Originally published on hexisteme notes.
On 2026-11-09 I sent the effect plans for two explainer shorts out for adversarial review, one about a famous painting and one about a mountain park. Codex reviewed both; agy also attacked the painting short, and Gemini read the mountain short's script in an earlier round. Many of the findings had the same shape: this effect, as drawn, states something that is not true.
A gray wash on a forest reads as "the forest died." A PHOTO label on a synthetic cleaning animation reads as a forged photograph. Fireworks read as a celebration that someone recorded. None of these effects were meant as statements. The reviewers read them as statements, and in each case I checked the reading against the sources before deciding.
This note is about how I sorted those findings, which ones I refused, the one line I left to a human because the two reviews pointed in opposite directions, and the rule that came out of it.
Why the effects were there at all
Some context on why the plans were full of effects. I had complained repeatedly, across channels, that my shorts had no reactions in them: "I need you to look at each line of the script carefully and mix in suitable effects, and there are none." On November 8 that became a written rule for all my shorts: every line gets at least one event where something on screen physically changes, stickers alone don't count (stickers at most 40 % of events), and feedback like "more effects" is never converted into a cap. Caps are soft fatigue warnings only; the floor is the contract.
My main Claude session also wrote a diagnosis of why the same complaint kept coming back. The existing gate only measured whether something changed: mean brightness difference inside the event box at least 4/255, a sound effect present for every change, and a width check of at least 7 %. In the diagnosis's words, nobody judged whether the effect matched the meaning, or whether it was big enough to react to, or overdone. There was also no step where effects were chosen. They were hand-coded line by line during the build, so the repertoire collapsed to moves, palette cards, color fills and sweeps. The diagnosis's fix was an effect plan written after the script freezes and before the build: for each line, the emotional beat, the image the words call up, the on-screen effect, the sound and the size. Then a check, then an adversarial review by Codex and agy, then measurement after render against the plan.
So by the time the reviews ran, the plans were deliberately dense. That matters for what I did with the findings.
What the reviewers flagged
The painting short's round 1 came back Codex FAIL/FAIL and agy FAIL/FAIL. Many of the script findings were ordinary: too many words for the length limit, a tangent outside the sources, an overclaim like "everyone thought." The effect findings are the ones this note is about. Here they are from the triage table, with what replaced them:
- A palette blob being "kicked" with a thump read as a violent act. The line became "left … off his palette," and the blob bounces away.
- Draining saturation from the street read as uniform gray, when the source describes colored grays. The replacement is a row of actual gray swatches sampled from the painting, standing like a choir.
- A BANNED chain drawn across the canvas read as "the painting was banned." The stamp moved to a separate tricolor flag card plus two crowns.
- Fireworks and cheering read as a documented celebration. They became the card's chain snapping and flying into the flag as a ring of color, labeled "REAL SUBJECT — Louvre".
- A synthetic cleaning animation carrying a PHOTO marker read as forged evidence. Only a real comparison of 2024 before and after photographs keeps PHOTO; the gray-to-yellow shift is drawn on a MODEL layer.
- A thermometer with a boiling sound emphasized heat and the body, and the "warmest spot" sat over a protected region of the figure. It became a gold halo behind the head plus warm and cool color tiles labeled "WARMEST SPOT," and the narration was changed so that what it points to is what is visible.
- A "#2" badge erased the "probably" in the narration. It became "MAYBE #2".
Each of these is an effect saying something the script did not say. The reviewers did not object to the effects being loud. They objected to what they asserted.
Some of these findings turned into checks on the plan itself. Each effect row now must state what it teaches. A row with photo status must carry evidence, a source id or a real file, and its motion is limited to comparison and zoom. An IMAGINE marker must have a size_px of at least 48, because a one-pixel chip technically labels the effect as imagined while nobody can see it. Text effects need text_px of at least 54 and dur_s of at least 1.5, so a label cannot flash past. One finding I did not turn into code: agy caught narration saying "up" while the screen showed the feet. That stays an item for adversarial review and eye checks.
The blind round, and the criterion
The mountain short went through more rounds. For round 3, Codex got a blind sheet generated from the reactions file: each line's dialogue and on-screen events, with the viewer's inner thoughts and the roles masked. One Codex run, verdict FAIL.
The criterion I applied was: fix only the places that read opposite to the facts. Findings of the "cut the effects down" kind would not be adopted, because they contradicted my own instruction (effects and sound on every line) and the rule that feedback is never turned into a cap.
