Performance learning
A production system that does not learn is a very expensive render farm.
The point of governing a production is that the record of what was decided can be compared against what happened. Without that comparison, episode forty is made with exactly as much information as episode one.
01 / Inside the workflow
The signals that matter
For serialized entertainment the commercial unit is the season, not the clip, and the principal value signal is the viewer's next-episode action. Views are a vanity number in a format designed around return.
- Three-second hold: whether the hook survives the thumb
- Per-episode completion
- Episode-to-episode conversion
- Pilot-arc completion across the five episodes
- Return rate across release days
- Language-track comparison on the same episode
- Whether recurring characters are actually recognised
For branded stories the metric is whatever was agreed in the brief - hold, completion, click-through, qualified traffic or conversion - alongside production-side measures like acceptance without material rework and the number of usable variants per approved master.
02 / Inside the workflow
The retention gate
The most valuable thing performance data does is stop production. A five-episode pilot arc exists so that the decision to render twenty to sixty episodes is made with audience evidence rather than optimism.
- 01Publish the pilot arc
Five episodes, released on a cadence the audience can follow.
- 02Measure the behaviour that predicts a season
Hold, completion, episode-to-episode conversion and return rate, per language track.
- 03Meet the gate, revise, or stop
Continuing is a decision, not a default. Stopping after five episodes is a far better outcome than discovering the same thing after forty.
- 04Feed what held into the next arc
Hook placement, episode length, cliffhanger type and which track led. These are production decisions, and they are the ones that change.
03 / Inside the workflow
What performance data is not allowed to do
Questions
Before you bring it into production
Where does the performance data come from?
From the platforms you publish on, supplied by you or by a distribution partner. Tosheo is not a viewer platform and does not have its own audience, so it reads the numbers you have access to rather than generating them.
What is the retention gate for?
To make continuing a deliberate decision. The pilot arc is five episodes precisely so that the cost of finding out the story does not hold is five episodes rather than a season. Continue, revise, or stop are all acceptable outcomes.
Can performance data change my characters?
No. It can inform what gets proposed - a different hook placement, a tighter episode, a different language track leading the release. Changing canon requires an approval from a named person, the same as every other canon change.
Do you publish performance benchmarks?
Not yet, and not until they are ours to publish. The acquisition sequence in our brief calls for publishing transparent production evidence - accepted-minute cost, repair examples, continuity results, language-track results - from our own demonstration production. Numbers from someone else's production are not evidence about Tosheo.
How is a branded campaign measured differently?
The metric is agreed in the brief before production, which is the part that most often does not happen. Alongside it we track acceptance without material rework, time from approved brief to delivery, cost per accepted asset, and how many usable variants came out of one approved master.