rate | 48.91% | 39.23% | -9.68 pp |
| Median latency | 3.80 s | 3.50 s | -0.30 s |
| 90th percentile latency | 7.48 s | 7.09 s | -0.39 s |
| Recorded agreement | 51.02% | 39.33% | -11.69 pp |
The absolute number of returned candidates increased by 2,109 even as the return rate fell. Latency also improved at the median and 90th percentile. Recorded volume, speed, and result return therefore moved in different directions; combining them into one performance score would hide the trade-off. The 80% month-over-month increase describes persisted eligible rows and should not be interpreted as growth in users or demand because telemetry coverage and persistence can also change.
Results by input type
| Input | Attempts | Returned | Unknown | July | June | P50 | P90 |
|---|---|---|---|---|---|---|---|
| Text description | 5,813 | 3,626 | 2,187 | 62.38% | 69.94% | 1.88 s | 3.71 s |
| Image upload | 4,635 | 2,023 | 2,612 | 43.65% | 47.20% | 4.80 s | 11.34 s |
| Instagram URL | 2,252 | 1 | 2,251 | 0.04% | 35.24% | 2.96 s | 4.48 s |
| YouTube URL | 2,017 | 1,162 | 855 | 57.61% | 67.69% | 4.01 s | 7.78 s |
| Facebook URL | 1,329 | 2 | 1,327 | 0.15% | 0.00% | 5.57 s | 8.43 s |
| TikTok URL | 697 | 2 | 695 | 0.29% | 0.66% | 4.20 s | 5.82 s |
| Other URL | 456 | 25 | 431 | 5.48% | 11.23% | 1.87 s | 3.44 s |
| X / Twitter URL | 286 | 19 | 267 | 6.64% | 10.86% | 2.76 s | 5.28 s |
| All eligible runs | 17,485 | 6,860 | 10,625 | 39.23% | 48.91% | 3.50 s | 7.09 s |
A product boundary changed late July 30 UTC
Beginning late July 30 UTC, VidScio deployed a flow that requests a screenshot before model inference when Instagram, Facebook, or TikTok exposes no inspectable frame, transcript, or useful public metadata. Those recovery prompts do not create terminal identification rows and are absent from this cohort.
July's social URL rates therefore describe the earlier URL-only behavior for nearly the entire month. They should not be projected onto the current screenshot-assisted flow. The change deliberately favors having evidence over producing a speculative title.
Operational telemetry, not model rankings
Provider telemetry was linked to 17,162 of 17,485 eligible attempts, or 98.15%. Those attempts generated 58,275 provider-call records because the production pipeline can consult several models for one request. This fan-out count describes orchestration activity, not identification quality.
Provider-by-provider candidate rates are not published as a leaderboard here. Routing was not randomized, providers received different input types and case difficulty, and July has no human-confirmed correctness labels. The historical rows also predate the current completed-snapshot marker, so their presence does not prove that every scheduled provider outcome was persisted.
Recorded agreement also declined
Of July's 6,860 returned candidates, 2,698 recorded matching normalized titles from at least two providers: 39.33%. June recorded agreement for 51.02% of returned candidates. This can reflect input mix, provider routing, available evidence, or model behavior; the telemetry does not isolate a cause.
Agreement is not correctness. Several providers can converge on the same wrong title, and a correct result can emerge from a production path that does not record two-provider agreement.
Methodology
Study window
July 1, 2026 at 00:00 UTC through August 1, 2026 at 00:00 UTC. Aggregates were queried on August 14, 2026.
Cohort construction
The month contained 21,756 persisted direct streaming-finder rows. We excluded 4,271 cache hits, plus deleted rows and correction retries, leaving 17,485 fresh terminal model attempts. This is an incomplete observational telemetry cohort, not a count of all identifications or unique people. Not every product flow writes to this table, and a handled request appears only when telemetry persistence succeeds. The cohort uses the direct finder's webapp source marker and excludes chat and extension traffic.
VidScio's roughly 30,000 monthly active users are measured separately and are not the denominator for this report. One person can create multiple rows, while another can use the product without creating an eligible row here.
Result return
A returned result requires a successful terminal status and a persisted movie record. Unknown requires an Unknown terminal status and no movie record. All status and movie-record fields agreed in the cohort.
Input groups
Image upload takes precedence, followed by YouTube. Remaining URLs are grouped by recognized platform domains into Instagram, Facebook, TikTok, or X/Twitter. Other URLs must begin with an HTTP or HTTPS scheme. Text containing an embedded non-platform link remains a text description; this rule keeps the July categories consistent with the published June methodology.
Latency
Server elapsed time runs from request handling to the terminal result. P50 is the median and P90 is the value at or below which 90% of attempts completed. No eligible row lacked latency data.
Provider telemetry
Aggregate provider-call coverage describes how often orchestration telemetry was linked to eligible attempts. It is not an accuracy measure. Provider routing was not randomized, so unlike routes are not treated as head-to-head comparisons.
Limitations
- This is an incomplete observational telemetry cohort, not a random sample or a controlled test set.
- There are no human ground-truth labels for the July attempts.
- A returned title may be wrong; an Unknown may be the safest correct behavior.
- Social recovery began changing late July 30 UTC, creating a within-month product boundary.
- Provider routes receive different inputs and are not randomized comparisons.
- Provider telemetry is absent for 323 eligible attempts, and completeness of the remaining historical provider snapshots cannot be established from the current completion marker.
- The telemetry does not identify staff tests, repeated users, or duplicate inputs, and this analysis does not infer them.
Privacy statement
This report uses aggregate counts, status fields, routing labels, agreement flags, and latency percentiles. No user identifiers, prompts, URLs, uploaded images, movie titles, IP addresses, or country-level records were selected for publication. No raw user submission was reviewed to write the report.
What July changed
July did not produce a flattering single score. It produced something more useful: evidence that recorded attempt volume grew faster than returned candidates, that latency held up under that volume, and that URL-only social-video inference needed a stricter evidence boundary.
The next quality step is not to relabel candidate return as accuracy. It is to build a balanced, human-labeled evaluation set and measure correctness separately from availability, speed, agreement, and abstention.
Try the direct finder
Submit a scene description, supported public URL, or screenshot and verify any returned candidate against the source evidence.
Identify a movie