How the record is made — and the 14 conditions and counter-arguments it publishes about itself.

One pass, once a day, in five steps. Nothing is estimated on the way through — the only numbers that exist are the ones YouTube reported.

1. 01

   YouTube's own APIs

   A channel owner connects their channel and grants read access. Three of YouTube's APIs are read: Data for the catalog, Analytics for views, impressions, click rate, watch time and traffic sources, and Reporting for the bulk daily files.

2. 02

   One read a day

   06:00 UTC daily

   A single pass runs each morning at 06:00 UTC. It only ever reads; nothing is written back to YouTube, and no video, title or thumbnail is ever changed.

3. 03

   The seam

   Recorded through Aug 25, 2026

   YouTube's reporting runs two to three days behind. Every chart and every total on this site stops at the last day that actually reported — the seam — rather than showing the pending days as zeros. Days that reported nothing are drawn as empty, never as a zero.

4. 04

   The file

   94 tables · schema v81

   Each channel's own numbers are written into a single SQLite file that describes itself: every table, every column, its grain and its caveats. That file is what this site reads, and it is downloadable in full.

5. 05

   The reading

   82 patterns

   Pure functions read that file and say what happened. Each one is gated: it stays quiet until the channel has enough of the kind of history it needs. Nothing on a page is written by hand, and nothing is generated by a language model.

Every number here has conditions attached — a reporting delay, a thin sample, an attribution gap, a field YouTube doesn’t expose. Each is named, with what it means when you read the number.

For each thing this site is willing to say, the strongest good-faith rebuttal to it, the guardrail already in the code, and why it still ships. These are arguments against the product, not a defence of it.

• Single-label channel-state compression

  What the app says

  The dashboard reduces the channel's current algorithmic situation to one label (e.g. suggested_run, search_tail, dormant) from a fixed set of states.

  Strongest counter-argument

  A channel can be in multiple states simultaneously — for example, in a suggested run on one video while other videos are dormant. A single label hides this co-occurrence and may assign the wrong dominant frame for a given reader's question.

  Guardrail in the code

  The state label is backed by two support metrics from event_channel_state_change that reflect the actual weighted evidence for the classification, and the compression is named explicitly wherever the state label is shown.

  Why it still ships

  A single-state label is more actionable for most readers than a multi-state matrix. The support metrics partially compensate for what the single label omits.

  Conditionssnapshot-reconciliation

• Traffic-source attribution gaps

  What the app says

  Traffic source breakdowns claim to show where views came from (Browse, Search, Suggested, External, etc.).

  Strongest counter-argument

  YouTube does not expose every source, and the per-source impression rows often don't sum to the per-video total. The unattributed share can be substantial on some channels or source types. Comparing sources by share may over- or understate the true split.

  Guardrail in the code

  Unattributed impressions are shown as their own number rather than being redistributed into the named sources. The limits registry names this as a first-class attribution limit.

  Why it still ships

  Attribution data is the best available signal for how videos reach viewers. The unattributed gap is visible, not hidden, so readers can calibrate accordingly.

  Conditionsunattributed-impressions

• Edit before/after observational limits

  What the app says

  The Changes page shows click rate before and after a title or thumbnail change, implying a before/after comparison.

  Strongest counter-argument

  Many factors change simultaneously with or shortly after an edit — seasonality, topic trends, the age of the video at the time of the edit, and YouTube's own ongoing impression testing. The before/after comparison observes correlation in time, not causation.

  Guardrail in the code

  The before-and-after copy states explicitly: "Whether the edit caused the change is observational, not provable — many things move in parallel." No causal language appears in the rendered prose.

  Why it still ships

  Knowing that click rate rose or fell around an edit is a factual observation that has informational value even without a causal claim. The hedge is present at the top of the page and in every era comparison.

  Conditionsctr-thin-sample

• Wisdom threshold flips near boundaries

  What the app says

  The Wisdom panel renders verdicts (agrees / disagrees / mixed) on conventional YouTube creator beliefs using a channel's own data.

  Strongest counter-argument

  Verdict boundaries are threshold-based. A channel sitting just below a threshold (e.g. a 0.1 pp difference in CTR) flips from 'agrees' to 'disagrees'. At that boundary the verdict carries less signal than the raw numbers do.

  Guardrail in the code

  Verdicts include the underlying numbers that drove them — the evidence is surfaced alongside each verdict, not just the label, so readers can assess near-boundary cases.

  Why it still ships

  Structured verdicts are substantially more scannable than raw numbers alone. The evidence display lets attentive readers catch boundary effects.

• Detector-family consolidation hiding module nuance

  What the app says

  The Moment chapter and Story timeline surface one insight per render, selected from the full detector inventory.

  Strongest counter-argument

  Collapsing dozens of detectors into one featured insight per render means many relevant signals are not shown. The selection algorithm prioritises by magnitude, recency, and certainty — but a less prominent signal may be more relevant for a specific reader's question.