That criterion sorted the round cleanly in most places.
Adopted, two findings:
- b11: the inside of the circle going ash-gray, to show land excluded from the return, read as the forest dying. The fill became a gray 45° hatch with the forest color still showing between the lines, plus an EXCLUDED label.
- b14: a card reading only CROWN APOLOGY let the apology attach to the previous line's claim about the reserve. The card became
APOLOGY FOR 1865.
Rejected on the facts, three findings:
- b15: Codex read the gold circle completing as the excluded land being returned and resolved. I kept it, because under the 2025 law the park land is vested in the new legal person; the circle is that land.
- b10: an inward arrow read as a return inside the circle. I kept it, because the next line ("skipped") bounces it back, and that correction is the reversal the narration is built on. The arrow itself had been introduced after round 2, when Codex read an outside fill as the extent of the return.
- b13: an 1881 stamp right after "wasn't" seemed to clash with the negation. The stamp reads RESERVE MADE · 1881 and the previous line's note reads Land returned · 1880, so the year order is on screen.
Rejected by rule: the remaining lines, b01–b04, b06–b08 and b12, where the findings asked to remove or soften effects and sounds (the radius figure, a cow emoji, a hiss, a blackout, a tilt). Audibility is measured separately on a phone band instead of being argued about in review.
Held for a human, two findings:
- b09: a red tint spreading with a shake evoked a bloody event at the summit. The color is explanatory, and the 1865 confiscation came after a war, so the effect is not opposite to the facts. Whether to make it less violent is a judgment I left to a person.
- b05: bringing the forest color back on "became" read as a forest created or recovered in 1881. A note saying FOREST RESERVE · 6-mile radius already appears on that word. And the forest color was there because Gemini, in round 2, said the earlier cream fill hid the contrast. Two reviews pointed in opposite directions on the same line.
I did not pick a side on b05. My criterion says fix what reads opposite to the facts; one reviewer says the current effect does, the previous reviewer said the previous effect failed for another reason, and the narration already carries a label. That is a judgment call, so it went to a human with both readings attached.
The two adopted fixes went into the same build, build 7. I did not record a pass for the script stage. The blind pass for the final contract is decided by human review.
The same failure in the words
The effect findings had a twin in the script reviews. In the mountain short's round 1, Gemini offered a rewrite of one line that opened with the Crown confiscating the land first, as the step toward marking out the boundary. It reads smoothly, and it is a causal claim: that the confiscation happened in order to mark the boundary. The sources do not say that, so I rejected it. Gemini's other suggestion on that line, replacing "the Crown" with "the government" because the term is unfamiliar, I adopted.
An effect that turns gray and a sentence that adds "in order to" fail the same way. Each adds a cause or an outcome the sources don't contain, and the viewer has no way to tell it was decoration.
What became a rule
Rule 3 of the reactions rule now reads, in short: effects follow the meaning of the line. The example is a line about firing at an atomic nucleus: the glow, the steam sound and the mushroom cloud happen inside a thought balloon, and the balloon pops on the next line. Alarm staging for a wrong guess (red heat, steam, smoke, clouds, warning lights) stays inside the imagination balloon only. Put it in the real scene and the false causation survives even after the narration says "it wasn't." The rule cites an adversarial review round 3 from that same day (11-09) as its origin.
The last sentence of that rule is the lesson of the round. An effect shown in the real scene is a claim about the real scene. A later line of narration can retract the words, but the image has already been shown as if it happened.
Rule 6 is the limit on all of this. Fun cannot be measured by a machine. What gets recorded is a blind script review, with inner thoughts and roles masked, and silent and sound-on viewings of the animatic. A gate PASS is never reported as a verdict on whether a short is fun. Nothing in this note says the shorts got better to watch. The reviews told me where effects asserted something false; they did not tell me whether anyone enjoyed them.
What is still open
b05 and b09 are with a human, and I have no outcome to report for either. The criterion "fix only what reads opposite to the facts" depends on a reviewer noticing the misreading in the first place, and round 3 was one Codex run; I do not know what a second blind reviewer would have flagged on the same sheet. The checks I added (teaches, evidence, marker size) make a plan state its claims, but whether a stated claim matches what a viewer reads still comes from review and eye checks, not from the checker.
Email list for these notes: hexisteme.beehiiv.com — no issue has gone out yet, so you would be on it before the first one. No welcome sequence, no course, no upsell.
More notes at hexisteme.github.io/notes.
Top comments (0)