The science
“Good” should be measurable. Every Chasca video is built to a published spec: eleven design laws graded by the strength of their evidence, a table of numbers each video must hit, and an eval harness that checks them before anything ships.
Based on the open explainer-evals spec (MIT, © Sketchie). Summaries below are ours; effect sizes and sources are the spec’s.
How to read it
Survived adversarial verification by independent reviewers, with a primary source cited.
Follows arithmetically from rules that are verified.
Practitioner canon or a careful extension of a known principle. Useful, not yet verified.
A product decision we made on purpose, not a research claim.
An engineering requirement that makes the other rules checkable by machine.
Part one
What the learning-science literature actually supports, and what each finding changes about how Chasca draws, narrates and cuts.
Diagrams built up stroke by stroke beat finished diagrams under identical narration: retention ηp² = 0.13–0.17 and transfer ηp² = 0.10–0.16, replicated across two experiments.
In Chasca: Nothing pops in. Icon outlines stroke on in pen order, color washes in after, text writes itself and arrows draw from tip to tail.
Source: PMC9898452
Drawing with a visible hand, pushing finished pictures in with a hand, or no hand at all changed motivation (p = 0.033) but made no measurable difference to learning.
In Chasca: No cartoon hand in the frame. The drawing does the teaching and the picture stays uncluttered.
Source: Smart Learning Environments 2023
Across 32 studies (N ≈ 3,597), cueing improved retention (d = 0.27) and transfer (d = 0.34) and cut cognitive load (d = −0.11). The bigger the drop in load, the bigger the gain (β = −0.70 for retention, −0.60 for transfer).
In Chasca: Arrows, labels and color are treated as cues, so each one lands on the word that names it, within half a second.
Source: PMC5576760
A meta-analysis of segmenting found significant, small-to-medium gains in retention and transfer, partly from the processing time that breaks between segments create.
In Chasca: One idea per scene, 12 to 30 seconds each, and at least 0.7 seconds of quiet after the last reveal before the cut.
Source: Rey et al. 2019
In 6.9 million edX sessions, engagement stayed near full under six minutes, fell to about 0.55 at 9–12 minutes and to about 0.2 past twelve. An independent 2026 study points the same way.
In Chasca: Videos top out at 6:00. Ask for more and Chasca plans a chaptered series instead of one long video.
Sources: Guo et al. 2014 · Nature HSSC 2026
Khan-style drawing held about 0.72 normalized engagement against roughly 0.52 for slides at 3–6 minutes, credited to hand-drawn motion and natural, unscripted-sounding speech.
In Chasca: Whiteboard drawing is the format, and scripts are written to be spoken, contractions and all.
Source: Guo et al. 2014
A narrative script improved transfer (ηp² = 0.12–0.20). With progressive drawing it also beat a plain informative script for retention (p = .021); over static visuals, the informative script won (η² = 0.21–0.24).
In Chasca: Scripts use story grammar (someone wants something, something stands in the way, here is how it resolves) precisely because every frame is drawn progressively.
Source: PMC9898452
Interesting but extraneous material does not improve learning.
In Chasca: Every element on screen is named in the narration. Unreferenced decoration fails the quality gate.
Source: Mayer, Springer 2020
Asking learners to produce an answer (the generative activity principle) improves learning more than handing them a recap.
In Chasca: Every video closes with a retrieval question, a pause of 1.5 to 2 seconds to think, then a one-line recap.
Source: Mayer, Springer 2020
More visual and narrative change points go with less mind-wandering.
In Chasca: Cuts land on idea boundaries, never in the middle of one.
Source: Nature HSSC 2026
Once channel size is controlled, explanation quality does not correlate with views (r = 0.23, not significant); only likes and relevant comments track it.
In Chasca: We grade learning design, not virality signals.
Source: arXiv 2207.05872
Part two
The spec every video has to hit. Grades show how firmly each number is grounded.
| Parameter | Target | Grade |
|---|---|---|
| Total length | 6:00 hard ceiling; 1:00 to 3:00 is the sweet spot for a single concept | Verified (L5) |
| Length steps | 30-second increments | Product |
| Scene duration | 12–30 s, one idea per scene | Verified (direction, L4) Heuristic (band) |
| Elements per scene | 3–7 visible when the scene ends | Heuristic |
| Reveal cadence | A new element every 2–5 s of narration, never two within 1.2 s | Derived (L1, L3, L4) |
| Reveal-to-word sync | Within ±500 ms of the trigger word | Derived (L3) |
| Word-sync coverage | At least 95% of elements carry a reveal word that resolves | Pipeline |
| Post-scene gap | At least 0.7 s of silence after the last reveal, before the cut | Verified (direction, L4) Heuristic (value) |
| Reveal style | Outline draws on (0.3–0.5 s), then fills wash in; arrows draw tip to tail | Verified (mechanism, L1) |
| Visible hand | None | Verified (unnecessary, L2) |
| Labels | Three words at most, ALL-CAPS keywords, never a sentence lifted from the narration | Heuristic |
| Narration pace | 140–160 words per minute, spoken register | Heuristic |
| Words per scene | About 35–75 | Derived |
| Script structure | Hook in the first 10 s, concrete analogy, mechanism, real-world anchor, retrieval question, recap | Verified (parts, L7 + L9) Heuristic (order) |
| Narrative grammar | Actor, goal, obstacle, resolution wherever the topic allows | Verified (L7) |
| Analogy | At least one concrete analogy before the first abstract term | Heuristic |
| Extraneous visuals | Zero elements the narration never mentions | Verified (L8) |
| Ending | Retrieval question, a pause of at least 1.5 s, then a one-line recap | Verified (L9) |
| Scene transitions | Hard cuts at idea boundaries only | Verified (L10) |
Part three
Cheap, objective checks catch most problems before a cent is spent; judgment and vision checks cover what rules can’t.
Layer A
Hard pass/fail assertions over the scene graph and its word timings. No model, no cost.
Every video, before narration; again with real timings; and in CI
Layer B
A model grades each video 1–5 against a rubric. Any score under 3 blocks release; the target mean is 4.
Sampled, and on every release of the generator
Layer C
Spot checks on the rendered MP4, looking at real frames and real audio.
On rendered output
Composite score: 50% Layer A, 35% Layer B, 15% Layer C. The ship gate is 85.
In practice
Layer A runs on the scene graph before any narration is recorded or any frame is rendered. When a check fails, the failure goes back to the script writer as specific feedback and the graph is regenerated. Once the voice is recorded, the timing checks run again against the real word timings.
The same checks run in CI, so a change to the engine that makes videos worse can’t quietly ship.
Chasca’s Layer-A gate is a port of the open explainer-evals harness (MIT, © Sketchie). github.com/Downshift/explainer-evals
Your first 2 videos are free, up to 1 minute each, watermarked. No card required. Full product, scene editor included.