From the desk

Clips are here

Create short practice clips from any video. Loop movements, tag techniques, and build your personal library.

Engineering notesVolume 5 of 8 · People in the loop

Volume 5

The Community Loop

The machine never guesses, so what it cannot decide piles up. Visitors, dancers, and a reviewer at one desk work the pile, and every decision they make flows back into the rules, the search, and the recommendations. This is how a catalog gets truer every day without anyone being paid to make it so.

questions waiting for a person: 122,059questions waiting for a person
taps from strangers verify a credit: 3taps from strangers verify a credit
trust needed to act alone, then to be believed: 5 · 20trust needed to act alone, then to be believed
piles on the desk: 8piles on the desk
the weekly edition goes out: Mon 10:00the weekly edition goes out

Part 01

A stranger is a witness with a grade

Someone objectsA missing dancer, the wrong song, a swapped role. It is a claim, not an edit.

Same rules as a machineThe claim is a clue in the notebook, and the visitor's trust is its grade.

Autonomy is earnedAccepted claims raise trust; rejected ones lower it. The ceiling is a reviewer's authority.

Crowdsourcing fails when it is a separate system with its own rules. Here a stranger is one more witness, held to the same five-line judge as a title parser or a fingerprint. A claim that names a dancer nobody has heard of goes through the same door as any new name in Volume 3.

Part 02

Three strangers who agree are a fact

One controlUnder the dancer names on every video: one tap that says “yes”.

Count, don't judgeThe system never decides whether a visitor is trustworthy. It counts agreement.

Ground truth for freeConfirmed credits tell us how precise each matching rule really is.

This is the same rule that verifies a recording when two sources agree, applied to people instead of parsers. Three is the number because two is a coincidence and four is friction. The nickname rule’s low grade in Volume 1 was set by comparing what it claimed to what people confirmed.

Part 03

One desk, two questions

Not a queueA queue sorted by age fills faster than it drains. Nobody empties a landfill.

HeartbeatCoverage, what was applied, and whether each known class of mistake has stopped arriving.

WorkEight piles, each one cheap signal. A reviewer names a festival's whole set in the time the queue took for one row.

The reviewer works in bills, not rows: a festival’s four videos of one couple in one week share the couple and the event but never the recording, because a set is four different songs. The desk copies what may be copied and forbids the rest. Every view is also a plain data feed at the same address, so an offline session, human or model, can read the heartbeat, decide a pile, and hand the decisions back as a file without logging in.

Part 04

Mine the disagreements, not the miles

Shadow replayLike a self-driving fleet: improvement comes from the disengagements, not from the miles driven.

Cluster by reasonWrong credits, missing credits, refusals, and correct refusals each get a pile.

Pin itA decision made once outlives the session that made it, as a test.

The heartbeat’s real question is not “how many facts were applied” but “did this kind of mistake stop arriving?” A quiet day with nothing new to learn is a good day.

Part 05

Follow a person, and read the paper on Monday

Follow the personThe catalog knows who is in the video, so you follow the dancer, not whoever uploaded it.

Gathered nightlyEach follow has its own level: everything, personalised, or muted. Nothing is sent twice.

The weekly editionThe week's new dances as a front page, with tour dates from the dancers you follow.

Recommendations come from the same tables and can always say why: “because you watch Di Sarli vals” is a join on band and genre; “more from this partnership” is the couples map; “people who saved this clip for practice also saved these” is a count. Fame stays a floor even on your personal shelf, because the ranking recipe from Volume 4 orders everything.

Take-aways, without the tango

01Treat a user as one more witness in the same notebook, with a grade, not as a separate editing system.
02Earn autonomy in tiers by a score only being right can raise. Verify by counting agreement, not by judging people.
03Replace scoreboards and backlogs with one desk that asks “did it get truer?” and “what is the cheapest next fix?”
04Mine disagreements, not activity. Decide once, pin it as a test.
05Let people follow the thing they care about, and build recommendations from joins that can explain themselves.