  Guardrail in the code

  The /help/detectors page lists every detector with its activation stage and sample-size threshold. The selection algorithm is transparent about its priority ordering. Story timeline events surface the most significant insight per week.

  Why it still ships

  Showing all active detector outputs simultaneously produces an unreadable wall of context-free numbers. One featured insight per render is the minimum to remain coherent. The full detector inventory — every detector with its activation stage and sample threshold — is listed for readers who want it.

Every pattern the reading engine can report, by the part of the channel it reads. Each one stays quiet until the catalog is large enough to read it honestly, so a pattern that has never matched has usually never had the history it needs — not a channel that lacks it.

• Milestones0 of 2 matched

  A milestone reached, An emerging signal

• Trajectory and outliers0 of 11 matched

  Where this channel sits against the shared pool, A video tracking off the channel norm, A video's view-curve deviating from its peer norm, Views accelerating, A week-over-week step up, Day-one numbers predicting longer-run reach, Cross-video outlier, A live video sitting outside the channel's day-N cone for its age, Retention and engagement ranking high relative to view mass, A per-metric trajectory consistent across videos, A title pattern correlating with CTR on this channel

• Discovery and traffic sources0 of 24 matched

  A traffic source proportion shifting, Browse / Homepage opening up, Browse / Homepage going quiet, YouTube Search reaching the channel, Suggested Videos lifting, Suggested Videos cliff, Shorts Feed lifting, Shorts Feed cliff, Where views arrive from, over time, Single-source-dominated day, Top-country flipping week-over-week, First External-source views, Channel-wide top traffic source flipping, Per-source typical days-to-activation, A specific video's source activation outside the channel's band, First impressions on a video, A video drawing most traffic from one source, Source mix widening, Source mix narrowing, A source going quiet on a video, A gap in watch time per view across sources, A gap in subs-per-view across sources, A previously-quiet organic source firing again on a video, A video drawing most views from a single country

• Per-video lifecycle0 of 11 matched

  This video's arc, A video gone quiet, Impressions without views, Days from publish to peak, Older videos vs newer videos, CTR rising late in life, Watch time rising late in life, Trend across recent debuts, Channel trend, week over week, Trend across recent videos, A video launched in a weak channel phase performing under the channel median

• Engagement0 of 14 matched

  Two metrics that usually move together pulling apart, A video drawing more engagement per view than the rest of the channel, Comments-per-view well above the channel's band, A trailing-3-day jump in shares on a video, Subscribe rate and reach moving in opposite directions, Subscribers stepping up on a routine view day, Subscribers lost in the trailing two weeks, Audience replacing itself faster than its trailing weeks, How much of last week's audience mix came back, Channel attribution share trending up, Per-video attribution share well above baseline, A sharp drop-off in the retention curve, Mid-video hold versus the channel's own catalog, Where this video differs across the four funnel stages

• Channel shape0 of 20 matched

  Lifetime distribution of views across the catalog, Channel's 5-dimensional state vector outside its 90-day envelope, Recent publishing gap notably longer than the channel's median, How often this channel publishes, Recent 7-day publish count diverging from the prior 7-day count, One or two weekdays consistently outperforming the rest, The catalog shape shifting, A metric outside its usual range, How retention distributes across the catalog, Share of the catalog that went the past week without views, Top-video and quiet-inventory shares trending up together, Share of videos below a view threshold, Share of fitted videos projected quiet within 30 days, A video projected quiet within two weeks, Before / after a title or thumbnail change, How a title or thumbnail change reshuffled traffic sources, How often a video was iterated on, Whether iterations moved the metric, How fast YouTube is testing a video, An impression spike and what followed it

Every projection shape this site can produce was replayed against what actually happened afterwards, across every connected channel. Two shapes earned their place. One did not, and does not render anywhere on this site as a result.

• A range for a number, at a date you pick

  Renders

  Subscriber ranges contained what happened 94% of the time across 415 past calls; daily-view ranges 88% across 306. Both clear the 80% these ranges are drawn to hold.

• A date for a number you pick

  Does not render

  Calls that named the day a channel would cross a milestone held 21% of the time across 212 past calls, and the middle miss was 12 days wide. That is far under the 80% bar, so no dated crossing is shown anywhere — not on a dashboard, not in an email, not on a share card.

• The pace a channel is moving at

  Renders

  Every channel that is moving carries a pace read. It states what the recent days did, so it is a description of the record rather than a claim about the future.

The replay ran on Aug 27, 2026 against the code at b6ca0901, and its full report is kept at .ai/notes/hindcast-report.md.

On @iamartharrison's own record, no past call has resolved yet — that record starts as soon as the first one does.

The steps of connecting a channel are the only thing this site counts, and the daily totals are published in full. Connect funnel counts.

Reading needs no account and sets no cookies of its own; a few of your own settings — the theme, whether motion plays — are remembered in your browser and go nowhere else. Signing in does set a session cookie, and that is the only place an account exists: the console a channel owner uses to choose what their channel shows.